Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Wrongful Convictions and Cumulative Evidentiary Error

Abstract. Wrongful convictions emerge from complex interactions among memory, interviewing, forensic interpretation, investigative selection, and adjudication. This fundamental research monograph examines cumulative evidentiary error: how early misinformation, suspect-centered hypotheses, or false confessions can alter later ostensibly independent evidence and make a case appear more strongly corroborated than its informational provenance warrants. The historical and conceptual analysis integrates primary experimental studies of eyewitness postidentification feedback, confession-induced identification changes, forensic context bias, and juror interpretation with scientific reviews, studies of DNA exonerees, exoneration registry data, and methodological work on forensic validation. It distinguishes experimental mechanisms from archival association, erroneous convictions from later exonerations, conditional dependence from genuine independent corroboration, and forensic classification from courtroom overstatement. The synthesis evaluates base rates, case selection, coding reliability, probabilistic evidence combination, risk of investigative tunnel vision, institutional incentives, and the preservation of information provenance. It examines procedures for recording witness and suspect accounts, controlling irrelevant forensic context, auditing corroboration, validating analytic methods, and managing digital evidence and artificial assistance. The conclusion identifies an integrated multistage error model supported by convergent but methodologically distinct evidence, and delineates a research agenda for prospective audits, international comparison, and causal evaluation of safeguards. This is a critical narrative research synthesis; no new trial, systematic review, meta-analysis, peer review, or overall wrongful-conviction prevalence estimate is claimed.

Keywords: wrongful convictions; cumulative evidentiary error; forensic psychology; false confessions; eyewitness identification; memory contamination; confirmation bias; forensic contextual bias; tunnel vision; exoneration; forensic science validation; corroboration inflation; information provenance; conditional evidence dependence; juror decision-making; expert testimony; DNA exoneration; legal judgment; probabilistic reasoning; digital forensic evidence; Artificial Era

Wrongful Conviction as a Psychological and Evidentiary Research Object

Wrongful conviction denotes the legal condemnation of a person who did not commit the offense for which that person was convicted, or the conviction of a person for conduct later established not to have constituted the alleged crime. Cumulative evidentiary error refers to the production, transmission, interpretation, and combination of inaccurate, misleading, or improperly weighted information across an investigation and adjudication. Its distinctive psychological question is how several seemingly corroborating items can emerge from connected cognitive and institutional processes. The unit of explanation is a sequence of decisions and information transfers, extending from the initial construction of a suspect through interrogations, witness procedures, forensic examinations, disclosure, adversarial presentation, and postconviction review. The National Registry of Exonerations documents several kinds of contributing factors, often co-occurring within a case, while laboratory research tests mechanisms through which one evidence item alters another National Registry of Exonerations (2026) Kassin et al. (2012) Findley & Scott (2006).

Three outcomes require separation. Factual innocence concerns the historical relationship between an individual and the charged conduct. A legally invalid conviction may be overturned because of procedural or constitutional defects without a definitive resolution of factual innocence. Exoneration is a documented postconviction event meeting the recording organization's criteria and can occur through several procedural routes. None of these is synonymous with the underlying prevalence of wrongful conviction: recorded exonerations depend on detection, access to legal assistance, evidence preservation, and institutional willingness to revisit cases. Research therefore requires explicit inclusion rules rather than treating all reversals or all unresolved accusations as equally established innocence cases Gross et al. (2014) National Registry of Exonerations (2026).

The monograph synthesizes forensic psychology, cognitive psychology, statistical reasoning, legal evidence scholarship, and institutional research. It compares controlled experiments on memory and professional judgment with archival analyses of exonerated defendants and scientific assessments of forensic disciplines. It evaluates source dependence, selection bias, task-relevant versus task-irrelevant contextual knowledge, disclosure of uncertainty, and feedback loops among practitioners. This is a critical narrative evidence synthesis. It neither creates a new individual case database nor reports an original experiment, preregistered systematic review, or estimated national rate of wrongful conviction National Research Council (2009) National Research Council (2014) Morgan (2023).

Historical Development of the Scientific Problem

Scientific concern with mistaken evidence preceded contemporary DNA testing. Work in experimental psychology established that observation, memory, and judgment depend on how information is encoded, retrieved, and socially interpreted; these principles later became central to eyewitness studies. Wells and Bradfield's 1998 experiments demonstrated that confirming feedback administered after a mistaken identification altered witnesses' later retrospective assessments of certainty and viewing conditions. The historical contribution was an experimentally specified route by which evidence could become more persuasive without acquiring greater accuracy. That finding made procedural history part of the evidentiary object: the confidence expressed at trial could reflect intervening feedback rather than confidence at initial identification Wells & Bradfield (1998) National Research Council (2014).

DNA-based exonerations created a second historical transformation because independently interpretable biological material sometimes revealed factual mistakes that had survived the ordinary court process. Studies of case transcripts then examined patterns in the evidence that had originally supported conviction. Garrett and Neufeld's analysis located trial transcripts for 137 cases in which later DNA exonerees had faced forensic testimony; they classified invalid testimony in 82 cases. The result concerns a selected set of trials rather than the error rate of forensic practice generally. Its conceptual significance is the demonstration that authoritative courtroom testimony can exceed the scientific strength of the underlying analysis, placing the translation of technical findings into legal propositions under scrutiny Garrett & Neufeld (2009) Hampikian et al. (2011).

The National Research Council's 2009 report on forensic science addressed the need for validated methods, governance, research, and standards; its 2014 report on eyewitness identification connected scientific findings to concrete police and courtroom procedures. In parallel, Findley and Scott analyzed tunnel vision as a recurrent institutional pattern, while Dror, Charlton, and Péron experimentally demonstrated susceptibility of fingerprint judgments to extraneous context. These developments converged on the idea that errors are shaped by the entire decision environment. The historical literature thus moves from isolated fallible witnesses or experts toward interacting information systems and institutional safeguards National Research Council (2009) National Research Council (2014) Findley & Scott (2006) Dror et al. (2006).

A Taxonomy of Evidentiary Error and Dependency

Errors can arise at acquisition, preservation, generation, classification, interpretation, communication, and combination. Acquisition errors include suggestive questioning or lineup procedures. Preservation errors include the failure to retain a witness's first account or original laboratory notes. Generation errors include a false statement elicited through coercive interviewing, while classification errors concern incorrect forensic source decisions. Interpretation errors turn a limited observation into an overconfident narrative; communication errors omit uncertainty in testimony or reports; combination errors treat products of a common informational origin as independent corroboration. This taxonomy is an analytic organization of documented mechanisms, not a claim that every case displays every category Kassin et al. (2010) Morgan (2023) Wells et al. (2020).

Dependence has several forms. Causal dependence occurs when knowledge of a purported confession changes a witness's later identification. Common-source dependence occurs when multiple accounts derive from the same news report or police suggestion. Institutional dependence occurs when analysts, investigators, and prosecutors receive a shared theory of the case before making nominally separate judgments. Statistical dependence concerns correlation among evidence conditional on guilt or innocence; it may exist even without direct influence because observations share latent causes. Procedural independence must accordingly be documented, and statistical independence cannot simply be inferred from the fact that two people signed separate reports Hasel & Kassin (2009) Kassin et al. (2013) Dror (2020).

Cumulative error also has a temporal structure. Early choices define which suspects and hypotheses receive attention, what evidence is collected, and what alternative observations remain available. An initial weak lead may become stronger through repetition, selective recording, and subsequent decisions that use it as a premise. Conversely, an early inaccurate inference can remain detectable when original observations are preserved, analysts are shielded from irrelevant facts, and decision records identify what was known at each stage. Reconstruction therefore requires a timestamped provenance for each claim: who first observed it, how it was recorded, what later disclosures occurred, and which actors relied upon it Findley & Scott (2006) Kassin et al. (2012) Morgan (2023).

Counting Exonerations and Understanding the Denominator

Exoneration registries are invaluable observational sources because they document discovered failures and can support rigorous coding of case histories. They remain ascertainment samples. The National Registry of Exonerations' report issued April 6, 2026 recorded 97 exonerations occurring in the United States during calendar year 2025, while explicitly warning that delayed discovery adds cases from earlier years. The report coded official misconduct in at least 70 of those 97 cases, false or misleading forensic evidence in 39, mistaken witness identification in 22, and false confessions in 19. These categories overlap and their sums are not independent counts of people. The figures describe cases entering a particular historically evolving registry, not the underlying probability that a random criminal conviction is wrongful National Registry of Exonerations (2026).

Selection is also shaped by offense type and technology. Biological testing can sharply alter the discoverability of certain crimes, while a case lacking preserved biological material may have no comparable path to definitive exclusion. Hampikian, West, and Akselrod examined 194 DNA exonerations and found mistaken identification in 75 percent of that selected sample, false confessions or admissions in 30 percent, and informant evidence in 22 percent. These proportions should be retained with their exact DNA-exoneration denominator, study era, and coding conventions. Applying them to all convictions would misrepresent both the case ascertainment process and differences across offense categories Hampikian et al. (2011) National Registry of Exonerations (2026).

Gross, O'Brien, Hu, and Kennedy used survival methods to estimate the rate of false convictions among defendants sentenced to death in the United States. Their 2014 model indicated that at least 4.1 percent would be exonerated if all remained under sentence of death indefinitely. This is a model-based lower-bound estimate for a narrowly defined death-sentenced population under particular assumptions about changing exposure to exoneration, not a general conviction error rate. The distinction illustrates a broader methodological lesson: observed exonerations, modeled latent innocence, identified contributing factors, and causal mechanisms are different inferential targets and should never be collapsed into one percentage Gross et al. (2014).

Eyewitness Encoding, Reconstruction, and Identification

Eyewitness evidence begins with perception of an event, but the later account is generated through memory, communication, and decision. Lighting, duration, attention, stress, prior familiarity, delay, and the similarity of lineup members can alter identification performance. An investigator must distinguish estimator variables outside immediate procedural control from system variables such as instructions, administrator knowledge, lineup construction, questioning, and feedback. The 2014 National Research Council synthesis emphasized both the scientific foundations of eyewitness identification and procedures that improve preservation of the witness's initial decision. The forensic outcome is not simply whether an individual eventually points to a suspect; it is whether the whole observation-and-identification history supports a specified inference under known conditions National Research Council (2014) Wells et al. (2020).

Memory is updated as people encounter information after an event. The timing of a later suggestion matters because a witness can mistake information acquired during an interview, conversation, or public report for information acquired at the scene. Repeated retrieval may stabilize a particular narrative even when its origin has become uncertain. Source monitoring is therefore a central construct for testimony: evidence of confidence in a remembered detail should be examined alongside records of what was disclosed to the witness before that detail first appeared. Recording an initial free narrative and maintaining an access history provide an evidentiary baseline that later interviews alone cannot reconstruct Wells & Bradfield (1998) National Research Council (2014).

An identification procedure presents a choice under uncertainty. A witness may infer that the perpetrator is among lineup members, select the person most similar to the remembered individual, or decline to identify anyone. The relevant comparison is between procedures and target-present or target-absent conditions, not simply between a witness who chose someone and one who did not. Wells and colleagues' 2020 scientific review recommends evidence-preservation practices based on this distinction, including unbiased administration, carefully constructed lineups, suitable instructions, a verbatim confidence statement at the time of the first identification, and a full record of the procedure. Implementation should be evaluated through both identification accuracy and downstream procedural fidelity Wells et al. (2020).

Confidence, Feedback, and the Changing Evidentiary Record

Eyewitness confidence is a judgment about a memory, and its evidentiary value depends on the conditions under which it is expressed. The first confidence statement, recorded before confirmatory feedback, is an observable response to the identification task. Confidence reported weeks later, after a witness has been told that police have arrested the person selected, is a response to a different information environment. In their 1998 experiments, Wells and Bradfield found that confirming feedback increased retrospective claims about certainty, the quality of the original view, memory clarity, and decision speed even though the lineup was constructed so that participants' identifications were erroneous. These observations directly establish feedback-induced changes in reported evidence quality Wells & Bradfield (1998).

The practical implication is temporal preservation. An investigation that records only the courtroom statement can lose evidence about the witness's earliest uncertainty and the evolution of that certainty. The same principle applies to viewing conditions and narrative details: a later statement may be lucid without revealing whether an earlier statement was hesitant or inconsistent. A full record can show whether increased confidence emerged before or after investigators communicated corroborative assertions. It also permits courts to compare witnesses who received different procedural treatments without equating the witness's final demeanor with initial perceptual accuracy National Research Council (2014) Wells et al. (2020).

Confidence research requires discriminating different scientific questions. Calibration concerns how often particular confidence levels correspond to correct identifications under specified testing conditions. Discriminability concerns how well a procedure separates guilty suspects from innocent suspects. A witness's subjective certainty can also affect a juror's perception of credibility. These three outcomes can diverge, particularly when feedback changes expressed confidence after the identification. Proper forensic interpretation uses the original decision, its initial confidence, and the identification procedure jointly rather than treating a high-confidence witness as a self-contained measurement instrument Wells et al. (2020) National Research Council (2014).

False Confession Formation and the Psychology of Compliance

Confession evidence occupies a distinctive place in criminal investigation because a person's apparent admission supplies an explanation for the offense as well as a statement against personal interest. Psychological research identifies circumstances in which an innocent person may nonetheless confess. Extended detention, perceived inevitability of conviction, misleading claims about incriminating evidence, minimization suggesting leniency, developmental vulnerability, and difficulty understanding the consequences of a statement can change a suspect's choice architecture. Kassin and colleagues' 2010 scientific review integrated confession research across experimental, clinical, and legal methods; the 2025 update elaborated the consequences of false confessions for plea decisions, witnesses, and forensic judgments Kassin et al. (2010) Kassin et al. (2025).

A confession may arise from instrumental compliance, when a person says what an interrogator wants to end an intolerable situation, without believing the statement. Other cases involve more complex pressures on recollection and inference, including uncertainty about one's own memory or acceptance of a suggested account. These categories describe possible mechanisms and demand case-specific evidence; they should not be inferred solely from an eventual recantation. The researcher should reconstruct interrogation duration, legal consultation, recording gaps, question wording, presented evidence, promises or threats, sleep and fatigue context, and the suspect's cognitive and developmental characteristics. Such process measures offer more explanatory value than a simple binary code of confession present versus absent Kassin et al. (2010) Kassin et al. (2025).

An apparently detailed confession can contain facts actually learned from investigators. Details that were public, disclosed during interrogation, supplied by photographs, or introduced through leading questions cannot automatically establish exclusive perpetrator knowledge. Full audiovisual recording from the first substantive interaction is necessary to compare volunteered information with information externally supplied before the purported admission. A reliable evidence history also asks whether investigators verified a detail that was genuinely unknown to the suspect, and whether verification was conducted independently. The 2025 scientific review explicitly recommends entire-interview video recording, science-based interviewing, limits on coercive tactics, legal safeguards for vulnerable suspects, and insulation of other evidence producers from confession content Kassin et al. (2025).

Confessions That Change Other Evidence

The experiment by Hasel and Kassin in 2009 directly challenged the legal intuition that two evidentiary sources can be combined as independent confirmation merely because they originated with different people. Participants witnessed a staged theft and initially made a decision from a target-absent lineup. Two days later they learned that a lineup member had purportedly confessed or denied culpability. Among those who had selected a person and were told another member confessed, 61 percent changed their identifications. Half of those who had initially made no identification subsequently selected the known confessor. The treatment was misinformation about a confession, and the measured outcome was subsequent eyewitness choice; the study did not itself measure false convictions in real courts Hasel & Kassin (2009).

These findings specify a mechanism of evidence contamination. A confession claim can revise a witness's interpretation of the original event and of an earlier lineup judgment. If the witness later testifies to an identification and the confession is presented separately, the fact finder may be invited to treat the two items as mutually corroborative even though the identification was changed by knowledge of the alleged confession. The observed concurrence then overstates the amount of independent information supporting guilt. This is a problem of causal provenance and inferential double counting, not merely a problem of the witness intentionally lying Hasel & Kassin (2009) Kassin et al. (2013).

Kassin, Bogart, and Kerner tested related predictions against an archival sample of DNA exoneration case files. They found that multiple evidence errors occurred more often in false-confession cases than in eyewitness cases and that confessions tended to precede other errors when multiple errors were documented. This sequence is consistent with a corruption hypothesis but cannot establish every individual causal link because archival case reconstruction is vulnerable to incomplete records and selection. The combination of controlled experiments and temporally organized case files is stronger than either alone: experiments show that a mechanism can operate; archives indicate the mechanism is plausible and salient in real cases while leaving room for common underlying causes Kassin et al. (2012).

An additional experimental demonstration comes from Kukucka and Kassin's 2014 research on handwriting judgments. Participants informed that a defendant had confessed were more inclined to classify handwriting samples as originating from the same writer and to judge guilt, compared with a no-confession condition. A second study examined judgments before and after a case summary. The result concerns judgments of handwriting evidence under controlled information exposure and complements the eyewitness experiment. It does not validate handwriting examination as an identification method; rather, it reveals how ostensibly supporting forensic evaluations may respond to information about other inculpatory evidence Kukucka & Kassin (2014) Kassin et al. (2013).

Scientific Validity, Source Attribution, and the Limits of Forensic Conclusions

Forensic evidence comprises very different measurements and interpretations. Some analyses detect specific chemical constituents, some compare biological profiles, and others require human judgments about similarity between patterns. A source-level conclusion about a trace's origin is not automatically an activity-level conclusion about when or why a person touched an object, and neither establishes every element of a charged offense. Valid interpretation requires identifying the exact proposition under test, how comparison material was selected, whether error rates have been measured for the specified task, and which alternative explanations remain plausible. The National Research Council's 2009 review called for stronger scientific foundations, empirical validation, consistent reporting, and appropriate institutional oversight across forensic disciplines National Research Council (2009) Morgan (2023).

Forensic information can be limited by sample quality, contamination, mixed origins, substrate, degradation, collection methods, and incompatible reference conditions. These limitations enter at different points. A measurement may be technically accurate while the chosen comparison or interpretation is misleading; the same apparently conclusive verbal statement can be built on strong or weak method validation. Scientific review must therefore examine the signal, the comparison population, the analyst's judgment process, and the communication to legal decision makers separately. Treating an expert's credentials or a discipline's popular reputation as a substitute for source-specific validation obscures the evidence actually available in the case National Research Council (2009) Garrett & Neufeld (2009).

Errors in reports are themselves heterogeneous. Morgan's 2023 analysis examined 732 exoneration cases marked by the National Registry of Exonerations as involving false or misleading forensic evidence and coded 1,391 forensic examinations. The paper distinguishes laboratory classification errors from misleading testimony, inadequate qualification, failures of communication, organizational weaknesses, and decisions by actors outside the laboratory. The sample consists of already identified exonerees associated with a registry classification and cannot yield the probability of error for a randomly selected forensic examination. Its significance is demonstrating that the term forensic error spans the laboratory, investigators, counsel, and courtroom, requiring quality assurance across boundaries Morgan (2023).

Contextual Bias in Forensic Examination

Expertise improves performance in validated tasks but operates through human perception, attention, interpretation, and decision. Case details that are irrelevant to the scientific comparison can still alter expectations. Dror, Charlton, and Péron's 2006 study revisited fingerprint impressions previously assessed by examiners, supplying new contextual information implying a nonmatch. In the experimental setting, many experts produced decisions inconsistent with their earlier identifications. This work demonstrated vulnerability to context under the tested conditions. It did not imply that every fingerprint match is wrong or that professional skill has no informational value Dror et al. (2006).

Contextual bias can influence what an examiner first notices, which features are recorded, what alternative hypothesis is considered, and when the examiner decides that an observed degree of correspondence is sufficient. These stages matter because an analyst may subsequently rationalize an initial feature selection as a purely data-driven outcome. Dror's 2020 analysis characterizes several common misconceptions about bias and identifies sources extending from the specific evidence and reference materials to organizational, personal, and environmental factors. Bias controls should be evaluated by the task relevance of information and by their effects on decision reproducibility, not by demanding that an examiner be psychologically unaffected by all context Dror (2020).

Task-irrelevant contextual information includes a suspect's confession when the analyst's assignment is to compare two fingerprint patterns, or police assertions about guilt when the analyst is evaluating an ambiguous trace independently. Some contextual information is necessary to select valid analytical methods, understand collection conditions, or interpret a sample's provenance. The scientific challenge is to distinguish necessary contextual facts from guilt narratives that can preload an answer. A documented sequential-unmasking procedure can preserve the analyst's observations and preliminary decision before potentially biasing contextual information becomes available, and quality assurance can record the justification for any later revision Kassin et al. (2013) Dror (2020).

The danger becomes cumulative when an investigator communicates an anticipated conclusion to the examiner, receives a supportive result, and subsequently treats that result as confirmation of the very hypothesis used to shape the examination. This feedback loop is especially concerning when results are described categorically and no contemporaneous record preserves the analyst's independent impressions. Research on forensic confirmation bias therefore belongs at the interface between laboratory psychology, information governance, and legal proof. Its prevention cannot be reduced to professional declarations of impartiality Kassin et al. (2013) Kukucka & Kassin (2014) Morgan (2023).

Invalid Testimony, Reporting Practices, and Inferential Inflation

Evidence enters court through human language. A laboratory report may contain an appropriately restricted comparison, while an oral presentation translates that conclusion into certainty beyond the method's demonstrated capacity. Garrett and Neufeld's 2009 transcript study found invalid testimony in 82 of 137 located trials of later DNA exonerees with forensic analyst testimony. The study identified statements that misstated empirical data or advanced inferences unsupported by the underlying discipline. Its denominator is explicitly the selected set of transcripts available for review, and the category of invalid testimony should be distinguished from an intentional falsehood or from a demonstrated laboratory misidentification Garrett & Neufeld (2009).

Misleading communication can involve overstatement of a source attribution, omission of uncertainty, selective mention of confirming results, failure to disclose limitations, or failure to explain that a presumptive analytical test requires confirmation. Morgan's analysis makes this courtroom pathway visible as part of the broader evidentiary error system. A defensible report should articulate the question tested, the material examined, the method and its validation, quality limitations, alternative outcomes, and the relation between a technical conclusion and the legal hypothesis. Where a probability or error rate is quoted, its underlying population, task, and study design should be reported accurately Morgan (2023) National Research Council (2009).

A court may hear several experts describing different observations from a shared laboratory process or common case file. This appearance of multiplicity does not establish independent corroboration if the interpretations relied on the same prior suspect attribution or one another's conclusions. It is essential to distinguish independent replication of a measurement from peer review of the same analysis, and both from a second expert who knew the first expert's answer. Independent review is strongest when analysts receive the necessary source materials, work blind to prior conclusions when feasible, record their observations before comparison, and document disagreements instead of resolving them through informal consensus Dror (2020) Kassin et al. (2013).

Discipline-Specific Lessons: Biological Evidence and Pattern Comparison

DNA evidence has transformed criminal investigation by enabling highly informative comparisons when samples, methods, interpretation, and databases are appropriate. Its power also depends on contamination controls, chain of custody, mixture handling, assumptions about contributors, and distinguishing a source association from the circumstances of transfer. DNA exonerations demonstrate the potential of independent biological evidence to overturn erroneous fact patterns, but the same exonerations reveal systematic differences in which convictions can later be challenged. The scientific lesson is that probative force is always tied to a specified proposition and method rather than to the generic prestige of a technique Hampikian et al. (2011) National Research Council (2009).

Other pattern-comparison disciplines have required extensive evaluation of their empirical foundations. The National Institute of Standards and Technology's 2023 scientific foundation review of bitemark analysis examined the premises needed to compare a mark in human skin with an individual's teeth. Its conclusions about insufficient empirical support illustrate the necessity of validating individualization claims, transfer conditions, and examiner performance before introducing strong courtroom attributions. This diagnosis is discipline-specific. An identified weakness in bitemark analysis must not be carelessly used as a numerical error estimate for fingerprints, DNA, or another method with different data and validation Sauerwein et al. (2023).

The comparison between disciplines is therefore a comparison of questions: Is the tested proposition identification, exclusion, association, or mere consistency? What alternative contributors could produce the observation? How often do examiners reach the same decision across repeated tests? How often does the method make mistaken identifications in ground-truth experiments? What uncertainty survives the analysis? Which contextual details were available when the conclusion was formed? These questions help distinguish scientific evidence quality from the persuasiveness of a categorical courtroom statement National Research Council (2009) Sauerwein et al. (2023) Morgan (2023).

Preservation, Chain of Knowledge, and Reanalysis

A modern forensic record must preserve more than the physical sample. It should preserve the analytical sequence: submission information, initial case hypothesis, contextual disclosures, examinations undertaken, reference material supplied, intermediate results, interpretations, revisions, and final communication. Without that sequence, later reviewers may be unable to determine whether apparent consistency between a confession, comparison result, and witness account arose independently. Evidence preservation is consequently a problem of informational provenance as well as physical chain of custody Kassin et al. (2013) Morgan (2023).

Reanalysis can improve reliability when it reexamines original material under methods with known performance and minimizes awareness of prior contested conclusions. It can also reproduce earlier errors if later experts begin with a strongly framed police narrative, receive only selected photographs, or are asked to verify an existing opinion rather than answer an independent question. Reproducibility should distinguish consistency of a result under repeated analysis from accuracy relative to ground truth. A series of agreeing analysts who share the same incorrect contextual information does not provide the same inferential evidence as genuinely independent measurements Dror et al. (2006) Dror (2020) National Research Council (2009).

Institutional reforms require measurable implementation criteria. A written promise of blind review is weak if intake forms routinely disclose confessions. A validated reporting template is insufficient if courtroom language departs from the report. Audits can instead examine how much task-irrelevant context analysts actually receive, the frequency of materially revised interpretations, documentation of uncertainty, and whether independent reviewers reproduce findings without access to predecessor conclusions. These process measures connect laboratory governance to the cumulative evidence mechanisms investigated in this dossier Kassin et al. (2013) Morgan (2023).

Tunnel Vision, Confirmation, and Investigative Selection

Investigative tunnel vision describes the process by which an emerging theory of guilt begins to govern which evidence is sought, which discrepancies are pursued, and how ambiguous observations are interpreted. A working hypothesis is necessary to organize investigation; the risk arises when provisional belief becomes a selection rule that excludes meaningful alternatives. Findley and Scott's 2006 analysis connected tunnel vision to eyewitness mistakes, confessions, forensic evidence, police and prosecutorial decisions, and defense failures. Their account identifies an institutional mechanism: separate actors can repeatedly encounter the same suspect-centered narrative and mistake agreement for independent confirmation Findley & Scott (2006).

Selection effects precede many visible evidentiary errors. Once one person becomes the principal suspect, detectives may devote interviews, testing budgets, surveillance, and witness attention disproportionately to that person. A witness may repeatedly view the same suspect's image; an analyst may be sent reference material from only that suspect; an alternative lead may be classified as unpromising before a comparable test. The resulting case file can appear richly inculpatory because the investigation generated more opportunities to observe inculpatory evidence for one hypothesis. This is a form of missing-not-at-random evidence collection, and it complicates any retrospective attempt to interpret the apparent balance of the file as a neutral sample of the relevant world Findley & Scott (2006) Garrett (2011).

Adversarial pressure can amplify confirmation. Prosecutors prepare theories of admissibility and proof, defense counsel respond under time and resource constraints, and experts may receive questions framed as requests to support a preselected proposition. A scientific audit should reconstruct every decision to collect or omit evidence and the operational alternatives available at that moment. Documentation of rejected leads, exculpatory results, changes of suspect, and reasons for abandoning hypotheses can reveal whether inquiry remained open to disconfirmation. The goal is disciplined hypothesis testing with an auditable record of selection, not a demand that investigators suspend judgment indefinitely National Research Council (2009) Morgan (2023).

Informants, Social Incentives, and Testimonial Dependence

Testimonial evidence may come from eyewitnesses, co-defendants, jailhouse informants, acquaintances, and officials. Their incentives, information access, and communication pathways differ. An informant may know facts through firsthand experience, public media, conversations with investigators, legal documents, or interactions with a detained suspect. A witness's ability to recount crime-specific facts cannot by itself establish direct access when those facts circulated through a case network. Source reconstruction therefore requires an inventory of opportunities to learn the information before it was reportedly volunteered National Registry of Exonerations (2026) Kassin et al. (2012).

Incentives to obtain favorable treatment, reduce personal exposure, or satisfy authority can shape testimony. That possibility does not entail that any particular witness is dishonest; the inference depends on the credibility of the account, documented exchanges, prior inconsistencies, and whether details were independently verified before disclosure. The evidentiary problem becomes cumulative when an informant repeats a theory previously communicated by an interrogator and the repeated story is subsequently used to corroborate an alleged confession. Separate persons have then produced statements with a shared informational ancestor. An audit should preserve contact records, benefits offered or understood, timing of reports, and access to case narratives National Registry of Exonerations (2026) Findley & Scott (2006).

Registry categories such as perjury or false accusation require careful interpretation. The National Registry's 2025 annual report coded perjury or false accusation in 71 of the 97 exonerations occurring that year. The category encompasses a defined range of knowingly false or accusatory statements according to registry coding and should not be treated as a direct experimental measure of memory error. It can co-occur with official misconduct, false confessions, and flawed forensic evidence. To study psychological mechanisms, the researcher must distinguish deliberate deception, mistaken belief, information contamination, strategic compliance, and institutional transmission rather than assigning a single mental process to every registry-coded false statement National Registry of Exonerations (2026).

Plea Decisions, Counsel, and the Hidden Conviction Pathway

Much of the criminal process is resolved before a contested jury verdict. Plea decisions are made in a setting structured by detention, expected sentencing differences, counsel's advice, perceived chance of acquittal, uncertainty about evidence, and the defendant's tolerance of risk. An innocent defendant can rationally prefer a lower guaranteed punishment to a possible severe outcome under conditions of perceived uncertainty. A plea therefore cannot be taken as an unambiguous behavioral indicator of factual guilt. The 2025 confession scientific review discusses increased willingness to plead guilty following false confessions and the institutional consequences of that choice Kassin et al. (2025).

Psychological analysis needs to measure the decision as a choice among constrained options. Relevant variables include anticipated loss, time pressure, decisional capacity, knowledge of disclosed evidence, trust in counsel, access to discovery, detention costs, and how evidence strength is communicated. A laboratory experiment on hypothetical plea offers can test selected behavioral mechanisms, but it cannot reproduce every feature of real detention and legal consequences. Observational comparisons of pleas and trials are also highly selected: cases reaching trial differ in charge, severity, available evidence, resources, and defendant expectations. Research should identify these distinctions before attributing an observed plea distribution to any single bias Kassin et al. (2025) Gross et al. (2014).

Inadequate defense can have a dual role in cumulative error. Missing investigative resources may allow contradictory forensic or testimonial claims to remain untested, while failure to preserve evidence can limit future opportunities for exoneration. The Registry's 2025 report coded inadequate legal defense in 32 of its 97 exonerations, but this count is not a population prevalence estimate of deficient representation. A credible causal account would require comparison of similarly situated cases, detailed records of counsel decisions, and a distinction between an outcome that appears erroneous after exoneration and representation that was deficient given information available at the time National Registry of Exonerations (2026) Garrett (2011).

Juror Reasoning, Evidence Weighting, and Statistical Misinterpretation

Jurors and judges must transform heterogeneous evidence into a conclusion about legally defined propositions. They encounter narratives, eyewitness testimony, expert estimates, credibility judgments, and procedural instructions that are difficult to integrate. A categorical source assertion may be easier to remember than the underlying uncertainty, while multiple convergent witnesses may appear persuasive even if all drew from the same initial disclosure. Kassin's 2012 account of confession evidence calls this apparent increase in corroboration a form of corroboration inflation, with consequences for how legal decision makers assess the independence and cumulative weight of evidence Kassin (2012).

A probabilistic approach clarifies the difficulty. Suppose a witness reports identifying the suspect after learning of a purported confession. The probability of that identification conditional on a false confession is different from the probability in a genuinely independent identification procedure. Multiplying likelihood ratios from the confession and identification as though the observations were independent risks overstating the joint evidence. The correct joint likelihood must account for dependence within the relevant guilt and innocence hypotheses. The practical problem is that court records often lack the information needed to estimate this dependence. Acknowledging it is more scientifically justified than supplying spurious precision Hasel & Kassin (2009) Kassin et al. (2012) Thompson & Newman (2015).

Thompson and Newman's 2015 experiment tested lay comprehension of random-match probabilities, likelihood ratios, and verbal equivalents for DNA and shoeprint evidence, finding that verdict sensitivity depended on evidence type and presentation. The study also identified common misinterpretations of forensic statistics. These findings demonstrate that format and prior beliefs affect translation from expert statement to verdict judgment. They do not imply that one numerical expression will always make every fact finder optimally Bayesian; communication should be evaluated against comprehension, the particular scientific question, and the danger of confusing probability of evidence with probability of guilt Thompson & Newman (2015).

Corroboration should be described by source and mechanism, not by a simple count of nominally distinct exhibits. Two independently obtained observations of different aspects of an event may carry substantial combined information. The same two observations may contribute far less if the second was revised after the observer heard the first. Conversely, two observations sharing some context need not be wholly redundant: conditional statistical dependence is a matter of degree and proposition. This graduated view is essential to avoid both blind multiplication of evidence weight and the opposite error of discarding informative but partially related evidence Thompson & Newman (2015) Hasel & Kassin (2009).

Harmless Error, Postconviction Review, and Institutional Learning

Legal review may evaluate the effect of an error by asking whether the remainder of the evidence could sustain a conviction. Psychological research on contamination complicates that assessment whenever the remaining evidence was influenced by the excluded item. If a false confession prompted a witness to revise an identification or influenced an expert's interpretation, removing the confession from a retrospective narrative does not restore the original independence of the other evidence. Kassin's 2012 review and the case-file analysis by Kassin, Bogart, and Kerner explicitly connect this problem to the legal concept of harmless error, while leaving the application of doctrine to the relevant legal system Kassin (2012) Kassin et al. (2012).

Postconviction reconstruction is often asymmetric: inculpatory information and courtroom testimony survive in official records, whereas unrecorded initial uncertainty, abandoned leads, and interactions before a formal confession may be irretrievable. The apparent completeness of a trial transcript should not be confused with completeness of the investigative history. An effective review protocol should compare the final evidentiary narrative with earlier timestamps, disclosure files, laboratory notes, interviews, and defense-investigation records. The strongest questions are often causal: what would the second evidence producer have concluded without exposure to the first claim, and which observations remain genuinely independent? Garrett (2011) Morgan (2023).

System learning requires tracking correction as well as conviction. A reversed judgment may reveal a contaminated process, a newly available scientific method, previously withheld material, or a revised understanding of a discipline's validity. Different pathways suggest different remedies. Exoneration databases can guide investigation of recurrent configurations, while independent audits of ordinary, non-exonerated cases are needed to estimate how often the same processes occur without discovery of innocence. A complete quality system investigates near misses, protected reporting of concerns, nonconfirming forensic results, and the fidelity of remedial procedures National Registry of Exonerations (2026) National Research Council (2009).

Research Designs for a Multistage Error Process

The scientific study of wrongful convictions uses complementary research designs because the complete causal process is difficult to manipulate and because ground truth in ordinary criminal cases is often inaccessible. Laboratory experiments can independently vary confession information, lineup feedback, contextual instructions, or framing of forensic evidence while observing changes in judgments. Archival studies reconstruct known exonerations through records and interviews. Prospective audits assess procedure implementation in ordinary cases; longitudinal designs document how assessments change after information disclosure; and statistical models represent uncertainty and dependence. Each design has its own inferential target and characteristic selection problems Hasel & Kassin (2009) Dror et al. (2006) Morgan (2023).

Experimental identification of a psychological mechanism depends on assignment, manipulation validity, outcome measurement, and an appropriate comparison condition. Hasel and Kassin's confession-information manipulation demonstrates that such information can alter subsequent identification, while Kukucka and Kassin's studies show an effect on handwriting judgments. Generalizing from participants, tasks, and stakes in the laboratory to criminal justice professionals and real consequences requires further evidence. Experiments with expert practitioners increase ecological relevance for particular forensic tasks, although expertise, recruitment selection, and repeated-stimulus exposure still constrain inference. Studies should report exclusion rules, demographic composition, preregistered outcomes where available, effect-size uncertainty, and whether the tested information was genuinely task irrelevant Hasel & Kassin (2009) Kukucka & Kassin (2014) Dror et al. (2006).

Archival exoneration research observes legally consequential outcomes and real institutional behavior, but its cases were selected through a multistage discovery process. Every variable depends on whether the necessary records survive, whether a researcher can obtain them, and how ambiguous evidence is coded. Causal sequences may be incompletely reconstructed from post hoc testimony. An archival finding that confessions typically preceded other mistakes is consistent with contamination but could also reflect an investigative focus that generated both the confession and later errors. Stronger studies combine temporal reconstruction with independent documentation of information exchange, sensitivity analyses for missing observations, and transparent coding definitions Kassin et al. (2012) Garrett & Neufeld (2009) Morgan (2023).

Measurement, Coding Reliability, and Construct Validity

An evidentiary error taxonomy must distinguish observed events from inferred mental states. A recorded investigator disclosure is an observed procedural event. An examiner's later changed conclusion is an observed decision. A claim that confirmation bias caused the change requires additional comparisons or process evidence. Likewise, an incorrect identification can be established when ground truth is known, while memory contamination describes a proposed mechanism that may or may not have caused it. Research protocols should create separate fields for event, evidence basis, causal inference, legal classification, and confidence of the researcher in each code National Research Council (2014) Morgan (2023).

Intercoder reliability matters because registry labels and narrative case files may use broad categories that admit different interpretations. Two trained coders can independently identify instances of nondisclosure, testimony overstatement, confession disclosure, or witness feedback. Disagreement should be documented and adjudicated by declared rules; adjudicated consensus alone does not reveal how reliably the coding procedure would perform with new cases. Denominators must identify whether the count refers to defendants, convictions, episodes, reports, forensic examinations, or exoneration records. Morgan's 2023 paper illustrates the difference by coding 732 cases and 1,391 examinations as distinct analytical units Morgan (2023).

Validation requires comparison of a proposed error category with independent records and outcomes. A registry coding of misleading forensic evidence should not be automatically equated with a laboratory misclassification when the underlying issue may have been the testimony, an external expert, or mishandled disclosure. A category such as official misconduct can reflect several behaviors in one case; counting categories then describes overlapping conditions rather than a partition. If research aggregates them, the aggregation rule should be specified and cross-classification retained. Construct clarity is crucial because different reforms target different failure mechanisms National Registry of Exonerations (2026) Morgan (2023).

Conditional Evidence, Bayesian Networks, and Double Counting

Statistical evidence accumulation depends on the joint probability of the observed information under the propositions being compared. When evidence items are conditionally independent, their likelihood contributions can be multiplied under appropriate assumptions. When an eyewitness has changed an identification after hearing of a confession, that independence condition may fail. A more appropriate representation connects the confession report to subsequent witness information, while allowing the facts of the case and the witness's original perception to affect the final response. This is a conceptual causal graph until its parameters are established from data. Treating its arrows as measured causal effects would exceed the evidence Hasel & Kassin (2009) Kassin (2012).

Correlation alone is insufficient to determine the direction of contamination. Two witnesses may agree because both saw the same perpetrator, because one learned the other's account, because both heard the same police description, or because the actual event had highly distinctive features. Distinguishing these explanations requires observations about communication, timing, exposure, and independently verifiable details. The strongest model keeps separate nodes for the latent event, original observation, recorded account, later information, judgment, and courtroom presentation. Without these distinctions, a single network diagram can disguise fundamental uncertainty Kassin et al. (2013) National Research Council (2014).

A simple numerical illustration exposes why dependence matters. Suppose a hypothetical witness report is strongly related to a prior confession claim rather than independently generated. A legal analyst who multiplies the nominal evidence strengths of both reports counts part of one information stream twice. The size of that overstatement depends on the conditional relationships, which must be measured rather than guessed. The illustration is about probability structure and makes no estimate for any real criminal case. Juror research on likelihood ratios and random-match probabilities adds a second difficulty: even correctly computed numbers can be misunderstood when translated into the chance that a defendant is guilty Thompson & Newman (2015).

The fact that claims are dependent does not determine which particular suspect is innocent or guilty. It changes the evidentiary weight and the need for scrutiny. Some dependent observations still contain unique information; complete redundancy is only one possible configuration. Conversely, an expert opinion may depend on a confession through subtle contextual exposure, making the statistical structure difficult to infer from final reports. The general methodological recommendation is to preserve informational ancestry and to distinguish observable paths of influence from merely hypothesized dependence Dror (2020) Kassin et al. (2013).

Generalizability, Selection, and the Limits of Case-Series Evidence

Known exonerations differ sharply from all convictions on access to new evidence, public attention, offense type, availability of legal representation, institutional review opportunities, and time since conviction. DNA exoneration samples add strong selection for crimes with testable biological traces and preserved samples. Researchers who divide erroneous identification cases by the number of DNA exonerees are measuring prevalence within that selected exoneration population. They are not estimating the population probability of mistaken identification among all cases adjudicated during the same years. Hampikian and colleagues' analysis and the 2025 National Registry report are informative when their sample definitions accompany every reported percentage Hampikian et al. (2011) National Registry of Exonerations (2026).

Survival-analysis approaches can partly address differential time at risk of exoneration but depend on assumptions about detection and case trajectories. The 4.1 percent estimate in Gross and colleagues' death-sentence analysis is explicitly tied to a hypothetical continued exposure to death-row exoneration processes. Its interpretation depends on the unusual visibility, stakes, and procedural history of capital cases. It should not be quoted as an estimated wrongful-conviction rate for misdemeanors, pleas, ordinary imprisonment, or any other national system Gross et al. (2014).

International research faces additional comparability problems. Countries differ in investigative interviewing, jury participation, legal aid, confession admissibility, scientific laboratory governance, appellate remedies, and public access to case records. A U.S. exoneration category may not have a direct analogue in another system. Nevertheless, psychological mechanisms involving memory, expectation, social influence, and information dependence can be studied across legal contexts by specifying equivalent experimental tasks and measurements. Cross-jurisdictional generalization should be earned through comparative evidence rather than assumed from a single national registry National Research Council (2014) Dror (2020).

Converging Evidence, Counterevidence, and the Strength of Inference

The strongest support for a general mechanism of cumulative evidentiary error comes from triangulation across methods: experimental manipulations showing that an alleged confession can alter a witness's choice, expert-judgment studies showing contextual effects on forensic interpretations, and archival observations in which multiple documented errors co-occur. Each answers a distinct question. The experimental literature provides causal evidence for particular information treatments; archival literature demonstrates historically consequential configurations; institutional scholarship identifies plausible organizational pathways. The combination supports a scientifically serious risk model while leaving the causal weight of each mechanism uncertain for any individual case Hasel & Kassin (2009) Dror et al. (2006) Kassin et al. (2012) Findley & Scott (2006).

Counterevidence and benign explanations deserve explicit consideration. Agreement among independent witnesses can reflect genuine corroboration, and separate validated forensic measurements can supply powerful additional information. Expert awareness of particular contextual facts can be necessary to interpret sampling or collection conditions. An early investigative focus may lead to efficient discovery of genuinely incriminating evidence. The research hypothesis is therefore conditional: contamination is possible when task-irrelevant or leading information changes a later judgment, and its magnitude varies with procedure, evidence strength, task, and participant. This precise proposition can be tested without presuming every consistent case narrative is spurious National Research Council (2009) Wells et al. (2020) Dror (2020).

Many proposed reforms have compelling mechanistic foundations but limited direct evidence that their adoption reduces wrongful convictions at the population level. A practice may reliably prevent postidentification confidence inflation in experiments while its effect on trial convictions remains difficult to isolate. Laboratory blinding may reduce contextual influence without resolving systematic weaknesses of the examination method itself. Improvements in legal disclosure may strengthen review even if official registry numbers increase because more errors become discoverable. Evaluations should distinguish process compliance, intermediate psychological outcomes, forensic error rates, wrongful conviction, and exoneration frequency Wells et al. (2020) National Research Council (2009) National Registry of Exonerations (2026).

Evidence synthesis should also resist retrospective inevitability. Once exoneration is known, an ambiguous old clue may appear obviously flawed, and decisions made under incomplete contemporaneous information may be judged with knowledge unavailable at the time. At the same time, hindsight bias must not excuse documented omission or contamination. A balanced analysis reconstructs what each decision maker knew, what alternatives were reasonably available, and how information paths shaped later observations. This perspective is both psychologically grounded and methodologically essential for fair classification of a causal error chain Findley & Scott (2006) Morgan (2023).

Prevention at the Earliest Investigative Stages

Preventing cumulative evidentiary error is most efficient before a weak initial inference becomes embedded in several subsequent records. An investigation can preserve an initial witness statement before suspect exposure, document alternative hypotheses, establish time-stamped records of communication, and record how each line of evidence was selected. The objective is to make the history of inference reconstructable. A written note identifying a suspect is useful only if it is accompanied by the observation, its source, the stage of investigation, and the alternatives considered. The procedural emphasis is supported by the National Research Council's recommendations on eyewitness evidence and by analyses of tunnel vision in criminal investigation National Research Council (2014) Findley & Scott (2006).

Investigative interviewing should be designed to acquire information with the least possible contamination. Open questions, explicit separation of volunteered detail from interviewer-supplied information, preservation of the first complete account, and clear records of repeated interviews provide a basis for assessing later changes. The 2025 review by Kassin and colleagues recommends science-based interviewing, full recording, protections for vulnerable suspects, and constraints on tactics associated with false confession. Whether a specific jurisdiction adopts each practice is a legal and institutional question; the scientific rationale is preserving the interpretability of a suspect's statement and avoiding information transfer into supposed guilty knowledge Kassin et al. (2025).

Documentation should include negative results. If a person is initially suspected and then excluded by independently generated evidence, the excluded result should remain accessible rather than vanish from a streamlined prosecution narrative. Investigators can audit whether comparable alternatives were pursued and whether disconfirming findings reached the relevant actors. A system that rewards only the accumulation of prosecution-supporting items creates pressure for selection and information asymmetry. Treating exculpatory findings as first-class research records makes the eventual assessment more complete and reduces the risk that an apparent abundance of inculpatory evidence merely reflects investigative attention Findley & Scott (2006) Garrett (2011).

Safeguarding Witness Evidence and Suspect Interviews

Lineup administration can be standardized to minimize information supplied by the administrator and to preserve the initial decision before feedback. The 2020 American Psychology-Law Society scientific review synthesized recommendations on lineup construction, instructions, administrator blindness, documentation, and statements of confidence. These elements should be implemented as linked procedures: a nominally double-blind lineup loses much of its protective value if a witness has already been shown the intended suspect through other channels. An audit should test the whole witness exposure history, not merely the presence of a checked procedural box Wells et al. (2020).

Confession safeguards must address both production and subsequent dissemination. Complete audiovisual recording permits evaluation of the account's origin; evidence-based suspicion requirements can constrain unjustified escalation; protection for youth and vulnerable adults can improve informed participation. The 2025 scientific review additionally recommends shielding other witnesses and forensic examiners from confession content when their tasks require independent evaluation. This last measure addresses the cumulative mechanism directly: preventing the first error from changing later evidence may reduce harm even in cases where the initial mistake has already occurred Kassin et al. (2025) Hasel & Kassin (2009).

Postprocedure feedback should be recorded and disclosed. If a witness later changes an identification after hearing of an arrest or confession, the original decision remains scientifically relevant. A later increased level of certainty should be interpreted as a new response produced under altered information conditions. Courts and investigators can then evaluate the event sequence and the degree of independence rather than treating the last narrative as a timeless account. The benefit of preservation is methodological even when the final identification proves correct, because it distinguishes genuine initial memory from later social influence Wells & Bradfield (1998) National Research Council (2014).

Laboratory Context Management and Independent Verification

Context management begins by defining the actual forensic question. An analyst testing a chemical sample may need to know its substrate, collection treatment, and possible contamination, but usually need not know that a suspect confessed. A laboratory should specify which contextual facts are analytically necessary and which are irrelevant to the determination. Where feasible, analysts can record observations and preliminary conclusions before exposure to potentially biasing information; later changes can be documented with the new information that prompted them. This creates a replicable trail without excluding scientifically necessary context Kassin et al. (2013) Dror (2020).

Blind or sequentially unmasked review is an implementation strategy whose value depends on adherence and the scientific quality of the underlying task. If a method is inadequately validated, isolation from contextual bias will not transform it into a reliable identification technique. Conversely, strong validation cannot guarantee impartial application in a highly suggestive case environment. The 2009 National Research Council report and NIST's 2023 assessment of bitemark analysis illustrate why method validity and case-level human factors must be evaluated together while remaining analytically distinct National Research Council (2009) Sauerwein et al. (2023).

Independent verification should specify what independence means. A second analyst who reads the first analyst's categorical opinion before examining the evidence has been exposed to a potent contextual cue. A stronger verification design can provide the same raw materials and task-relevant parameters while withholding predecessor conclusions until a documented preliminary assessment is complete. Discrepancies should be preserved, resolved through declared procedures, and available to subsequent case review. This arrangement addresses both accuracy and the transparent provenance of agreement Dror et al. (2006) Morgan (2023).

Courtroom Communication and the Audit of Corroboration

Expert testimony must preserve the limits of the underlying observation. The court needs to know the level of proposition addressed, the method and its validation, the relevant uncertainty, and whether contextual exposure altered interpretation. Training jurors to understand these distinctions is useful but cannot substitute for accurate testimony. Thompson and Newman's 2015 experimental results show that lay judgments depend in part on the discipline and the numerical presentation of evidentiary strength. Information communication can therefore be studied empirically through comprehension, decision changes, and error detection rather than assumed effective because a likelihood ratio or cautionary sentence appears in a report Thompson & Newman (2015).

A corroboration audit should identify each evidence item, its time of origin, the information available to its producer, and any route by which it influenced or was influenced by other items. The audit must then distinguish independent confirmation from repetition, shared-source accounts, and retrospectively amplified confidence. Such an audit is especially important when a purported confession was acquired before witness or laboratory judgments, since controlled experiments demonstrate plausible channels of influence. It should also look for truly independent exculpatory observations, including alternative identifications, trace exclusions, and alibi information that received less investigative attention Hasel & Kassin (2009) Kassin et al. (2012).

Adjudication requires a complete record rather than a simple score of independent-looking items. A court can receive two accurate observations, an inaccurate but independent inference, and a second inference distorted by the first. The explanatory challenge lies in tracing causal and informational dependencies. In some cases, the strongest contribution of an audit may be exposing how little is known about the original procedure. This uncertainty should be communicated as part of evidence quality, rather than filled with unsupported assumptions of procedural regularity or improvised numerical adjustments National Research Council (2014) Morgan (2023).

Evaluating Safeguards and Their Unintended Consequences

Reforms should be evaluated through a chain of outcomes. First, measure whether personnel followed the prescribed procedure. Second, measure whether the procedure improved a proximal construct such as preservation of initial confidence, independent forensic classification, or completeness of recording. Third, examine broader error indicators and decisions. Finally, where ethically and practically possible, monitor outcomes such as reversals, verified false allegations, and independently investigated near misses. A decline in recorded exonerations is an ambiguous measure of success because it may reflect reduced wrongful convictions, diminished access to review, or altered recognition of mistakes National Registry of Exonerations (2026) Wells et al. (2020).

Randomized or stepped implementation studies may be feasible for some interviewing instructions, blinding practices, report formats, or training interventions, provided they avoid exposing people to unnecessary legal risk. Difference-in-differences and interrupted time-series designs can evaluate system-wide procedural changes but require explicit assessment of changing case composition, concurrent reforms, data completeness, and preexisting trends. Independent audits of routine cases are especially valuable because they sample processes outside the highly selected population of exonerations. No single design resolves all identification problems, so protocols should be registered and outcomes transparently reported Morgan (2023) Dror (2020).

Potential unintended consequences also deserve study. A narrowly designed context shield could prevent an examiner from learning necessary sample information. More extensive recording may improve transparency while creating new privacy and security obligations. A rigid numerical reporting rule may replace one source of misunderstanding with another. Reform evaluation should use both scientific validity and implementation quality as criteria, and should treat informational access as a controlled variable whose benefits and risks vary by task National Research Council (2009) Thompson & Newman (2015).

Artificial Systems, Digital Evidence, and New Channels of Contamination

The Artificial Era broadens the network through which case evidence is generated, selected, summarized, and circulated. Investigative records increasingly include digital traces, machine-generated comparisons, automated transcription, search-ranked material, and computationally prepared case summaries. Their forensic relevance depends on the origin of the input data, the system's validated task, the error characteristics of the output, and the degree to which subsequent human decisions rely upon it. A digital system may pass along an earlier factual error with extraordinary consistency, making repetition appear to be confirmation. Equally, carefully validated tools may expose discrepancies and preserve an auditable evidence trail. These are research hypotheses about specific human–artificial workflows and require direct empirical assessment rather than general claims that automation necessarily improves or diminishes justice Hargreaves et al. (2024) National Research Council (2009).

Digital forensic examination extends conventional questions of chain of custody to data acquisition, software version, extraction settings, logs, access permissions, and transformations between representations. Hargreaves, Nelson, and Casey proposed an abstract model for digital forensic tools intended to support systematic error mitigation. The value of such a framework is methodological: it directs attention to potential errors in the tool's inputs, internal operations, human interpretation, and communicated conclusions. For the present dossier, digital provenance should record whether a later statement was composed from an original source, machine-produced transcription, paraphrase, or secondary investigative summary, and whether each stage was verified against the underlying material Hargreaves et al. (2024).

Algorithmic assistance also poses distinctive questions about conditional dependence. If an investigator, an examiner, and a prosecutor all consult the same automated summary, their subsequent agreement can reflect a shared computational source. If a machine-generated report quotes an earlier witness statement and a later witness is shown that report, an information loop can emerge even when no participant intends to manipulate the record. Measuring these pathways requires logging which people and systems accessed which versions of information before making their decisions. The appropriate unit is an information lineage with both human and artificial processing stages, rather than a tally of how many outputs repeat the same assertion Kassin et al. (2013) Hargreaves et al. (2024).

Machine-assisted screening can be evaluated through ground-truth datasets, predeclared operating points, subgroup performance, distribution shift, human oversight, and prospective decisions under realistic workloads. NIST's ongoing Face Recognition Technology Evaluation is an example of a domain in which algorithm-specific performance and demographic differentials are empirically assessed. Those benchmark results should never be transposed into an error estimate for a different tool, jurisdiction, or task. Especially in high-stakes criminal investigation, a benchmark score is a property of a defined evaluation, not a free-standing probability that a given suspect is guilty NIST (2026) Morgan (2023).

Institutional Transparency, Disclosure, and the Epistemic Record

Information provenance can be defined as the documented history of a claim from observation or generation through every material transformation, disclosure, and decision that relied upon it. For an eyewitness account, provenance includes first perception, initial interview, lineup exposure, feedback, public communication, and later testimony. For a forensic result, it includes original item, collection, analytical method, task-relevant context, preliminary interpretation, verification, and final report. For a confession, it includes the full course of questioning, disclosed facts, suspect statements, and subsequent dissemination to other evidence producers. These histories permit later reviewers to identify which elements were truly independent National Research Council (2014) Kassin et al. (2025) Morgan (2023).

Provenance should be usable, not merely voluminous. A digital case management system may contain thousands of documents but still conceal who read a misleading summary before making a crucial decision. A useful record distinguishes a source observation from a later annotation, captures access and revision history, and permits reconstruction of the information available at specified times. Confidentiality, disclosure rights, and evidence access are legal-system-dependent, so an implementation should document the legitimate scope of access and retain a way to audit contested transformations. The scientific objective is a complete causal history of evidentiary claims, subject to appropriate legal and privacy safeguards Findley & Scott (2006) Hargreaves et al. (2024).

Transparency also helps identify genuine counterevidence. If an analyst excluded the principal suspect but that exclusion was not disclosed to subsequent witnesses or fact finders, the evidentiary narrative has been filtered. If an investigator recorded a witness's early low confidence and a later statement gives greater certainty, a full history permits examination of what changed. These cases demonstrate why a trial presentation is a curated representation of a prior process. Institutional quality assurance should retain earlier versions, negative results, unresolved discrepancies, and the rationale for interpretive revisions Wells & Bradfield (1998) National Research Council (2009).

Research Priorities for Cumulative and Comparative Forensic Psychology

A first priority is to estimate the frequency and magnitude of cross-source contamination under conditions resembling actual investigation. Controlled studies can vary whether a witness hears of a confession, whether an examiner receives suspect-related context, and whether later reviewers see the original record or only a summary. Repeated assessments can reveal how the same information changes across procedural stages. Professionally realistic tasks should preserve ethical safeguards and report both process outcomes and task-grounded accuracy, with measures of effect heterogeneity across expertise, case complexity, and evidence ambiguity Hasel & Kassin (2009) Kukucka & Kassin (2014) Pena et al. (2024).

A second priority is the measurement of information dependence. Researchers can construct case-flow networks with timestamped nodes for source observations, investigative hypotheses, witness interviews, reports, machine-generated summaries, expert judgments, and court submissions. Comparing these networks to independent ground-truth assessments can identify which procedures permit contamination and which preserve separable informational contributions. Statistical methods can then distinguish correlations arising from common events from dependence created by subsequent communication. Network models should be openly specified, experimentally tested where feasible, and accompanied by sensitivity analysis for unobserved contacts and lost records Kassin et al. (2012) Thompson & Newman (2015).

A third priority is prospective evaluation of safeguards outside documented exonerations. Routine audits can sample ordinary forensic cases and record the rate at which administrators inadvertently provide feedback, analysts receive irrelevant suspect narratives, original interview materials are incomplete, or reports overstate validated conclusions. Studies should include cases that were dismissed, acquitted, resolved by plea, or never prosecuted, not only those later exonerated. This broad sampling would improve knowledge of the prevalence of error-prone conditions without asserting that every observed procedural lapse causes a wrongful conviction National Registry of Exonerations (2026) Morgan (2023).

A fourth priority is international comparison. Psychological studies can measure similar operations across different legal systems while preserving the particular meaning of local legal standards and procedures. Comparative research should distinguish lineup administration, confession recording, laboratory organization, disclosure obligations, and the availability of independent review. Cross-national replication of specific cognitive effects is valuable, but claims about ultimate conviction outcomes require accompanying institutional and legal evidence. A multi-jurisdictional research program could identify which protections operate robustly, which depend on local institutions, and where common scientific measurement standards are possible National Research Council (2014) Dror (2020).

A fifth priority follows the expansion of computational assistance: compare human-only, machine-assisted, and mixed evidence workflows with independent performance and provenance measures. Experiments should test whether algorithmic summaries improve factual recall or amplify an earlier false claim; whether human reviewers independently inspect source material; and how benchmark calibration changes across populations and tasks. The Artificial Era supplies new information architectures whose reliability must be established at their real points of use. Evaluation must distinguish the technical accuracy of a component from the accuracy and fairness of the whole investigative decision process Hargreaves et al. (2024) NIST (2026).

An Integrated Account of Cumulative Evidentiary Error

The evidence reviewed in this monograph supports a multistage account. Investigative selection determines what information is sought. Interview and identification procedures affect what witnesses and suspects report. Analysts work with scientific material while interpreting signals through expectations and contextual information. Testimony translates technical judgments into propositions that influence jurors and judges. At every transition, people and institutions may preserve an observation's provenance or replace it with a more confident narrative. Cumulative error develops when earlier mistakes become inputs to later evidentiary production and when their resulting agreement is interpreted as independent support Findley & Scott (2006) Kassin et al. (2012) Morgan (2023).

Controlled experiments establish identifiable components of this process. Confession information can alter a witness's lineup choice; contextual knowledge can change forensic judgments; confirming feedback can inflate later claims of eyewitness certainty. Exoneration research shows that multiple error categories occur together and that their discovery is strongly selected by access to review and independent evidence. These lines of research reinforce each other but retain their separate inferential limits. The general proposition that error can accumulate through information transfer is supported; the frequency of such accumulation across all criminal adjudication and its contribution to specific conviction rates remain unresolved empirical questions Hasel & Kassin (2009) Dror et al. (2006) Wells & Bradfield (1998) Gross et al. (2014).

Prevention follows from the same structure. Preserve the earliest uncontaminated evidence, record all substantive interviews, control irrelevant contextual exposure, validate forensic methods, disclose analytical limitations, document alternative hypotheses, and audit the informational independence of corroboration. Each measure targets a specific link in the potential error chain. Their collective effectiveness should be assessed through implementation fidelity, accurate intermediate measurements, independent audit, and case outcomes wherever ascertainable. Institutional accountability requires an explanation of how evidence was formed and combined, including the information available to every consequential decision maker National Research Council (2009) National Research Council (2014) Kassin et al. (2025).

Conclusion

Wrongful convictions expose a foundational problem in the psychology of evidence: the perceived quantity and consistency of proof can exceed the number of genuinely independent observations from which it derives. Eyewitness confidence may change after feedback, confession claims may alter later identifications, forensic opinions may reflect extraneous case knowledge, and legal judgments may treat these related products as corroboration. The relevant scientific question concerns the origin, causal history, and conditional information value of each claim. This focus unites cognitive, social, forensic, and institutional explanations without presuming that every evidentiary agreement is contaminated Wells & Bradfield (1998) Hasel & Kassin (2009) Kassin et al. (2013).

The strongest research integrates experimental mechanisms, validated scientific methods, complete procedural histories, and carefully delimited archival observations. Exoneration registries reveal important failures in cases that have been discovered, while their changing and selective denominators limit general estimates of wrongful conviction. Probabilistic evidence integration requires recognizing dependence, and effective communication demands preserving method-specific uncertainty. These distinctions make the difference between a scientifically supported inference and a legal narrative that has gained certainty through repetition Gross et al. (2014) Morgan (2023) Thompson & Newman (2015).

The central conclusion is that evidentiary reliability is a property of the entire information-producing process. It depends on perception, recording, questioning, laboratory analysis, institutional incentives, disclosure, interpretation, and adjudication together. As artificial systems become involved in preparing and transmitting evidence, the provenance of claims grows still more important. Psychology contributes rigorous methods for exposing dependencies, testing safeguards, and separating actual independent corroboration from the accumulated expression of an earlier mistake. A trustworthy justice system must be capable of reconstructing how it came to believe what it believes Hargreaves et al. (2024) National Research Council (2014) Findley & Scott (2006).

References

  • Dror, I. E. (2020). Cognitive and human factors in expert decision making: Six fallacies and the eight sources of bias. Analytical Chemistry, 92(12), 7998–8004. https://doi.org/10.1021/acs.analchem.0c00704
  • Dror, I. E., Charlton, D., & Péron, A. E. (2006). Contextual information renders experts vulnerable to making erroneous identifications. Forensic Science International, 156(1), 74–78. https://doi.org/10.1016/j.forsciint.2005.10.017
  • Findley, K. A., & Scott, M. S. (2006). The multiple dimensions of tunnel vision in criminal cases. Wisconsin Law Review, 2006(2), 291. https://repository.law.wisc.edu/s/uwlaw/item/30977
  • Garrett, B. L. (2011). Convicting the innocent: Where criminal prosecutions go wrong. Harvard University Press. https://doi.org/10.4159/harvard.9780674060982
  • Garrett, B. L., & Neufeld, P. J. (2009). Invalid forensic science testimony and wrongful convictions. Virginia Law Review, 95(1), 1–97. https://scholars.duke.edu/publication/1344926
  • Gross, S. R., O'Brien, B., Hu, C., & Kennedy, E. H. (2014). Rate of false conviction of criminal defendants who are sentenced to death. Proceedings of the National Academy of Sciences, 111(20), 7230–7235. https://doi.org/10.1073/pnas.1306417111
  • Hampikian, G., West, E., & Akselrod, O. (2011). The genetics of innocence: Analysis of 194 U.S. DNA exonerations. Annual Review of Genomics and Human Genetics, 12, 97–120. https://doi.org/10.1146/annurev-genom-082509-141715
  • Hargreaves, C., Nelson, A., & Casey, E. (2024). An abstract model for digital forensic analysis tools—A foundation for systematic error mitigation analysis. Forensic Science International: Digital Investigation, 48(Suppl.), 301679. https://doi.org/10.1016/j.fsidi.2023.301679
  • Hasel, L. E., & Kassin, S. M. (2009). On the presumption of evidentiary independence: Can confessions corrupt eyewitness identifications? Psychological Science, 20(1), 122–126. https://doi.org/10.1111/j.1467-9280.2008.02262.x
  • Kassin, S. M. (2012). Why confessions trump innocence. American Psychologist, 67(6), 431–445. https://doi.org/10.1037/a0028212
  • Kassin, S. M., Bogart, D., & Kerner, J. (2012). Confessions that corrupt: Evidence from the DNA exoneration case files. Psychological Science, 23(1), 41–45. https://doi.org/10.1177/0956797611422918
  • Kassin, S. M., Cleary, H. M. D., Gudjonsson, G. H., Leo, R. A., Meissner, C. A., Redlich, A. D., & Scherr, K. C. (2025). Police-induced confessions, 2.0: Risk factors and recommendations. Law and Human Behavior, 49(1), 7–53. https://doi.org/10.1037/lhb0000593
  • Kassin, S. M., Drizin, S. A., Grisso, T., Gudjonsson, G. H., Leo, R. A., & Redlich, A. D. (2010). Police-induced confessions: Risk factors and recommendations. Law and Human Behavior, 34(1), 3–38. https://doi.org/10.1007/s10979-009-9188-6
  • Kassin, S. M., Dror, I. E., & Kukucka, J. (2013). The forensic confirmation bias: Problems, perspectives, and proposed solutions. Journal of Applied Research in Memory and Cognition, 2(1), 42–52. https://doi.org/10.1016/j.jarmac.2013.01.001
  • Kukucka, J., & Kassin, S. M. (2014). Do confessions taint perceptions of handwriting evidence? An empirical test of the forensic confirmation bias. Law and Human Behavior, 38(3), 256–270. https://doi.org/10.1037/lhb0000066
  • Morgan, J. (2023). Wrongful convictions and claims of false or misleading forensic evidence. Journal of Forensic Sciences, 68(3), 908–961. https://doi.org/10.1111/1556-4029.15233
  • National Institute of Standards and Technology. (2026). Face Recognition Technology Evaluation: 1:1 verification (ongoing public evaluation results, accessed November 3, 2026). https://pages.nist.gov/frvt/html/frvt11.html
  • National Registry of Exonerations. (2026). Exonerations in 2025: 2025 annual report (April 6, 2026). https://n2t.net/ark:/88112/x2v16v
  • National Research Council. (2009). Strengthening forensic science in the United States: A path forward. The National Academies Press. https://doi.org/10.17226/12589
  • National Research Council. (2014). Identifying the culprit: Assessing eyewitness identification. The National Academies Press. https://doi.org/10.17226/18891
  • Pena, M. M., Stoiloff, S., Sparacino, M., & Schreiber Compo, N. (2024). The effects of cognitive bias, examiner expertise, and stimulus material on forensic evidence analysis. Journal of Forensic Sciences, 69(5), 1740–1757. https://doi.org/10.1111/1556-4029.15565
  • Sauerwein, K., Butler, J., Reczek, K., & Reed, C. (2023). Bitemark analysis: A NIST scientific foundation review (NISTIR 8352). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.IR.8352
  • Thompson, W. C., & Newman, E. J. (2015). Lay understanding of forensic statistics: Evaluation of random match probabilities, likelihood ratios, and verbal equivalents. Law and Human Behavior, 39(4), 332–349. https://doi.org/10.1037/lhb0000134
  • Wells, G. L., & Bradfield, A. L. (1998). “Good, you identified the suspect”: Feedback to eyewitnesses distorts their reports of the witnessing experience. Journal of Applied Psychology, 83(3), 360–376. https://doi.org/10.1037/0021-9010.83.3.360
  • Wells, G. L., Kovera, M. B., Douglass, A. B., Brewer, N., Meissner, C. A., & Wixted, J. T. (2020). Policy and procedure recommendations for the collection and preservation of eyewitness identification evidence. Law and Human Behavior, 44(1), 3–36. https://doi.org/10.1037/lhb0000359