When Confidence Outruns Evidence: What 1929 Can Teach Us About AI in Police Interviews

Source and credit

This article was prompted by Sarah Gibbons’ interview with Dr Brandon May, published by Policing Insight on 24 August 2026, and by the Boring History for Sleep episode, “The Great Depression of 1929: When the American Dream Collapsed”, released on the same date.

The Policing Insight feature is a subscriber article. This is independent commentary and does not reproduce it. The podcast supplied the historical spark; historical propositions below have been checked against official Federal Reserve, FDIC and US Securities and Exchange Commission sources.

The short version

• The 1929 crash did not by itself cause the Great Depression. It exposed weaknesses—debt, speculation, fragile institutions and inadequate safeguards—which then transmitted harm through the financial system.
• AI can create a similar transmission problem in evidence. A leading prompt, inaccurate transcription or overconfident summary may feed the next tool, officer, prosecutor or court.
• Dr Brandon May told Policing Insight that the current empirical evidence does not support operational AI interviewing and warned of possible miscarriages of justice. His statement about errors becoming undetectable at a third review stage was expressly described as preliminary research.
• UK adoption is already moving quickly. A June 2026 research report mapped 70 AI tools deployed, trialled or in development across law enforcement and criminal justice in England and Wales.
• A nominal “human in the loop” is not enough if the human sees only an AI-shaped document, lacks the original source, or is primed to confirm a polished output.
• Police material also enters private and public family proceedings. An error born in a criminal investigation can therefore affect safeguarding assessments, fact-finding and decisions about children even where no criminal conviction follows.
• The answer is not a blanket rejection of AI. It is a hard evidential boundary: no unvalidated AI should create, elicit or materially rewrite testimony capable of affecting life, liberty or family life.

1. The warning that arrived at the right moment

The history episode begins with a blunt proposition: the disaster did not arrive without warning. The signs were there, but the prevailing story of progress was more attractive than the evidence of fragility.

Within minutes, LinkedIn supplied a modern version of that warning.

Policing Insight reported Dr Brandon May’s view that police forces remain “decades” of research away from safely deploying AI in interviewing contexts. In the accompanying post, his most important boundary was expressed in stark terms:

“AI should not be deployed if it has an impact on life, limb or liberty.”

That is not a claim that every use of AI in policing is harmful. Dr May’s own work identifies possible benefits in training and carefully bounded assistance. His warning is directed at operational use where an unvalidated system can shape an account, create evidential material or influence what happens to a person.

He also described preliminary findings suggesting that once an error enters the process, further errors tend to follow and may become undetectable by a third review stage. That distinction matters. This is an emerging research programme, not a final peer-reviewed conclusion. But it identifies a risk that criminal justice cannot responsibly wait to discover through wrongful cases.

The immediate question is therefore not whether AI can ask a grammatically correct open question. It plainly can. The question is whether the system can obtain a reliable human account without leading, validating, elaborating, contaminating or quietly rewriting it—and whether every later user can still see exactly what happened.

2. What 1929 actually teaches us

The popular story says the stock market crashed in October 1929 and the Great Depression followed. The reality is more instructive.

The Federal Reserve’s historical account explains the speculative expansion, the role of easy credit and the collapse in share prices. Its wider history of the Great Depression describes how banking failures, deflation, policy errors and institutional weakness deepened and transmitted the crisis. The money supply fell by nearly 30 per cent between autumn 1930 and winter 1933. A severe shock became a prolonged social catastrophe because the system could not contain it.

The consequences were not carried only by the people who had made speculative bets. Bank failures destroyed savings. Businesses closed. Workers lost employment. Families who had never traded a share paid for risks created elsewhere in the system.

Protective institutions followed the damage. The Securities Act 1933 and Securities Exchange Act 1934 imposed disclosure and market safeguards. Federal deposit insurance was created in 1933 amid thousands of bank failures. Regulation was built from the wreckage.

3. The same institutional pattern

There is a temptation to dismiss historical comparison as rhetoric. So the limits should be stated clearly. A probabilistic language model is not a bank. A witness account is not a share certificate. A wrongful prosecution is not an economic recession.

What can properly be compared is the pattern by which risk is created, normalised and passed downstream.

This is why the fashionable instruction to “move fast” is unsuitable at the point where information becomes evidence. Speed can be valuable in administration. In an evidential workflow, speed without provenance simply enables an error to travel further before anybody notices it.

4. Why an interview is not ordinary data collection

A witness does not contain a complete, fixed record waiting to be downloaded. Human memory is reconstructive. The form of a question, information encountered after an event, another witness’s account and the interviewer’s responses can influence later recall.

That is why scientifically informed interviewing took decades to develop. The Ministry of Justice’s Achieving Best Evidence in Criminal Proceedings guidance addresses how child and adult victims and witnesses should be interviewed and adopts a victim-centred, trauma-informed approach. The process is not merely a list of open and closed questions. It includes rapport, communication needs, pacing, context, vulnerability, clarification and professional judgment.

In a June 2026 analysis, Dr May and Professor Becky Milne explained the particular risk of a system that validates or elaborates an account. A conversational AI designed to be helpful may agree too readily, fill a silence, suggest coherence or repeat a detail in a way that strengthens it. The result may feel supportive while changing the evidence.

That creates at least four distinct risks:

1. Elicitation error: the system asks a leading, confirmatory or contextually inappropriate question.
2. Memory contamination: the witness encounters a detail or narrative not present in the original memory and later incorporates it.
3. Representation error: transcription or summarisation changes uncertainty, sequence, attribution or wording.
4. Authority error: later readers give the machine-shaped account greater weight because it is coherent, formatted and embedded in an official record.

The last risk is especially dangerous. Human recollection is often hesitant, incomplete and non-linear. An AI system is rewarded for producing an answer. It can therefore remove the very uncertainty that a court needed to see.

5. How one error becomes an evidential chain

Imagine a witness gives an early account through an AI-assisted interface. The system asks a question that subtly assumes a fact. The resulting conversation is transcribed automatically. A second tool turns the transcript into a witness statement. A third summarises it for an investigator. The investigator’s report informs a charging document. The prosecutor and later the court see the final, polished version.

1. Human experience and memory — incomplete, contextual and sometimes uncertain.
2. AI-mediated questioning — capable of introducing assumptions or post-event information.
3. Automated transcript — capable of mishearing names, negation, pauses, overlap or emotional nuance.
4. AI-drafted statement — capable of converting speech into a smoother and more definite narrative.
5. AI or human summary — capable of selecting facts that fit an emerging theory and omitting ambiguity.
6. Official reliance — the transformed account informs investigation, charge, trial or safeguarding decisions.

No single person in that sequence needs to act dishonestly. Each may believe they are checking the work of the previous stage. But if each reviewer receives only the transformed output, nobody is checking the original account.

This is the problem of chaining. A major June 2026 report, The Use of Probabilistic AI in Law Enforcement and the Criminal Justice System of England and Wales, warns that connected AI stages may inherit and compound earlier errors. It also found that the familiar safeguard of a “human in the loop” can be performative rather than real.

6. The current position in England and Wales

This debate is not about a distant future.

The Northumbria-led PROBabLE Futures project mapped 70 AI tools already live, being trialled or under development across law enforcement and the criminal justice system in England and Wales. It identified 27 as live and around 34 at trial or pilot stage; 52 per cent came from commercial vendors. The report found genuine benefits in carefully designed uses such as transcription, redaction, crime analysis, vulnerability identification and reducing officers’ exposure to traumatic material. Its central warning was that adoption is moving faster than the institutional, evidential and governance arrangements needed to support it.

On 10 June 2026, the Government formally launched PoliceAI, backed within a wider commitment of £115 million over three years. Its first-year programme includes pilots in up to ten forces to triage, disclose and summarise digital evidence, with wider scaling planned from 2027. The Government has promised independent testing for accuracy and bias and a public registry of AI tools used in policing.

Those safeguards are welcome. The announced pilots are not the same as autonomous witness interviewing, and they should not be described as such. But triage, disclosure and summarisation sit close to the evidential core. A summary can influence which material is read, which line of inquiry is pursued and which document is treated as important. Efficiency at that stage is valuable only if omission and distortion can be detected.

There is also a live factual warning. In June 2026, Derbyshire Police confirmed that a criminal investigation had been launched into an allegation that an officer used AI systems to create evidential material in a number of cases. The force said it was working with the Crown Prosecution Service in relation to potentially affected cases. The officer was removed from frontline duties and, at that stage, no arrest had been made.

That remains an allegation under investigation. It proves neither guilt nor that any conviction was unsafe. It does prove that provenance, authorised use and case-remediation procedures can no longer be treated as theoretical governance questions.

7. Why existing legal safeguards may not be enough

England and Wales already has substantial protections governing interviews and evidence.

• Section 76 of the Police and Criminal Evidence Act 1984 governs the admissibility of confessions and requires exclusion where the statutory test concerning oppression or unreliability is met.
• Section 78 of PACE gives the court discretion to exclude prosecution evidence where, having regard to all the circumstances, admission would adversely affect the fairness of proceedings.
• PACE Code E requires the whole of relevant suspect interviews to be audio recorded, subject to defined exceptions, and contains safeguards for preserving the integrity of the master recording.
• The Criminal Procedure and Investigations Act 1996 disclosure regime requires investigators to pursue reasonable lines of inquiry and to retain and reveal material according to the statutory test. Depending on relevance, AI inputs, prompts, outputs, edits and audit records may be capable of undermining the prosecution or assisting the defence.
• The CPS’s published approach to AI commits to human oversight, transparency, continuous evaluation and accountability.

But those protections were largely designed around identifiable human acts and stable records. AI creates a more elusive problem. The original audio may be intact while a derived statement subtly changes its meaning. A final decision may be made by a human while the material presented to that human has already been selected, compressed or contaminated by a machine.

Data-protection law also provides safeguards for significant decisions based solely on automated processing. The Information Commissioner’s Office stresses that human intervention must be genuine rather than token. Yet the most consequential AI influence may occur before the “decision”: in the construction of the evidence on which a human later decides.

That is the regulatory gap. A court can test an identified defect. It cannot test an invisible transformation nobody recorded.

8. Why this matters in the Family Court

Police evidence does not remain inside the criminal justice system.

The 2024 protocol on disclosure between criminal and family jurisdictions applies to private and public family law proceedings and to material held by the police. Police logs, interviews, body-worn video, statements, risk assessments and investigation records may all become relevant to decisions about domestic abuse, safeguarding and children.

Family proceedings also have different evidential features. Under the Children (Admissibility of Hearsay Evidence) Order 1993, hearsay evidence concerning a child’s upbringing, maintenance or welfare may be admissible. Facts are generally determined on the balance of probabilities, as confirmed in Re B (Children) [2008] UKHL 35.

None of that means family judges accept police material uncritically. Admissibility is not the same as reliability or weight. It does mean that an AI-shaped police record could influence family proceedings even where no criminal court has tested it and no conviction has followed.

The pathway may look like this:

• an early account influences a police risk assessment;
• the police record informs a local authority or Cafcass safeguarding response;
• the record is disclosed into private children proceedings;
• the wording affects whether the court orders fact-finding, interim contact or protective arrangements;
• the resulting procedural course shapes a child’s relationship with a parent long before the original source is examined in detail.

This can cut in every direction. AI contamination could strengthen an allegation beyond what a witness said. It could also flatten a frightened, fragmented account into something that appears inconsistent; omit coercive context; misattribute a speaker; or make a victim sound more certain or less credible than the original recording supports.

For family justice, provenance is therefore a safeguarding issue. The court must be able to distinguish:

• what the witness actually said;
• what the interviewer asked;
• what a machine transcribed;
• what a machine or officer summarised;
• what was later edited; and
• what is evidence, professional opinion, inference or generated wording.

9. Questions that should be asked now

Where AI may have touched police material relied upon in criminal or family proceedings, “a human checked it” is not a sufficient answer. The relevant questions are concrete.

1. Was AI used at all? This includes interviewing, live prompts, transcription, translation, summarisation, statement drafting, risk scoring, report writing, disclosure review and decision logs.
2. Which system and version? Models change. A product name alone may not identify the system that produced the output.
3. What was the exact task? Correcting punctuation is different from eliciting facts or rewriting a witness account.
4. What entered the system? The original prompt, source recording, documents and contextual instructions matter.
5. What came out? Every material output, not only the final selected version, should be preserved where it forms part of an evidential process.
6. Who edited or approved it? The record should identify human changes, the basis for them and the reviewer’s access to the source.
7. Can the final wording be traced to the original? Factual propositions should be auditable back to time-coded audio, video, notes or documents.
8. Were uncertainty and disagreement preserved? Pauses, qualification, ambiguity and corrections may be evidentially important.
9. Was the use authorised and validated for this population and task? A tool tested on clear adult speech is not thereby validated for a traumatised child, an interpreter-mediated interview or a person with a cognitive disability.
10. What happened when the tool failed? There should be an incident route, correction process and means of identifying every downstream document or case affected.

Requests in live proceedings must remain relevant and proportionate; this is not an invitation to speculative disclosure applications in every case. But where there is a proper basis to suspect material AI involvement, the provenance of the evidence may itself be a live issue of fairness and reliability.

10. Eight red lines for safe adoption

History’s most expensive lesson is that reform after collapse costs more than restraint before it. For AI in policing and evidential work, the minimum safeguards should be clear.

1. No unvalidated AI interviewer in an evidential role. A system should not conduct a first account or substantive witness or suspect interview capable of affecting life, liberty or family life unless it has been independently validated in realistic conditions and for the populations it will encounter.
2. The original source must remain the evidential anchor. Audio, video, contemporaneous notes and source documents must be preserved. A generated transcript or statement must never become the only surviving record.
3. Every material transformation must be labelled. The document should disclose that AI was used, the task performed, the system and version, the human reviewer and the date of review.
4. Review must be source-first and independent. The reviewer should actively compare the output with the original rather than merely read along with AI-generated wording, which risks confirmation bias.
5. Prompts, outputs, edits and logs must be retained. If AI touches evidence, the audit trail may be as important as the final document. Procurement terms must permit lawful scrutiny and disclosure.
6. Testing must measure evidential harm, not just average accuracy. Evaluation should include negation, names, accents, children, trauma, disability, interpreters, emotional speech, conflicting accounts and error propagation across connected systems.
7. There must be a public incident and remediation framework. Authorities need a way to identify affected cases, notify prosecutors, defence teams and family courts, correct records and review outcomes. A registry should record withdrawals and failures as well as approved tools.
8. Use AI where the risk is genuinely lower. Training simulations, administrative support and carefully governed tools can offer value. The correct response to risk is disciplined task selection, not technological theatre or a blanket ban.

PoliceAI’s promised testing and public registry can provide part of that architecture. But governance must be capable of saying no to a tool that is fast, impressive and commercially attractive when it is not evidentially safe.

Final thoughts

The Great Depression was not one bad decision made on one bad day. It was the collapse of confidence in a system whose fragility had been accumulating in plain sight. The crash exposed the weakness; the chain transmitted it; ordinary people carried it.

AI in policing is at a much earlier point, and that is precisely why the comparison matters. We still have the opportunity to place safeguards before systemic harm rather than build them from overturned cases.

Artificial intelligence may help officers organise vast disclosure, transcribe recordings, train for difficult interviews and spend more time with the public. Those possibilities deserve rigorous development. But the moment a system begins to shape what a witness says, what the official record means or what a later decision-maker is allowed to see, efficiency is no longer the only measure.

The measure is justice.

We should not ask only whether the tool performs better than a tired human on an average task. We should ask whether the process preserves memory rather than contaminates it; uncertainty rather than erases it; source evidence rather than generated narrative; accountability rather than diffuses it; and a person’s practical ability to challenge what the machine has done.

Sources and further reading

• Sarah Gibbons, Policing Insight, interview with Dr Brandon May, 24 August 2026 (subscriber article)
• Dr Brandon May, summary of key interview points, LinkedIn, 25 August 2026
• Boring History for Sleep, “The Great Depression of 1929: When the American Dream Collapsed”, 24 August 2026
• Brandon May and Rebecca Milne, “Investigative interviews are key to solving crimes—should AI be helping police with their inquiries?”, 29 June 2026
• Lawal et al, The Use of Probabilistic AI in Law Enforcement and the Criminal Justice System of England and Wales, 23 June 2026
• Home Office, “PoliceAI to speed up investigations and fight crime”, 10 June 2026
• Home Office, Police use of artificial intelligence factsheet, 9 June 2026
• ITV News Central, Derbyshire Police statement concerning alleged AI-created evidential material, 14 June 2026
• Ministry of Justice, Achieving Best Evidence in Criminal Proceedings
• Police and Criminal Evidence Act 1984, section 76
• Police and Criminal Evidence Act 1984, section 78
• PACE Code E: audio recording interviews with suspects
• Disclosure of information between family and criminal agencies and jurisdictions: 2024 protocol
• Children (Admissibility of Hearsay Evidence) Order 1993
• Re B (Children) [2008] UKHL 35
• Federal Reserve History, Stock Market Crash of 1929
• Federal Reserve History, The Great Depression
• US Securities and Exchange Commission, laws governing the securities industry
• Federal Deposit Insurance Corporation, historical timeline

Sources checked on 26 August 2026. Ongoing investigations are described as allegations and no finding of wrongdoing should be inferred. The historical comparison, legal analysis and reform proposals are the author’s commentary.

How JSH Law can help

JSH Law supports litigants in person to organise, analyse and present evidence in family proceedings. Where a police record, transcript, summary or professional report is important, we can help you separate source material from later interpretation, build a clear chronology, identify evidential gaps and formulate focused questions for disclosure or case management.

JSH Law does not provide forensic certification of AI systems. Where specialist digital-forensic analysis or regulated legal representation is needed, that should be obtained from an appropriately qualified professional.

Book a free 15-minute introductory call:

Important information

This article provides general legal information and commentary about England and Wales. It is not legal advice and should not be relied upon as a substitute for advice on the facts of an individual case. Ongoing investigations may develop after publication. Laws, court rules, guidance and technology also change, so current primary sources should always be checked.

JSH Law Ltd is not regulated by the Solicitors Regulation Authority. It does not conduct litigation, exercise rights of audience or undertake reserved legal activities. JSH Law provides practical, structured and procedural support to litigants in person.

Connect with JSH Law

Follow JSH Law’s work on family justice, access to justice and responsible legal technology.

JSH Law links and social channels