AI Hallucination ResearchPartnership › Regulators

Regulators — adoption playbook

How financial, markets, and conduct regulators use RegLegBrief findings to detect industry deviation from their administered rules, run right of reply at scale, and sensitise the regulated population to AI hallucination risk in their framework.

Last updated 14 Jun 2026 · For: central banks, securities regulators, financial supervisors, conduct regulators, prudential authorities

Bottom line for the regulator

The industry has started using AI to read your rules. AI is structurally degrading. You need a way to see where the industry has deviated.

The industry that you supervise is rapidly mediating its compliance work through AI. The AI that the industry uses is structurally degrading — model collapse is a documented, peer-reviewed phenomenon — and the rare, precise technical details of your administered instruments are exactly the parts that disappear first.

That means industry interpretation of your rules is drifting from what you wrote, in directions you cannot see from the top down. RegLegBrief is the verification layer that surfaces this drift, against your primary sources, and gives you a structured intake of every documented case.

In this playbook
  1. Why this matters now
  2. How the industry deviates from your rules
  3. Right of reply at scale
  4. Industry sensitisation
  5. Cross-regulator pattern detection
  6. In their own words — the regulators
  7. What RLB delivers to a regulator
  8. Engagement model
  9. FAQ

Skip to engagement: If you already lead supervision, policy, or conduct, jump straight to the inquiry form. The supervisory liaison route is the shortest.

Engage as Regulator partner →

1. Why this matters now

The case for a regulator to engage with AI-output verification is no longer hypothetical. Three forces have arrived together:

One: regulator AI rules are now law in the regulated population. Across major jurisdictions, supervisors and conduct authorities have shifted from issuing guidance on responsible AI adoption to fixing liability for AI-driven non-compliance on the regulated entities themselves and on licensed professionals acting in their capacity. The legal exposure for an AI hallucination in regulated work has crossed from theoretical to documented, with sanctions orders and bar-association formal opinions on record.

Two: AI itself is structurally degrading. A July 2024 peer-reviewed paper in Nature by researchers from Oxford, Cambridge, Imperial College London, the University of Toronto, and the University of Edinburgh named the phenomenon: model collapse. When generative AI models are trained on data that is itself AI-generated, the rare cases — the precise technical thresholds, the minority interpretations, the carve-outs — disappear first. By April 2025, 74.2% of new web pages contained AI-generated content (Ahrefs, 900,000-page study). Each generation of AI compresses the regulatory edge cases out of what it knows.

Three: the deviation is silent. The model collapse compounds without alerting anyone. Each AI cycle produces output that is more confident, more polished, more professionally formatted — and progressively more averaged-away from the precise instrument text. A regulated firm using a Generation-4 model gets the same confident tone it got from Generation 1, but the answer has drifted further from your primary source.

The regulated population is therefore moving toward your rules through a verification layer (AI) that is itself moving away from the source. Without an external verification mechanism, the industry's understanding of your administered framework will deviate without alerting either you or the industry.

→ Read the RLB Panel Speak on model collapse and regulatory drift

2. How the industry deviates from your rules

RLB has documented industry-AI deviation patterns across 18 published regulations and 94 verified hallucinations. The patterns are consistent across jurisdictions and AI subjects:

Failure modeWhat the industry-AI doesWhy your rule reads otherwise
Inference driftStates a number, threshold, or scope that sounds in-range but is not in the instrumentTraining corpus has averaged the actual threshold toward a plausible mid-point — the precise figure has been lost
Misstated ruleRestates your rule with a wording shift that flips an inclusion to an exclusion, or vice versaSecondary summaries (compliance blogs, vendor copy) routinely simplify; the AI is now training on the simplification, not the rule
MisattributedCites your rule for a proposition it does not contain, or attributes a proposition to a sibling regulationCo-citation patterns in the corpus have produced false associations the AI now treats as primary
OutdatedQuotes a superseded version, a withdrawn FAQ, a consultation draft, or a pre-amendment thresholdThe current version sits behind a portal; the AI defaults to the version that was scraped into open corpora

Each is a deviation from your administered framework that the regulated entity may not know it is making — because the AI's confidence is undisturbed by the drift.

3. Right of reply at scale

Every RLB finding is publishable. Before publication, the regulator whose rule is affected is invited to respond. The regulator decides:

The right-of-reply intake is a structured workflow, not an ad-hoc email exchange. The regulator gets a standardised package per finding: the primary source extract, the AI subject's output, the failure-mode classification, and a comment field. Response can be on-attribution (named regulator official) or on background (institutional response, no attribution).

Operational benefit to the regulator: every documented industry hallucination on your administered rule becomes either a regulator-endorsed clarification record or a published-with-context finding that the regulated population can see. Both outcomes harden industry understanding of your framework.

4. Industry sensitisation

RLB collaborates with regulators on industry-facing sensitisation work. Two formats:

Joint advisories — a regulator-issued advisory referencing published RLB findings as illustration of the AI risks under that regulator's framework. RLB provides the underlying evidence base; the regulator authors the framing and the supervisory expectation. Typical use: an existing operational-resilience or model-risk advisory adds AI hallucination as a new risk surface, with RLB findings as the worked examples.

Industry roundtables — closed-door briefings to regulated population subsets (large banks, asset managers, insurance, broker-dealers) on hallucination patterns documented under that regulator's framework. RLB presents the findings under NDA; the regulator's policy or supervision lead facilitates. Outcome: regulated entities walk out knowing which specific provisions their AI is most likely getting wrong, and what the supervisor expects them to do about it.

Both formats convert the RLB finding base into a regulator-led industry conversation, with the regulator retaining full editorial control over framing and supervisory expectation.

5. Cross-regulator pattern detection

RLB findings are tagged by jurisdiction, body, audience, and failure mode. For regulators participating in international fora (IOSCO, BCBS, IAIS, FSB) or in cross-border cooperation arrangements, this generates a cross-regulator analysis layer:

This is supervisory intelligence on rule-construction patterns — independent of any single regulated entity, available to the regulator under engagement, useful for policy refinement and for international peer comparison.

6. In their own words — the regulators

Senior regulators and global standard-setters are increasingly explicit that AI hallucination is a supervised risk surface, not an emerging curiosity. A representative selection:

"Hallucinations may not be intended, but they can represent a critical risk for financial services, where trust and credibility are paramount."
— International Organization of Securities Commissions (IOSCO), Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, Consultation Report CR/03/2025, March 2025. IOSCOPD788
"Today's large language models can produce answers that are fluent, confident — and wrong. In supervision, 'wrong but confident' is outright dangerous."
— Speech delivered at ECB Banking Supervision (SSM), Artificial intelligence and supervision: innovation with caution, 14 October 2025. ECB Banking Supervision
"New AI models … untested nature and the fact they can 'hallucinate' mean they should not become robo-ratesetters … We like to hold humans accountable."
— Cecilia Skingsley, Head, BIS Innovation Hub (former First Deputy Governor, Sveriges Riksbank), press remarks around the BIS Annual Economic Report 2024, 25 June 2024. BIS
"AI models such as ChatGPT can actually invent fake case studies sometimes referred to as 'hallucination bias'. This was visible in a recent New York court case with case citations by one set of lawyers being based on fake case material."
— Nikhil Rathi, Chief Executive, Financial Conduct Authority, speech: Our emerging regulatory approach to Big Tech and Artificial Intelligence, 12 July 2023. FCA
"The most frequently identified risks related to AI models were those related to explainability and interpretability, model bias, complexity, robustness and resiliency, hallucinations, and conflicts of interest."
— IOSCO, Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, CR/03/2025, March 2025. Survey of 24 IOSCO members. IOSCOPD788

7. What RLB delivers to a regulator

  1. Structured finding intake. Every documented case where an AI subject misrepresents a provision of your administered framework is delivered as a packaged finding: primary source extract, AI output, classification, failure mode, audience.
  2. Right of reply on every finding. Before publication, every finding affecting your framework goes through the right-of-reply intake. The regulator chooses how to respond.
  3. Industry-facing voice. Joint advisories or industry roundtables that convert documented findings into regulator-led sensitisation, with the regulator owning the framing.
  4. Cross-jurisdiction pattern intelligence. Tagged finding base supports cross-regulator analysis on rule-construction patterns, useful for policy and international peer fora.
  5. Model evaluation evidence. Structured documentation of AI subject behaviour on your administered rules — useful supervisory evidence when AI labs claim their models perform reliably in regulated work.

8. Engagement model

Services-led. Regulator engagements are scoped by:

Scope dimensionTypical regulator engagement
Frameworks in scopeOne regulation, one topic cluster, or one administered framework (e.g., conduct rules)
AI subjects under auditRLB's standard subject set (Sonnet 4.6, Opus 4.7, third subject) — additional subjects on request
Right-of-reply intake cadencePer-finding rolling intake, or batched (monthly / quarterly)
Sensitisation workJoint advisory, industry roundtable, both, or neither
Attribution policyNamed, institutional, or background — regulator's choice per output
ConfidentialityNDA governs the engagement; regulated entities are never named in public findings

Typical first engagement: a single administered regulation or a topic cluster, with right-of-reply intake on the resulting findings and one joint advisory or industry roundtable at the end of the cycle.

9. FAQ

Does RLB issue supervisory expectations or interpretations of regulator rules?

No. RLB classifies AI subject output against the regulator's own primary source. Interpretation of the rule remains with the regulator. RLB findings are observations of AI behaviour, not interpretation of your framework.

How does RLB choose which regulations to audit?

By a combination of supervisory salience (rules with broad regulated population impact), AI subject performance signals (rules that produce the highest hallucination yield in initial probes), and partner regulator priorities. A regulator partner can nominate frameworks for prioritised audit under the engagement.

Can a regulator commission a private finding base without public publication?

Yes. Default is publish-on-completion (with right of reply); commissioned engagements can be held private for a defined embargo period before publication, or kept off-publication entirely if the regulator's framework requires it.

How does RLB protect regulated entities whose AI surfaced a finding?

Public findings name the regulation and the failure mode. They never name the regulated entity whose AI subject surfaced the finding. RLB's standard subjects (Sonnet 4.6, Opus 4.7, and a third) are foundation models tested directly, not deployed in-firm. Findings on the foundation-model layer apply to every regulated entity using that model class.

Is the right-of-reply intake compatible with my jurisdiction's consultation procedures?

The intake is editorial, not procedural. It does not substitute for, and does not interfere with, a regulator's statutory consultation, rulemaking, or supervisory communication processes. It runs alongside them.

What happens to a finding the regulator considers wrong?

If the regulator's right-of-reply response demonstrates the finding misreads the primary source, RLB withdraws the finding from the public register. RLB's no-substrate-no-audit and publish-only-negative-findings rules already filter for primary-source-bound conflicts; the right-of-reply layer is the final correction layer.

Ready to engage? The supervisory liaison route is the shortest. The inquiry form pre-fills with the regulator track selected; describe your administered framework and the priority topics.

Engage as Regulator partner →

Related: methodology · hallucination register · partnership tracks overview · banks and financial institutions playbook · licensed practitioners playbook