Executive Summary
AI assistants tested against the CPMI-IOSCO 2026 consultation on updated initial margin guidance produced a materially misleading characterisation of a CCP's public disclosure obligations — converting a "should" standard into a "must" and inventing specific disclosure sub-elements not found in the consultation text. For lawyers advising CCPs, infrastructure operators, or clearing members on the disclosure architecture this guidance is shaping, that distinction is not a drafting nuance: it is the difference between a best-practice expectation and a binding obligation.
The single confirmed finding carries direct liability exposure — an opinion or compliance memo built on the AI-generated version of this obligation would misstate both the nature of the duty and its granular content.
How AI gets this regulation wrong
The dominant error pattern on this consultation is modal misrepresentation combined with content fabrication: AI tools restated a recommendation-grade standard as a mandatory requirement and populated that invented obligation with specific disclosure sub-elements nowhere present in the source text. When the same AI was challenged on its answer, it conceded it had not been working from reliable knowledge — a pattern that is only useful as a safeguard if the practitioner already knows to push back.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your practice
For lawyers in international jurisdictions, the failure translates directly into liability and professional indemnity exposure: an opinion letter, board memo, or client briefing that mischaracterises the obligation's binding force — or that lists fabricated disclosure requirements as if they were mandated — creates the conditions for a negligent advice claim. The risk is acute precisely because the AI's answer was internally consistent and confidently framed, giving no visible signal that the modal upgrade and the invented sub-elements were not sourced from the document.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Liability / PI exposure | 1 | Finding#1 |
When this affects Lawyers
Lawyers advising CCPs or clearing members on the 2026 consultation are most likely to reach for AI when scoping the disclosure obligations the guidance is expected to crystallise into — particularly when advising on how to build or audit an override framework that will withstand regulatory scrutiny once the final guidance is published. The AI's answer looks like precisely the kind of structured, obligation-level summary a practitioner would use to anchor a compliance memo or draft a board-level briefing on what public disclosure the CCP needs to put in place.
The practical risk surfaces at two moments: first, when a lawyer uses the AI output to advise a client on whether public disclosure of override practices is a hard requirement or a best-practice expectation — the difference between "you must" and "you should" directly shapes how the client budgets, prioritises, and defends its disclosure posture to regulators.
Second, when a lawyer or a senior associate uses the AI's enumerated sub-elements (categories of override warranting disclosure, decision-maker identification, permissible adjustment types) to draft a disclosure framework template or vendor-facing terms — those sub-elements are not in the consultation text, and building a framework around fabricated granular requirements creates a compliance programme that answers obligations the regulator did not actually impose.
The consultation's current "should" framing also matters for cross-border advice, since international lawyers frequently need to assess whether CPMI-IOSCO standards translate into binding obligations through domestic implementation or remain at the expectation level. An AI that has already resolved that question in the wrong direction — before the practitioner even asks — pre-poisons the analysis.
The findings at a glance
The table below summarises the confirmed AI failure on this consultation, including the question area, the nature of the error, and the risk category it triggers for lawyers in international jurisdictions.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | CCP override framework disclosure: 'should' vs 'must' and fabricated sub-elements | Hallucination | RLB-F-INT-BIS-CPMI-IOSCO-INITIAL-MARGIN-DISCLOSURE-CONSULT-2026-Q005 |
Aggregate impact
The single confirmed finding on this consultation illustrates a failure pattern that is particularly hazardous on consultative documents: AI tools trained up to a knowledge cutoff lack clean access to the consultation text itself, so they reconstruct the answer from adjacent knowledge — in this case, probable familiarity with CPMI-IOSCO's general disclosure philosophy and prior margin guidance. The reconstruction is coherent enough to pass a non-specialist review but wrong in two distinct ways simultaneously: the obligation's modal weight is upgraded from a recommendation to a mandate, and the specific content of that mandate is populated with invented sub-elements.
For lawyers, this is not a random error — it is a systematic one. CPMI-IOSCO guidance habitually uses modal language ("should", "are expected to", "it is recommended that") as a deliberate instrument for calibrating the regulatory expectations CCPs face across the PFMI implementation landscape. Getting "should" vs "must" right is a core competency of international financial regulatory practice, and it is precisely the kind of distinction AI tools consistently flatten when they are working from incomplete source access.
The fact that the AI conceded uncertainty only when challenged — not when first responding — means the error is invisible to any workflow that does not include an adversarial verification step.
The fabricated sub-elements compound this: a lawyer who accepts the AI's answer ends up advising on disclosure requirements that have no basis in the consultation, potentially driving a client to over-engineer its disclosure programme or — worse — to treat a fabricated checklist as a reliable basis for a legal sign-off on adequacy.
What your team should do
The default position for any AI-assisted work on this consultation should be: never allow an AI-generated characterisation of a CPMI-IOSCO obligation's binding force to survive into a client deliverable without direct verification against the consultation text. This applies with particular force to modal language — if an AI says a CCP "must" do something under this guidance, verify before repeating that characterisation to a client. Treat AI-generated obligation summaries on consultative documents as drafts that require paragraph-level checking against the source, not as reliable starting points.
For the override framework disclosure question specifically, the practical safeguard is to go directly to the consultation text for the "should / must" determination and to treat any enumeration of specific disclosure sub-elements in an AI response as suspect unless you can locate each one in the source. If a junior on your team has used AI to scope the disclosure requirements and the answer includes a structured list of categories — decision-maker identification, permissible adjustment types, instances warranting disclosure — ask them to point to the paragraph. If they cannot, the list is likely AI-generated fill, not consultation content.
Where AI tools are reliably useful for lawyers working on this consultation: horizon-scanning for how CPMI-IOSCO's prior guidance on model governance and transparency has evolved, identifying analogous domestic implementation frameworks across jurisdictions, and drafting initial skeleton structures for comment letters or client briefings — provided the specific obligation characterisations in those drafts are replaced with verified language before issue. AI performs well on structural and contextual tasks; it performs poorly on exact obligation characterisation from source documents it cannot reliably access.
How RLB Can Help
RegLeg's published Hallucination Research is available as a free pre-flight check for lawyers working on regulatory matters. Before relying on AI-assisted output — whether for advice, drafting, or due diligence — lawyers can consult the research to understand which failure modes have been observed for the specific regulation in question. This is not a substitute for legal judgement, but it is a structured, independent reference that flags where AI tools have historically misfired, allowing practitioners to focus their human verification effort on the highest-risk points.
For firms where multiple lawyers work across the same regulatory portfolio, RegLeg offers bespoke deep-dive engagements. These go beyond the published research to examine the specific regulations, jurisdictions, and question types most relevant to the firm's practice. The output is a tailored briefing that legal teams can use as a standing reference — updated as the regulatory landscape evolves — giving the whole team a shared, consistent picture of where AI tools should be treated with caution and where they have performed reliably.
RegLeg also works with legal teams on training and CPD-aligned content. This covers the categories of failure lawyers are most likely to encounter — including outdated regulatory text, cross-jurisdictional confusion, and misattributed citations — framed around real regulatory examples rather than abstract AI theory. Separately, RegLeg can conduct a confidential review of a firm's existing AI-use policy, assessing it against the failure-mode catalogue the research has surfaced. The output is a structured gap analysis: which risks the policy already addresses, which it does not, and where practical amendments would strengthen the firm's position.