AI Hallucination on Revisions to Business Conduct and Swap Documentation Requirements for Swap Dealers and Major Swap Participants for Compliance teams at Hedge Funds firms in the United States

Executive Summary

Compliance teams at hedge funds in the United States operate under the CFTC's December 2025 final rule revising business conduct and swap documentation requirements for swap dealers and major swap participants, a rule that touches pre-trade disclosure obligations, counterparty classification mechanics, and the treatment of swaps on non-US execution venues. Across two questions drawn directly from this regulation's operational detail, AI tools produced incorrect answers on both.

The failures were not edge-case omissions: in one instance AI mapped a staff no-action letter to familiar US execution venue categories instead of its actual scope covering eligible UK trading venues, and in another it extrapolated "eliminated in its entirety" to mean an obligation had been removed for swap types that were never within its scope in the first place.

Both errors went undetected in initial responses and were only surfaced when the AI was pressed further, meaning a compliance analyst who accepts a first-pass AI answer and moves on would carry a substantively wrong understanding into policy work, training materials, or internal legal sign-off.

How AI gets this regulation wrong

Both failures on this regulation share the same fingerprint: AI tools answered confidently with plausible-sounding detail, but the detail was wrong, and the errors only surfaced when the AI was explicitly challenged to reconsider. In each case the AI had anchored on a more familiar framing (US execution venues, broad product-scope elimination) and applied it to a narrower, jurisdiction-specific or technically precise regulatory construct it had not correctly internalised. The table below maps this pattern across the findings.

What that means for your team

Every failure identified on this regulation carries regulatory enforcement exposure, not hypothetical operational friction, but direct risk of a compliance function advising its business incorrectly on obligations that the CFTC's examination staff will test against. For a hedge fund trading through a swap dealer relationship, errors in understanding venue-scoped no-action relief or misreading which disclosure obligations were actually eliminated can translate into deficient written supervisory procedures, incorrect training certifications, or faulty regulatory capital and documentation assessments passed to counterparties. The table below sets out where enforcement risk lands for each finding.

When this affects your department

Hedge fund compliance teams reach for AI tools when mapping regulatory changes onto existing swap trading workflows, most commonly when a final rule or staff letter amends an obligation that the fund's swap dealer counterparties rely on, and the compliance function needs to quickly understand what changed, for which products, and whether the fund's own internal controls or onboarding documentation need updating.

The December 2025 business conduct and documentation rule is exactly that kind of instrument: it reorganises disclosure obligations, addresses cross-border execution certainty through staff guidance, and amends the §23.431 framework in ways that require a compliance analyst to correctly understand what prior requirements said before they can accurately characterise what changed.

The failure modes identified here hit precisely those use cases. When a compliance analyst asks an AI tool about the scope of CFTC Staff Letter 25-49, for instance while drafting a counterparty communication, updating a product approval memo for UK-venue-executed swaps, or advising the trading desk on documentation treatment, and receives a confidently wrong answer anchoring on US SEF/DCM venues instead of the letter's actual UK MTF/OTF scope, the firm risks building an incorrect understanding of when the no-action relief applies.

That error propagates into written supervisory procedures, counterparty representations, and potentially into representations made to external legal counsel or, in examination contexts, to CFTC staff.

Similarly, when compliance is advising on whether the PTMMM disclosure obligation has been eliminated for a specific swap type, a question that surfaces in documentation reviews, pre-trade workflow sign-offs, and training sign-offs for sales and structuring staff, an AI response that overstates the scope of elimination (incorrectly asserting that cleared CDS and other product types are now exempt when those products were outside the prior rule's scope anyway) hands the compliance function a technically incorrect framing.

The risk is not that the firm discloses more than required; it is that the compliance team frames its internal controls and training around a mischaracterisation of the regulation's architecture, which can be difficult and costly to remediate when tested under examination.

The findings at a glance

The two findings below cover the specific questions where AI tools produced incorrect answers on this regulation, each is the kind of targeted technical question a compliance analyst would ask when scoping obligations or advising on a recent regulatory change.

Aggregate impact

Both findings on this regulation cluster on the same underlying dynamic: the AI had a general understanding of the regulatory domain, ITBC swaps, pre-trade disclosure mechanics, §23.431, but resolved ambiguity in the specific question by defaulting to a more familiar or more general framing rather than the precise, bounded answer the source document supports.

In Finding 1, the AI's mental model of "ITBC swap treatment" mapped naturally onto the US SEF/DCM framework it had encountered repeatedly in broader swap dealer materials, causing it to mis-describe the actual scope of Staff Letter 25-49 as US-venue-focused when the letter specifically addresses UK trading venues. In Finding 2, it took the phrase "eliminated in its entirety" as a product-scope statement rather than a provision-structure statement, missing the crucial prior constraint that §23.431(a)(3) only ever applied to uncleared swaps, FX forwards, and FX swaps in the first place.

The systemic risk for a hedge fund compliance function is that these are exactly the categories of error that internal review is unlikely to catch without primary source verification. Both AI responses were internally coherent and referenced accurate surrounding context, the error was in one specific, loadbearing factual claim that a reviewer unfamiliar with the precise source text would have no reason to question. Compliance teams under time pressure, producing regulatory change analyses, training updates, or counterparty advisories, are the target workflow where this pattern of confident-but-wrong AI output causes the most damage.

For this regulation specifically, the two errors also reinforce each other thematically: one concerns the cross-border scoping of swap execution obligations, and the other concerns the technical architecture of what was and was not changed in the December 2025 rule. Both are areas where a fund's compliance team is likely to be producing work product, onboarding documentation updates, product approval memos, WSP amendments, and passing that work product upstream to legal or senior management for sign-off without independent source verification at every step.

What your team should do

The default position for compliance teams using AI tools on this regulation should be: AI is useful for orientation and drafting scaffolding, but any response that characterises the specific scope of a staff letter, no-action relief, or the technical architecture of a rule amendment requires primary source verification before it enters a work product. Staff Letter 25-49 and the December 2025 final rule text are both available directly through the CFTC's website, and the specific provisions at issue, the venue scope in the staff letter, the restructuring of §23.431(a), are short enough that verification takes minutes, not hours.

The cost of that verification is trivially low relative to the cost of a compliance function operating under a materially wrong understanding of either.

For AI-assisted work on this regulation, the safest use cases are summarising the structural changes the rule made at a high level, identifying which subsections of Part 23 were affected, generating a first-draft issue checklist for a regulatory change analysis, or producing training materials where the substantive claims are independently reviewed before distribution. The failure pattern here, confident anchor on a generalised framing, specific factual error in one claim, means AI output on venue-specific scoping questions and technical "what changed" questions should always be treated as a draft for verification, not a source of record.

Where the compliance function is advising a trading desk, counterparty onboarding team, or senior management on the specific scope of a no-action letter or the correct characterisation of what disclosure obligations exist post-amendment, the appropriate workflow is to verify the specific claim against the CFTC source text or through specialist legal counsel before relying on it.

AI tools we tested on this regulation were willing to correct themselves when challenged, but that means the error only surfaces if someone knows to push back, which is exactly the skill gap that makes AI hallucinations on technical regulatory questions dangerous for junior analysts working without that prior knowledge.

How RLB Can Help

RegLeg's published Hallucination Research functions as a pre-flight check before your Compliance team acts on AI output for regulatory questions. The research is public and regulation-specific, if you're using AI tools to interpret SEC, CFTC, or FINRA requirements, you can cross-reference the exact failure modes those tools have demonstrated against the relevant instrument before you rely on the output. That's a faster, more defensible control than ad hoc prompt testing, and it gives your team a documented rationale for the confidence level they attached to a given AI-assisted analysis.

Beyond the public research, we work with Compliance functions directly. The shape that typically adds most value for a hedge fund: a structured mapping exercise against your live AI-supported workflows, trade surveillance documentation, regulatory exam prep, investor disclosure review, AML/KYC policy interpretation, ranked by hallucination exposure based on what the research has surfaced for each regulatory domain. Funds that have built AI use into their Compliance processes often find the risk is concentrated in a narrower set of tasks than they expected, and that knowing where it sits is most of the work.

We can also run a confidential review of your firm's existing AI-use policy against the failure-mode catalogue and return a prioritised remediation list: gaps that need hard controls now, workflows that are low-risk with light documentation, and edge cases worth monitoring as regulatory guidance on AI use matures.

If your team needs to build internal capability, whether for CPD purposes, for onboarding new staff who are already AI-native, or to satisfy a regulator's expectation that you can demonstrate AI governance competency, we can develop training material grounded in the research findings. The content is calibrated for practitioners: it covers specific failure patterns, the regulatory contexts where they've appeared, and the judgment calls a Compliance professional needs to make when AI output touches a regulatory question.

It's not a general AI literacy programme; it's scoped to what your team actually needs to assess risk in the workflows you're already running.