Executive Summary
The CFTC's December 2025 final rule revising business conduct and swap documentation requirements for swap dealers is among the most technically dense updates in the post-Dodd-Frank framework—and its January 2026 correction narrows the compliance window further by reinstating provisions that briefly fell out of the codified text. Legal teams at US investment banks are the primary internal function responsible for translating these amendments into updated policies, swap documentation playbooks, and guidance to trading desks and counterparty-facing business lines.
Across three tested questions on this regulation, AI assistants produced hallucinated answers in every case—wrong on the identity of the reinstated appendix, wrong on the geographic scope of a key staff letter, and wrong about which swap types the eliminated PTMMM provision previously covered. The failures split between AI quietly omitting critical regulatory precision and AI giving confident, surface-plausible answers that collapsed when challenged—neither of which a Legal team working under time pressure is likely to catch before the output reaches an internal memo or policy update.
How AI gets this regulation wrong
AI tools we tested on this regulation displayed two distinct failure patterns. In one, the AI correctly acknowledged a regulatory event—the January 2026 correction—but silently dropped the specific appendix identity and the provisions it governs, producing a materially incomplete research brief. In the other two, the AI delivered confident, syntactically competent answers about regulatory scope, then retracted when challenged—substituting familiar regulatory templates (US SEF/DCM venue scope, broad product coverage) for the specific and narrower facts the actual regulatory text establishes.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 2 | Finding#2 · Finding#3 |
| Inference Drift | 1 | Finding#1 |
What that means for your team
Every failure identified on this regulation produced the same category of risk: a wrong deliverable—a compliance memo, internal policy brief, or regulatory gap analysis that contains a material factual error before it leaves the Legal function. For Legal teams at US investment banks, those outputs gate decisions by trading desks, product structuring, counterparty-facing business lines, and, in examination contexts, the bank's own position before CFTC staff—making the cost of an embedded error disproportionately high relative to the research shortcut AI appears to offer.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 3 | Finding#1 · Finding#2 · Finding#3 |
When this affects your department
The December 2025 rule and its January 2026 correction intersect with several standing Legal workflows at US investment banks. The reinstatement of Appendix A to Subpart H—governing §§23.434 and 23.440 on suitability for recommendations to counterparties and special entities—is exactly the kind of amendment that triggers a Legal review of the bank's internal suitability policy. When business lines, product structuring, or compliance ask Legal to confirm what the rule currently says, and which guidance remains operative, the resulting memo becomes the authoritative internal record.
If that memo relies on AI research that identifies the correction but omits which specific appendix was reinstated and what provisions it covers, the policy review it supports will be built on an incomplete foundation—leaving the bank's suitability documentation out of step with what the Commission currently expects.
Cross-border swap documentation is a second high-frequency trigger. Legal teams at US investment banks with UK-facing desks or FCA-regulated entities are regularly asked to map which documentation relief or safe harbour applies to which execution venues and instrument types. CFTC staff letters are a key input to that mapping, and their geographic scope determines where documentation requirements shift. An AI tool that characterises Staff Letter 25-49 as covering ITBC swaps executed on US SEFs and DCMs—when the letter specifically addresses eligible UK trading venues—produces a gap analysis with the wrong perimeter.
The bank either fails to capture the relief where it applies, or incorrectly assumes it applies to domestic venue execution where it does not. Either version of that error reaches the documentation team before it reaches Legal review, and the window for catching it is narrow.
The PTMMM elimination sits inside the bank's swap documentation playbook update cycle. When a rule eliminates a disclosure requirement, trading desks and operations teams need Legal's guidance on which disclosures are no longer required and from when. The relevant precision question—what the prior requirement actually covered—is precisely where AI tools on this regulation produced their most consequential error. An AI that over-reads "eliminated in its entirety" to include cleared swaps that were never within PTMMM's scope in the first place produces playbook guidance that mischaracterises the firm's pre- and post-rule documentation posture.
In a regulatory examination, the bank then has to explain why its internal records suggest the firm thought cleared CDS were previously subject to a requirement they were not.
The findings at a glance
Three tested questions on this regulation produced confirmed hallucinations—each involving a regulatory precision detail that would have passed silently into a Legal deliverable if accepted at face value.
Aggregate impact
The three failures on this regulation cluster around a structural gap in how AI tools handle regulatory precision: they perform well at the level of narrative description—a correction was issued, a staff letter was published, a provision was eliminated—but they systematically fail at the level of specificity that Legal compliance work actually requires.
The difference between knowing a correction reinstated "an appendix" and knowing it reinstated "Appendix A to Subpart H, titled Guidance on the Application of §§23.434 and 23.440 for Swap Dealers That Make Recommendations to Counterparties or Special Entities" is, from an AI output perspective, invisible; from a Legal deliverable perspective, it is the entire point. A compliance memo built on the first answer is not a compliance memo.
The two exposed fabrications share a second identifiable pattern: both involved the AI substituting a familiar regulatory template for the specific operative facts. In one, the AI mapped a staff letter onto the US SEF/DCM venue framework it already knows—rather than the narrower UK-venue relief the letter actually provides. In the other, it extrapolated an elimination of a provision to cover instrument types the provision never reached. Both initial answers were syntactically fluent and internally consistent. Both retracted when challenged.
The practical risk for Legal teams is not that experienced counsel is deceived; it is that a junior associate writing a first-pass research memo treats the AI's confident framing as an answer to transcribe rather than a hypothesis to verify. By the time the memo reaches a supervising attorney, the AI's wrong answer has been through a drafting pass and looks like research.
Across all three findings, the firm's exposure runs through the same channel: AI error embedded in a Legal deliverable that then gates downstream decisions. The CFTC's enforcement record on swap dealer business conduct and documentation obligations—particularly around suitability for special entity recommendations and pre-trade disclosure practices—means that policy misalignments and gap analysis errors on this rule carry direct regulatory risk, not just operational inefficiency. The January 2026 correction's narrow window and technical specificity make it exactly the kind of post-final-rule development that AI tools are most likely to get partially right and critically wrong.
What your team should do
The default position for any Legal team consulting AI on the December 2025 rule and the January 2026 correction is to treat AI output as a drafting scaffold, not a regulatory source. The three failures on this regulation all involve precision details—specific appendix identity and governing provisions, specific geographic scope of a staff letter, specific pre-rule product scope of a deleted paragraph—that AI tools reliably elide or replace with plausible-sounding substitutes.
For initial orientation on the broad shape of what changed, AI can usefully summarise the rule's major themes; that summary should not drive a compliance memo without verification against the Federal Register text and, for the January 2026 correction specifically, the correction notice itself.
For questions about staff letters and no-action relief—where the operative question is the exact scope of which venues, counterparties, or instruments are covered—Legal teams should treat AI answers as requiring primary-source verification before use. Staff letters are narrow and frequently supersede prior letters in partial ways that AI tools conflate with the broader codified rule. For Staff Letter 25-49, the right verification step is the published letter on the CFTC's website, not an AI summary, before any cross-border documentation gap analysis is finalised.
Where a letter is recent enough that the AI's training may not fully capture it, the risk of a confident but wrong answer is highest—and the team should assume the answer has not been verified until the primary source has been read.
For questions about the scope of eliminated provisions, a standing safeguard is to ask the AI to characterise what the prior requirement covered before describing what the amendment removed. The PTMMM failure here was not about the elimination itself—the AI correctly identified that the provision was deleted—but about what it assumed the provision previously applied to. Asking the AI to state the pre-amendment scope as a separate step surfaces that assumption and creates a checkpoint before it becomes embedded in a playbook update.
For this regulation more broadly, the Federal Register preamble is the most reliable guide to the Commission's intent on each amendment, and should be the primary reference for any update to the firm's swap documentation standards or suitability policy.
How RLB Can Help
RegLeg's published Hallucination Research is available now, free of charge, as a pre-flight check before your team relies on AI output on any regulatory question we've tested. If your attorneys are using AI tools to answer questions on FINRA rulebooks, SEC disclosure requirements, Dodd-Frank swap-dealer obligations, or cross-border capital treatment, the published findings tell you concretely where those tools fabricate citations, invert positions, or confuse jurisdictional scope — before that output reaches a brief, an opinion, or a client memo. That is not a theoretical risk catalogue; it is a documented failure log against the actual regulatory text.
For firms that want to go further, we run bespoke regulator deep-dives scoped to the specific AI-supported workflows your Legal function is running today. That means mapping your actual use cases — regulatory change monitoring, red-line drafting, internal compliance Q&A, deal-specific regulatory opinion research — against the hallucination failure modes we have characterised for the relevant US and cross-border instruments, and returning a ranked exposure assessment your team can act on. The output is workflow-specific, not a generic AI-risk framework rehash; it reflects the regulations your deal teams and compliance counsel are actually touching.
We also conduct confidential reviews of existing AI-use policies against our failure-mode catalogue. If your firm has already drafted or deployed an AI governance policy for the Legal function, we will tell you where it underspecifies the risk relative to what we have observed in practice and return a prioritised remediation list — sequenced by the workflows carrying the highest exposure, not by document structure.
Where your team needs internal training or CPD-aligned content on AI reliability in regulatory practice, we can develop that material directly from our findings, so attorneys understand not just that AI tools can hallucinate regulatory content, but the specific failure patterns most likely to surface in the work they do daily.