AI Hallucination on the OECD Merger Review Recommendation (2025) for Accountants (CA/PA) in international jurisdictions

Executive Summary

The OECD's 2025 Recommendation on Merger Review sets out revised standards for competition authorities when assessing concentrations, including the conditions that must be satisfied before a failing firm defence can be accepted. For Accountants (CA/PA) advising multinational clients on cross-border merger transactions, understanding exactly what that defence requires in OECD member jurisdictions is material to deal structuring, remedies strategy, and signed professional opinions.

Across the question tested on this recommendation, AI tools produced answers that were wrong in ways that would not be immediately visible to a practitioner who did not have the primary text in front of them: the three-condition structure was reproduced with apparent confidence, but condition 3 was mischaracterised and the non-exhaustive nature of the test was dropped entirely. Both failures were confirmed only when the AI was challenged and acknowledged it could not verify the precise English formulation, a candid admission that came too late for any practitioner who had already drafted on the basis of the initial response.

How AI gets this regulation wrong

The failure pattern on this recommendation is one of confident partial accuracy: AI tools reproduced a recognisable three-part structure for the failing firm defence, then quietly substituted a traditional counterfactual test for the comparative harm standard the 2025 text actually imposes. Compounding this, the AI omitted the "inter alia" qualifier in the text, converting an explicitly non-exhaustive list into a closed cumulative test, and only conceded uncertainty about the primary source when pressed in a second call.

What that means for your practice

For Accountants (CA/PA) on international mandates, the dominant risk from the failure on this recommendation is professional liability: an opinion or advice memorandum that incorrectly frames the failing firm defence conditions exposes the practitioner if the client proceeds to file on that basis and the defence is rejected, or, worse, if the client's legal counsel relied on the accountant's framing to instruct local competition counsel across multiple jurisdictions simultaneously. The risk is concentrated at the point of sign-off, not discovery.

When this affects Accountants (CA/PA)

Accountants (CA/PA) on international M&A engagements reach for AI tools most acutely at the scoping stage of a cross-border deal, when a target's financial distress is first surfaced and the question arises whether a failing firm argument is even viable. At that point, the practitioner needs a reliable read on what the OECD standard actually demands: is condition 3 a straight counterfactual (assets leave the market regardless) or a comparative harm test (exit causes more harm than the merger does)?

That distinction determines whether to invest in a full failing firm economic analysis, how to structure the information request to local counsel across OECD member states, and whether a distressed vendor's position is commercially presentable as a defence at all.

A second and more acute exposure arises at the advice memo or opinion-letter stage. If a junior team member drafts the failing firm section by querying an AI tool without cross-checking the primary text, the condition 3 mischaracterisation and the dropped "inter alia" qualifier will propagate cleanly into the draft, they look like learned legal framing, not an error. The partner reviewing that draft has to know the 2025 revision's precise wording well enough to catch a plausible-sounding substitution.

In a multi-jurisdiction engagement where the opinion is being issued across eight or ten signatory states simultaneously, the blast radius of that error is not confined to one filing.

Training and internal knowledge-base applications raise a third exposure. Firms that use AI tools to build internal briefing notes or training materials for staff on the updated OECD recommendation will embed the mischaracterisation at scale. Staff who then advise clients in OECD member jurisdictions, particularly those whose domestic competition law tracks the OECD framework closely, will carry the wrong standard into every engagement until the error is caught, which typically happens only when a regulator pushes back.

The findings at a glance

The table below summarises the finding tested on the 2025 OECD Merger Review Recommendation, the question asked, the nature of the AI failure, and the risk category it maps to for Accountants (CA/PA) practice.

Aggregate impact

With one finding, the pattern is tight and the lesson is precise: AI tools on this recommendation are capable of generating structurally plausible answers that embed a substantive error at the level of a single word, the substitution of a counterfactual test ("assets would exit the market") for the comparative harm test the 2025 text requires ("exit causes more harm than the merger"). That kind of error is particularly dangerous because the surrounding architecture of the answer (three conditions, cumulative, each stated with appropriate technical vocabulary) looks authoritative.

A practitioner without the primary text in front of them has no reason to suspect that condition 3 has been silently rewritten.

The second dimension of the failure, dropping "inter alia" and presenting the three conditions as an exhaustive closed test, matters operationally because it forecloses anticipation of additional evidentiary requirements. Competition authorities in OECD member states may require supplementary evidence beyond the three conditions; an opinion that treats the list as closed will not prepare the client for that possibility, and the practitioner who signed off on that opinion carries the exposure if the defence fails on a ground they told the client was not live.

The candid acknowledgment of uncertainty that emerged only on challenge, the AI conceding it could not confirm the exact English formulation from primary sources, is diagnostic of a broader risk profile: initial confidence, latent uncertainty, and disclosure only under adversarial pressure. For Accountants (CA/PA) in international jurisdictions who use AI tools in time-pressured deal contexts, "adversarial pressure" is not a step that makes it into the workflow.

What your team should do

The default position on the failing firm defence under the 2025 OECD Recommendation is straightforward: do not use an AI-generated summary as the authoritative statement of the conditions. The text is short, is publicly available from the OECD, and the operative paragraph (Section III.11.b) can be read in under a minute. Any engagement where the failing firm argument is live, even as a fallback, warrants a direct read of the primary text by whoever is signing the opinion.

The AI's framing of condition 3 and its omission of "inter alia" are not detectable by a practitioner who is only cross-checking the structure of the answer against their memory of the pre-2025 doctrine.

For team use, the practical safeguard is to treat AI output on this recommendation as a starting scaffold for identifying which questions to ask, not as a source for what the answers are. AI tools are reliable for: drafting client-facing executive summaries once you have confirmed the conditions from the primary text; generating a first-cut checklist of diligence items for a distressed target; and flagging which OECD member jurisdictions' domestic merger control regimes expressly track the OECD framework.

They are unreliable for: stating the precise conditions of any defence that has been updated in a recent revision; characterising whether a list of requirements is exhaustive or illustrative; and identifying the precise standard of comparison the regulation imposes.

Where a junior prepares the initial draft of any failing firm section, the review step should explicitly include a line-by-line comparison of condition 3 against the OECD text, not just a structural check for whether three conditions are present. On multi-jurisdiction mandates, designate one person to hold the primary text and act as the check on AI-assisted drafts from local teams across OECD member states. That function does not need to be expensive; it needs to be assigned.

How RLB Can Help

RegLeg's published Hallucination Research is available as open reference, use it as a pre-flight check before relying on AI output on regulatory questions that matter to your sign-off. The findings are organised by regulation and failure mode, so if you are working across IFRS application guidance, PCAOB standards, or cross-border group reporting obligations, you can pull the relevant regulation page and see, specifically, where AI tools have fabricated citations, misstated effective dates, or collapsed jurisdiction-specific carve-outs into a single incorrect answer. That is faster and more defensible than discovering the error after the advice has gone out.

For firms running multiple Accountants on the same regulatory portfolio, group reporting, audit quality frameworks, independence requirements across jurisdictions, RegLeg offers bespoke deep-dives. We work through the specific regulations in scope, map the failure modes that surface most consistently in that regulatory space, and produce a structured briefing your team can use as a standing reference. This is not a one-size engagement: the output is scoped to the regulations you are actually using AI tools against, and framed around the workflow decisions those findings affect, materiality judgements, disclosure drafting, cross-border reconciliation.

We also produce training and CPD-aligned material built around the failure modes your team should be stress-testing in their own AI use. Not generic AI literacy content, specific failure patterns documented against the regulations accountants in international practice touch most, presented in a format that maps to the professional judgement calls your team makes daily. If your firm has an existing AI-use policy, we can review it confidentially against RegLeg's failure-mode catalogue and flag where the policy's assumptions about AI reliability are not supported by what the research actually shows.