Executive Summary
Legal teams at Corporate Banking firms operating across international jurisdictions use the OECD Merger Review Recommendation as a baseline framework when advising on cross-border M&A transactions, particularly when client mandates span multiple OECD member jurisdictions and the team needs to map defensive arguments available to a target or acquirer under each regime's merger control process.
Across the single substantive question we put to AI tools on the 2025 Revision, the AI produced a confidently wrong answer: it mischaracterised the third condition of the failing firm defence and simultaneously presented the defence's conditions as an exhaustive, closed test when the Recommendation explicitly signals they are not. When challenged, the AI tools acknowledged they could not confirm the exact formulation from primary sources, meaning the initial confident answer had no reliable foundation.
For a Legal team advising a distressed-asset deal or defending a merger notification where the failing firm defence is in play, that combination of confident error and pretextual sourcing is precisely the shape of failure that ends up in a regulator's file.
How AI gets this regulation wrong
The dominant failure pattern on this regulation is AI presenting authoritative-looking doctrine that subtly substitutes traditional common-law formulations for the standard actually codified in the 2025 text, and simultaneously collapsing a non-exhaustive list into a closed, exhaustive test. The AI tools cited real OECD documents for support, but those sources do not establish the proposition being asserted; the citations function as window dressing rather than verification. The table below maps how that failure manifests across the tested questions.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 1 | Finding#1 |
What that means for your team
For the Legal function at a Corporate Banking firm, the practical risk from this failure concentrates in a single category: a wrong deliverable reaching a client, a transaction counterparty, or a regulator. The AI error on the failing firm defence is the kind that survives initial review, it looks like settled doctrine, and surfaces only when the competent authority pushes back on the third condition or requires a more demanding evidentiary showing than the team anticipated. The table below maps that risk exposure across the findings.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
Corporate Banking Legal teams reach for AI tools on the OECD Merger Review Recommendation in a narrow but high-stakes set of situations: mapping the cross-jurisdictional merger control landscape for a complex multi-party transaction, drafting internal guidance notes on available defences for distressed-asset acquisitions, and pressure-testing a client's or counterparty's proposed arguments before a notification filing.
In a distressed-asset context, where a Corporate Banking client is acquiring a failing counterparty or where the bank itself is disposing of a non-performing loan portfolio to a single acquirer, the failing firm defence often determines whether the transaction clears at all or requires extensive remedies. The team's internal memoranda and external submissions on that defence are built, at least in part, on their understanding of the standard the competent authority will apply.
The specific risk here is that the 2025 Revision's third condition is not the traditional "assets would inevitably exit the market" formulation that counsel trained on U.S. or pre-2025 OECD doctrine will recognise. Section III.11.b requires evidence that the exit of the firm's assets would cause more harm to competition than the merger itself, a comparative harm test that is materially more demanding and more fact-specific than the traditional counterfactual.
If a junior lawyer uses an AI tool to draft the team's internal position paper or a submission outline, and the AI substitutes the traditional formulation for the 2025 text, the error is invisible until the competent authority applies the correct standard and finds the evidentiary showing inadequate.
Equally, the AI tools tested dropped the "inter alia" qualifier that Section III.11.b explicitly includes, converting a non-exhaustive evidentiary list into a closed, exhaustive three-condition test. Competent authorities applying the 2025 standard retain discretion to require additional evidence beyond the three named conditions. A Legal team that advises the client the defence is mechanically available upon satisfying three conditions, without flagging the authority's residual discretion, has understated both the evidential burden and the regulator's enforcement latitude.
In a multi-jurisdictional transaction where the bank is advising on simultaneous filings across OECD member jurisdictions, that misread propagates across every jurisdiction-specific submission the team produces.
The findings at a glance
The table below summarises the single finding tested on this regulation, the AI's handling of the failing firm defence standard under the 2025 Revision, with outcome, failure mode, and risk category at a glance.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Failing firm defence: condition 3 mischaracterised, conditions presented as exhaustive | Hallucination | RLB-F-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q005 |
Aggregate impact
With one finding in this cell, there is no cross-finding pattern to aggregate, but the single finding is structurally significant because it combines two independent error modes in one answer. The first is a substantive mischaracterisation of the third condition: the AI substituted the traditional counterfactual test (target assets would exit regardless) for the 2025 Recommendation's comparative harm test (exit would cause more harm to competition than the merger). These are not synonyms.
The comparative harm test requires the parties to affirmatively demonstrate that the competitive harm from the merger is less severe than the harm from an uncontrolled exit, an evidential burden that points toward market impact analysis, not just the inevitability of the target's failure. The second error mode, dropping "inter alia" and presenting the three conditions as exhaustive, removes the competent authority's explicit residual discretion from the Legal team's working model of the defence.
The combination is more dangerous than either error alone. A Legal team that has both misread condition 3 and believes the test is closed at three conditions will under-resource the comparative harm analysis and over-commit to the defence's availability before the authority has been tested. In a transaction context, that translates to an internal sign-off memo that overstates the strength of the failing firm defence, a client being advised the deal will clear when the evidentiary showing is actually inadequate, or a notification submission that the authority finds falls short on the most demanding condition.
The sourcing pattern compounds this: both cited sources are real OECD documents, and a junior reviewer cross-checking the AI's answer by looking at the cited URLs would find genuine OECD material, nothing that obviously signals a fabricated or displaced standard. The error is not caught at the citation-check stage; it is only caught by reading the 2025 text directly against the AI's formulation of condition 3. For a Legal team relying on AI to accelerate first-draft research, that catch has to be a deliberate step, not an incidental one.
What your team should do
The default position for this regulation should be that AI tools are not reliable for generating the operative standard on specific defences or conditions without a direct text check against the 2025 Recommendation itself. The failing firm defence finding is a clean example of why: the AI answer sounds right and cites real documents, but the operative legal test it states for condition 3 is wrong, and the framing of the conditions as exhaustive is wrong. Neither error is self-revealing.
The team's workflow needs a mandatory step, not a best-practice suggestion, that requires any AI-assisted summary of a specific OECD standard to be verified line-by-line against Section III.11.b of the 2025 text before it moves into a draft deliverable.
Practically, that means the AI is useful for the tasks that sit upstream and downstream of the legal standard itself: identifying which jurisdictions have adopted the Recommendation, flagging the regulatory timeline for a multi-jurisdictional filing, structuring the outline of a defence memorandum, or summarising publicly available authority decisions on the defence. It is not reliable as the source of the standard the authority will actually apply.
For any transaction where the failing firm defence is being seriously evaluated, the team should work from the OECD text directly for condition 3, and the internal memo that frames the available defences should carry a note that the 2025 Revision's comparative harm test is more demanding than the traditional common-law counterfactual formulation that prior training materials or AI tools may default to.
Where the risk is highest is in time-pressured situations: a distressed counterparty emerging late in a financing transaction, a rapid-turnaround regulatory mapping exercise for a client's board presentation, or a junior lawyer producing a first-draft position paper overnight. Those are exactly the situations where AI tools get the heaviest use and where the error described here, a plausible-sounding but wrong statement of a legal standard, backed by citations that don't establish the proposition, is most likely to survive into a deliverable.
Building the primary-text check into the team's standard research protocol for this regulation, rather than treating it as optional, is the control that prevents the error from propagating.
How RLB Can Help
RegLeg's published Hallucination Research gives Corporate Banking Legal teams a concrete pre-flight check before placing weight on AI-assisted regulatory analysis. Rather than relying on internal validation alone, you can cross-reference AI output against a documented catalogue of failure modes, specific instances where AI tools have misrepresented scope, inverted obligations, or confabulated enforcement thresholds across the regulatory frameworks your team works with daily.
That's a faster and more defensible due-diligence step than building your own test suite from scratch, and it gives Legal a defined evidentiary basis for where AI output can be trusted and where independent counsel or a primary-source read is non-negotiable.
Beyond the published research, RegLeg works with Corporate Banking Legal functions to map their specific AI-supported workflows against the hallucination risk profile for the regulatory perimeter they operate in, cross-border capital requirements, sanctions screening obligations, loan documentation standards, and the cross-jurisdictional licensing frameworks that sit at the intersection of prudential and commercial regulation. The output is a prioritised exposure map: which workflows carry material hallucination risk, where the failure modes tend to cluster (entity mis-scoping, numeric threshold drift, temporal applicability errors), and which regulatory instruments have the densest documented failure history.
That scoping exercise informs how Legal allocates human review time and where AI reliance caps should sit.
Where a firm already has an AI-use policy in place, RegLeg can run a confidential review against our failure-mode catalogue to surface gaps, particularly around regulatory research, contract review, and compliance sign-off workflows where Legal teams have the most exposure if AI output is wrong. We can also develop training material and CPD-aligned content calibrated for Legal professionals in Corporate Banking: not generic AI literacy, but practitioner-level material on how specific failure modes manifest in regulatory contexts your team actually encounters, and what a sound review protocol looks like when AI is part of the workflow.