Executive Summary
AI assistants tested against the 2025 OECD Merger Review Recommendation (OECD/LEGAL/0333) produced wrong answers on all four questions put to them, every failure carried the same risk tag: wrong deliverable. For Legal teams at Private Equity & Venture Capital firms navigating multi-jurisdictional merger clearance, that failure profile is not academic: the Recommendation's five-section operative structure, its Section IV remedy hierarchy, and its Section III failing firm defence are precisely the reference points that anchor pre-signing risk assessment, deal condition analysis, and investment committee sign-off.
The dominant pattern was confident fabrication, AI tools invented operative sections (adding a non-existent "Cross-Jurisdictional Co-operation" RECOMMENDS block), imported EU merger-control doctrine as if it were OECD text, and presented an explicitly non-exhaustive failing-firm test as a closed, three-limb cumulative standard. When challenged, the models typically retracted or expressed uncertainty, confirming the initial confidence was unfounded. A team that treats any of these AI outputs as a reliable first-cut reference for deal-level legal work is carrying material structural error into its deliverables from step one.
How AI gets this regulation wrong
The failures on this Recommendation cluster around two modes: AI tools that invented structural content, adding operative sections the Recommendation does not contain and omitting sections it does, and one tool that grafted a well-known regional practice hierarchy onto the Recommendation's remedy text as if it were the operative standard. Both modes share the same surface characteristic: the answers sounded authoritative, were internally consistent, and collapsed only under direct challenge with primary source access. The table below maps each failure to the specific mechanism.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 3 | Finding#1 · Finding#3 · Finding#4 |
| Misattributed | 1 | Finding#2 |
What that means for your team
Every failure in this cell lands in the same category: wrong deliverable. For Legal teams at PE and VC firms, that means deal memos, condition-negotiation briefs, investment committee packs, and multi-jurisdictional clearance analyses built on incorrect legal foundations, errors that compound as downstream teams rely on the initial characterisation without re-checking primary text. The table below breaks down where in the deal and regulatory workflow each failure type would surface and what remediation looks like once the error is discovered.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 4 | Finding#1 · Finding#2 · Finding#3 · Finding#4 |
When this affects your department
Legal teams at PE and VC firms reach for AI-generated summaries of the 2025 OECD Merger Review Recommendation in three recurring contexts: pre-signing jurisdictional mapping (which OECD adherent states are in scope, what procedural obligations apply, how the Recommendation frames notification thresholds and review timelines); deal-condition analysis (whether a proposed structural or behavioural remedy sits within the Recommendation's accepted framework and whether a failing-firm defence is viable for a distressed target); and investment committee and LP briefing materials that synthesise the regulatory risk profile of a cross-border transaction.
In each case, the AI output is typically treated as a trusted first-cut reference that gets embedded into a memo or deck before a senior lawyer reviews it, and "senior review" in a deal environment often means a quick read against memory, not a line-by-line check against OECD primary text.
The consequences of carrying AI error into these workflows differ by context but are uniformly serious. A mischaracterised operative structure (e.g., treating a non-existent "Cross-Jurisdictional Co-operation" section as an operative RECOMMENDS clause) skews a firm's regulatory mapping, leading to gaps in multi-jurisdictional filing checklists or misaligned negotiating positions with co-bidder legal teams in other adherent states.
A wrong remedy hierarchy, importing EU fix-it-first doctrine as OECD standard, would directly mis-set expectations with portfolio company management and advisers about what a competition authority in an OECD adherent state is likely to require and in what priority order, creating both negotiation risk and delay risk at a point in a deal process where every week costs money.
And a misfired failing-firm analysis, treating an explicitly non-exhaustive test as a closed three-limb standard and mischaracterising the third limb's comparative-harm requirement, could lead a deal team to conclude a failing-firm defence is either available or unavailable when the correct analysis points the other way, affecting both pricing decisions and authority engagement strategy.
The cross-jurisdictional dimension amplifies all of this: PE and VC firms running simultaneous filings across multiple OECD adherent states need a consistent, accurate baseline for the Recommendation's framework precisely because local authority practice diverges from it in ways that matter, and that divergence analysis starts from the correct baseline text, not an AI-synthesised version of it.
The findings at a glance
The four findings below cover the questions Legal teams are most likely to direct AI tools at when working with the 2025 OECD Merger Review Recommendation, its operative structure, its Section IV remedy hierarchy, its Section III failing-firm standard, and the scope boundaries of what the instrument actually addresses.
Aggregate impact
The pattern across these four findings is not random error, it is a consistent failure to correctly characterise the 2025 Recommendation's operative structure. Three of the four findings involve AI tools adding a "Cross-Jurisdictional Co-operation" or "Transnational Co-operation" block as if it were one of the five numbered RECOMMENDS sections, while simultaneously omitting Section V's ex-post assessment mandate. This is not a fringe characterisation: it is a structurally coherent but wrong model of the instrument, and it will propagate consistently into any Legal output that relies on AI to map the Recommendation's scope.
For firms that use AI to generate the initial regulatory framework summary, which then shapes all downstream analysis, the error is present from the first document in the deal file.
The Section IV remedy hierarchy failure adds a different layer of risk. The Recommendation's Section IV.3 sets a clear internal priority within structural remedies: standalone-business divestitures first. AI tools instead imported a practice-based timing hierarchy from EU merger control, fix-it-first, then upfront-buyer-with-trustee backstop, then crown jewel, and presented it as the operative text of the OECD standard.
These frameworks are not the same, and a Legal team at a PE firm negotiating conditions with competition authorities in OECD adherent states that apply the Recommendation's framework would be working from the wrong set of expectations if the AI-generated characterisation is not corrected. The misattribution is particularly dangerous because the EU and OECD frameworks look adjacent from a distance; junior Legal team members without deep EU competition law training may not notice the substitution.
The failing-firm finding compounds this: the AI tools correctly identified three conditions but mischaracterised the third (substituting an "assets inevitably exit the market" test for the Recommendation's actual comparative-harm standard, "that the exit of the firm's assets would cause more harm to competition than the merger") and, critically, presented the three conditions as an exhaustive closed test despite the Recommendation's explicit "inter alia" qualifier.
For a PE firm considering a distressed acquisition where a failing-firm defence may be the deal's regulatory clearance strategy, getting the third limb wrong and treating the test as closed could produce a materially incorrect legal opinion at a moment when that opinion is directly affecting deal economics.
What your team should do
The default position for any Legal team at a PE or VC firm should be: do not use AI-generated characterisations of the 2025 OECD Merger Review Recommendation as a working reference without verifying the operative text directly. The errors identified here are not edge-case misreadings, they are structural mischaracterisations of the Recommendation's five numbered RECOMMENDS sections, its Section IV.3 remedy priority rule, and its Section III.11.b failing-firm conditions. Each of those elements is foundational, not marginal.
The fix is not a senior review layer that catches errors on the way out; it is primary-source verification before any AI summary is used to anchor deal analysis.
For structural mapping, understanding the Recommendation's operative scope, what its five sections cover, and what was deliberately excluded or addressed elsewhere, the answer is simply to read the RECOMMENDS clause in OECD/LEGAL/0333 directly. It is a short document. AI tools are not saving meaningful time on this task, and the structural errors they produce (inventing sections, omitting sections) will take longer to correct than the time saved.
For remedy analysis under Section IV, AI tools can usefully surface the broad behavioural-over-structural preference and the standalone-business priority within structural remedies, but any more granular hierarchy should be checked against Section IV.3's text; the EU fix-it-first framework should never be assumed to replicate the OECD standard. For failing-firm analysis under Section III.11.b, AI tools can frame the three core conditions as a starting point, but the "inter alia" qualifier must be front-and-centre in any Legal memo, the test is not closed, and the third condition's comparative-harm framing matters for authority engagement strategy.
Where AI tools are reliably useful in this workflow: searching the OECD competition law library for related instruments (background papers, working party reports, competition committee decisions), drafting initial questionnaires for local outside counsel to map how adherent-state authorities implement the Recommendation's standards, and synthesising publicly available authority guidance on specific procedural requirements (notification thresholds, review periods) that are operationalised at the national level. In each of those uses, the AI is doing retrieval and first-draft structuring rather than characterising the operative standard, which is where the failures concentrated.
How RLB Can Help
RegLeg's published hallucination research functions as a pre-flight check your team can run before placing any weight on AI-assisted output in regulatory matters. The findings are regulation-specific and failure-mode specific, they tell you not just that an AI tool got something wrong on a given instrument, but how it got it wrong and on what category of question.
For a Legal function fielding fund structuring queries, cross-border regulatory-status questions, or investor protection disclosure obligations across multiple jurisdictions simultaneously, that granularity matters: it lets you calibrate reliance by question type rather than applying a blanket haircut to every AI output or, worse, applying no haircut at all.
Beyond the published corpus, we work with PE/VC Legal teams on bespoke regulator deep-dives scoped to their actual workflow exposure. The failure modes that tend to surface in securities-regime interpretation or fund-marketing-rules analysis are not the same as those in prudential liquidity work, and the risk profile shifts again when the question touches GP liability, fiduciary duty framing, or co-investment structural constraints under a specific regulator's framework.
We map which AI-supported tasks in your Legal function carry the highest hallucination exposure, investment-memorandum regulatory sections, LPA compliance sign-offs, regulatory-change horizon scanning, and quantify where the gap between what an AI tool says with confidence and what the instrument actually requires is largest and most costly.
We also offer a confidential review of your firm's existing AI-use policy against RegLeg's failure-mode catalogue, with prioritised remediation recommendations, identifying which policy commitments are under-protected given what the research shows AI tools actually do when pressed on the instruments your team relies on. For firms building or refreshing internal Legal AI governance frameworks, we can develop training material and CPD-aligned content calibrated to the jurisdictions and regulatory domains your team operates in, so that competence standards keep pace with the tools your lawyers are already using.