Executive Summary
Across five questions put to AI tools about the OECD's 2025 Merger Review Recommendation (OECD/LEGAL/0333), every single response produced a hallucination, a rate that should give any Legal team at an international law firm serious pause before using AI-generated analysis of this instrument in client-facing work.
The failures are not superficial: AI tools consistently miscounted and mislabelled the Recommendation's operative RECOMMENDS sections, invented a standalone "International Co-operation" or "Transnational Co-operation" block that does not exist in the text, and in the same breath omitted Section V's ex-post assessment obligation, which is the provision most likely to drive future enforcement convergence among Adherents.
On the remedies side, AI imported the EU merger-control fix-it-first hierarchy directly into the OECD framework, presenting EU practice as OECD text; and on the failing firm defence, AI collapsed the Recommendation's deliberately open-ended "inter alia" standard into a closed three-condition exhaustive test, stripping the qualifier that leaves authorities discretion to demand additional evidence. For law firms advising multi-jurisdictional clients on OECD-aligned regimes, these are the structural errors most likely to travel undetected into opinion letters, merger control matrices, and jurisdictional coverage memos.
How AI gets this regulation wrong
The dominant pattern across this regulation is AI confidently constructing answers that look structurally correct, right number of clauses, plausible subject-matter headings, appropriate regulatory vocabulary, but whose substantive content is either invented or cross-contaminated from other frameworks. In the most consequential instance, AI cited real OECD sources to support a remedies hierarchy it had actually imported wholesale from EU merger practice, meaning the citations appear to validate analysis that those sources do not in fact contain.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 4 | Finding#1 · Finding#3 · Finding#4 · Finding#5 |
| Misattributed | 1 | Finding#2 |
What that means for your team
For a law firm's Legal team, the preponderance of risk sits with professional indemnity and client-liability exposure: four of the five failures concern substantive advice on the Recommendation's operative architecture or defence standards, where an AI-generated error that reaches an opinion letter or a client briefing note creates a direct line to a negligence claim. The remaining failure, mischaracterising the Competition Committee's reporting cycle, carries lower immediate commercial stakes but would produce a materially wrong deliverable in any regulatory mapping exercise or compliance timeline that a client is relying on.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Liability / PI exposure | 4 | Finding#1 · Finding#2 · Finding#4 · Finding#5 |
| Wrong deliverable | 1 | Finding#3 |
When this affects your department
International law firms encounter the 2025 OECD Merger Review Recommendation at three recurring points. The first is regulatory mapping on multi-jurisdictional transactions, where competition counsel needs to characterise the OECD framework's operative obligations against local implementing rules, particularly in Adherent jurisdictions that have amended their merger control regimes in response to the 2025 revision. The second is advising clients on the substantive standard for clearance or remedy negotiation: what the Recommendation says about structural versus behavioural remedies, how the failing firm defence is framed at the international-benchmark level, and whether an authority's domestic approach diverges from that benchmark.
The third is benchmarking and training: senior associates and younger partners working on merger control in multiple jurisdictions routinely use the Recommendation as the reference architecture for how a sound merger review framework is supposed to be structured.
Where AI errors bite hardest is in the first two contexts. If a junior lawyer uses an AI tool to draft the "OECD Recommendation framework" section of a multi-jurisdictional merger control matrix and the AI invents a standalone "Cross-Jurisdictional Co-operation" operative section, or omits the ex-post assessment obligation entirely, that error propagates directly into the client deliverable without the partner reviewing OECD/LEGAL/0333 against the AI's characterisation.
On failing firm defence work, the difference between an open-ended "inter alia" standard and a closed three-condition test is not cosmetic: it directly affects the advice a firm gives a client about the evidentiary threshold and the risk of a challenge failing on a condition the authority raises that isn't in the AI's closed list.
The remedies mischaracterisation is the most litigation-adjacent risk. AI presenting the EU fix-it-first hierarchy as OECD text could cause a firm to benchmark a client's remedy proposal against the wrong standard in a non-EU jurisdiction that follows the OECD framework rather than the EU Merger Regulation practice. The harm is not just a wrong answer, it is an answer supported by real OECD citations that appear to validate it, making downstream verification less likely and the PI exposure worse if the error reaches an opinion.
The findings at a glance
The table below summarises each verified AI failure on the 2025 OECD Merger Review Recommendation, together with the specific provision involved and the risk category most relevant to a law firm's Legal team.
Aggregate impact
The errors across this regulation share a common structural feature: AI tools consistently produced answers that were internally coherent and surface-plausible but were calibrated to a version of the instrument that does not exist. The most telling example is the repeated invention of a "Co-operation" operative section, two separate AI tools, independently, produced a Recommendation with six operative RECOMMENDS blocks rather than five, inserting international co-operation as a standalone section and dropping Section V's ex-post assessment obligation.
This is not a single bad answer on a hard question; it is convergent confabulation on the basic architecture of the instrument, which suggests the model's representation of OECD/LEGAL/0333 in training data is systematically incomplete or cross-contaminated with the superseded 2005 version.
The remedies finding adds a different risk dimension: cross-framework contamination. The EU fix-it-first hierarchy is a genuine and well-documented practice, which is exactly what makes it dangerous when an AI presents it as the OECD text. The fabricated answer is not implausible to a reader who knows EU merger practice, it is the kind of answer a senior EU competition lawyer might give if asked to describe "standard" structural remedy priorities.
For a firm advising on a deal in an OECD Adherent jurisdiction that is not in the EU, that contamination produces advice that is wrong about the applicable benchmark standard and simultaneously very difficult for a junior to identify as wrong.
The failing firm defence findings compound this: the "inter alia" qualifier in Section III.11.b is the provision's load-bearing hedge. Its removal, and its replacement with a closed cumulative list, changes the practical advice to a client about evidentiary risk. Across the five findings, the Legal team at a law firm faces a consistent pattern of AI answers that would pass a plausibility check but would fail a primary-source verification. The practical implication is that any AI-assisted work on this instrument needs a mandatory primary-source review step for every substantive proposition, not just a sanity check on structure.
What your team should do
The default position for any AI-assisted work on OECD/LEGAL/0333 should be that AI can draft structure, suggest comparator frameworks, and flag issues, but the operative text of every RECOMMENDS section must be verified against the published instrument before it leaves the Legal team's hands. This is not a general "check your sources" platitude: the specific risk here is that AI confabulations on this instrument are structurally plausible and well-formatted, which reduces the probability that a reviewer catches the error without actually pulling the text. Build the primary-source verification step into the workflow, not the review step.
For the specific failure modes on this instrument: on the Recommendation's operative structure, treat any AI-generated section list as unverified until compared against the five RECOMMENDS sections (I–V) in the published text. On remedies, any AI answer that describes a tiered internal hierarchy within the structural remedies category (fix-it-first, upfront buyer, crown jewel) should be flagged as likely EU practice cross-contamination and checked against Section IV.3's actual text before use. On the failing firm defence, check that any AI-generated conditions list preserves the "inter alia" qualifier, if the AI presents a closed exhaustive test, that is the error signal.
On reporting timelines, verify the two-tier cycle (five years initial, ten years thereafter) against Section VIII.c directly.
AI tools are genuinely useful for background orientation on the broader merger control context, understanding how the Recommendation sits alongside other OECD competition instruments, identifying which Adherent jurisdictions have implemented specific provisions, or drafting the non-substantive sections of a regulatory mapping memo. The risk is concentrated in the substantive operative content, particularly the structure, remedies framework, and defence standards. Keep AI in the drafting seat for scaffolding and orientation; keep qualified competition lawyers in the seat for any proposition about what the instrument actually requires.
How RLB Can Help
RegLeg's published Hallucination Research gives Legal teams at law firms a ready pre-flight check before placing weight on AI-assisted output in regulatory matters. Each research entry documents a confirmed failure mode against a specific instrument, the type of provision involved, how the AI went wrong, and the risk consequence, so lawyers can run a quick cross-reference against the regulation they are working with before finalising advice, drafting submissions, or briefing clients. The research is freely available and requires no engagement to access.
For firms that want to go further, RLB offers bespoke regulator deep-dives scoped to the specific bodies and instruments your Legal function works with most. These engagements map which AI-supported workflows, regulatory research, precedent checking, cross-border compliance comparison, client advice drafting, carry the highest hallucination exposure in your practice context, and produce a ranked risk register the team can act on immediately. The output is confidential and is tailored to the jurisdictions and regulatory perimeters your firm operates across.
RLB also conducts confidential reviews of existing AI-use policies against its failure-mode catalogue, identifying gaps between the controls a firm has documented and the classes of error its AI tools are most likely to produce on regulatory questions. Each review closes with a prioritised remediation plan. Alongside policy work, RLB can supply training materials and CPD-aligned content, structured around real failure cases, that Legal teams can deploy internally to build consistent, defensible AI literacy across practice groups and seniority levels.