Executive Summary
Legal teams at mainboard and premium-listed issuers operating across OECD member states reach for the 2025 Recommendation as the reference architecture when mapping multi-jurisdictional merger review obligations, drafting internal M&A governance frameworks, or briefing the board on cross-border regulatory exposure ahead of a significant transaction. Across the two aggregated findings tested on this Recommendation, AI assistants hallucinated on both, producing a confident but structurally incorrect account of the instrument's operative sections.
In both cases, the AI invented a standalone "Cross-Jurisdictional" or "International Co-operation" operative section that does not appear in the Recommendation's five RECOMMENDS clauses, and simultaneously dropped Section V, the ex-post assessment obligation, from its account entirely. When challenged, neither tool corrected to the right answer; both admitted uncertainty rather than providing the accurate structure.
For a legal team using AI output to draft internal M&A policy, jurisdictional screening protocols, or training materials on the 2025 revision, the structural error flows directly into the deliverable and will not surface until a transaction, an audit, or a regulator creates the occasion to check the source text.
How AI gets this regulation wrong
Both failures documented here share the same underlying pattern: AI tools confidently invented a plausible-sounding but nonexistent operative section, framed around cross-border cooperation, while dropping a real one in its place. The invented section was thematically coherent with merger control practice, borrowing from related OECD instruments, which makes it especially resistant to detection without independent reference to the Recommendation's text. The table below shows how this failure mode distributed across the questions tested.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 2 | Finding#1 · Finding#2 |
What that means for your team
Both findings land in the same risk category: a deliverable built on AI's structural characterisation of this Recommendation will be wrong in ways that are invisible until the source text is checked. For a legal team at a mainboard issuer, that deliverable could be an internal M&A governance policy, a board briefing on cross-border regulatory exposure, or training materials for transaction teams, each of which then propagates the error across the firm's deal workflow. The table below maps where that exposure sits.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 2 | Finding#1 · Finding#2 |
When this affects your department
Mainboard issuers with active acquisition programmes, or those subject to incoming bids, use the OECD Merger Review Recommendation as the reference standard when building internal M&A playbooks, defining jurisdictional screening protocols, or briefing the board and transaction committee on multi-jurisdictional regulatory exposure. The 2025 revision is a natural candidate for AI-assisted research precisely because its structural changes from the 2005 version are non-obvious: what was consolidated, what moved into other instruments, and what new operative obligations were introduced.
Asking AI to characterise those changes is exactly the kind of research task where a time-pressured lawyer or a secondee reaches for an AI tool first, often to produce a first-draft summary that will go into a board pack or policy update without independent verification of the instrument's text.
The specific scenario where this failure bites hardest is early in a transaction, when the legal team is scoping multi-jurisdictional filing obligations and building the regulatory chapter of the deal timetable. If the AI-generated structural summary of the Recommendation is used to anchor that chapter, framing the operative obligations the firm must track across adherent states, the firm's deal governance is calibrated to the wrong instrument architecture.
A policy or training document that describes the Recommendation as having six operative sections, with one anchored on cross-border cooperation as a standalone RECOMMENDS clause, will misdirect junior lawyers' attention and produce gaps in any compliance mapping exercise that follows.
The omission of Section V carries a distinct operational risk that the invented section does not. Ex-post assessment of completed merger decisions is a live operative commitment under the 2025 revision. A legal team whose internal framework does not flag Section V has not merely misdescribed the instrument, it has failed to account for the possibility that adherent-state authorities may revisit a completed transaction under this standard. For a mainboard issuer that has closed a cross-border deal in an OECD member state, that gap in the firm's risk map is material and not easily recovered once the deal is done.
The findings at a glance
The two findings below cover the structural questions tested on this Recommendation, both catching AI tools constructing the same phantom operative section while omitting the actual Section V on ex-post assessment.
Aggregate impact
The failure pattern across both findings is strikingly consistent and non-random. AI assistants characterised the 2025 Recommendation's operative structure as containing six RECOMMENDS sections rather than five, anchoring the invented sixth, or in one case fifth, section on cross-border or transnational cooperation. The actual five RECOMMENDS sections address framework maintenance, notification and review procedures, substantive merger analysis, remedies, and ex-post assessment; cooperation is addressed within those sections as a procedural and analytical matter, not extracted into a standalone operative clause.
Section V, ex-post assessment of merger decisions and remedies, is the section that AI systematically dropped to make room for the invented one.
The clustering matters because the invented section is thematically credible. Multi-jurisdictional merger review is inherently a cross-border exercise, and OECD instruments do separately govern competition cooperation, OECD/LEGAL/0408 being the relevant instrument. An AI that constructs a plausible-sounding six-section architecture by borrowing thematic language from the surrounding OECD competition law corpus will produce output that reads correctly to a practitioner who is not independently checking OECD/LEGAL/0333. The error does not look like noise; it looks like a reasonable structural summary of an instrument the reader expects to address cross-border cooperation.
That is precisely what makes it dangerous in a legal drafting or advisory context.
The omission of Section V is the sharper operational risk. A legal team that produces a structural summary of the Recommendation, for internal policy, board briefing, or training, without flagging the ex-post assessment obligation has built an incomplete risk map. The 2025 revision's inclusion of Section V signals that adherent-state authorities can revisit completed transactions and remedies; a firm whose merger governance framework does not account for this exposure is not operating with an accurate picture of the Recommendation's operative scope.
That gap will not be visible in the deliverable itself, it will surface when a post-closing review by an adherent-state authority raises an issue the firm's legal team was not tracking.
What your team should do
Do not use AI output as a substitute for direct review of OECD/LEGAL/0333 when drafting any document that characterises the 2025 Recommendation's operative structure, the scope of its RECOMMENDS clauses, or the changes from the 2005 version. The fabrication pattern documented here is not an edge case, it appeared across multiple AI tools on the most fundamental structural question about the instrument. Any internal policy, board paper, or training document that purports to describe the Recommendation's operative sections must be verified against the source text before sign-off, regardless of how plausible the AI's account reads.
For jurisdictional screening exercises, mapping which adherent states require notification for a proposed transaction, what thresholds apply, and what the deal timetable looks like, AI tools remain useful for initial scoping and for gathering publicly available national-level merger control rules. The failure mode documented here is specific to characterising the OECD Recommendation's own operative framework, not to jurisdiction-specific merger control legislation. Teams can continue to use AI for the latter while applying a direct-source check to the former.
The practical safeguard is simple and should be built into the team's standard operating procedure for any M&A regulatory mapping exercise: treat AI output on the structure of OECD instruments as a hypothesis to be verified, not a citation. Where the deliverable will be relied on by the board, transaction counsel, or external regulators, require a lawyer with direct access to OECD/LEGAL/0333 to confirm the structural description before the document is finalised.
This is not a general disclaimer about AI reliability, it is a specific, documented failure mode on this instrument, confirmed across multiple AI tools, that should be briefed to any junior or secondee conducting merger control research using AI assistance.
How RLB Can Help
RegLeg's published hallucination research is available without charge as a pre-flight check your team can run before relying on AI-assisted output on specific regulatory questions. If your Legal function is already using AI tools to triage disclosure obligations, review continuous reporting timelines, or cross-reference listing rule requirements across jurisdictions, the research gives you a structured view of where those tools have demonstrably failed on the same regulatory texts, before those failures surface in a board pack or an exchange submission.
That is a different kind of assurance than a vendor's accuracy claim: it is adversarial evidence, publicly documented, tied to specific instruments.
Beyond the published corpus, we work with Legal teams on bespoke regulator deep-dives scoped to the workflows that carry the heaviest hallucination exposure for a Mainboard or Premium-Listed Issuer. In practice that tends to cluster around multi-jurisdictional disclosure synchronisation, prospectus and circular drafting where AI tools are being used to check completeness against exchange and securities law requirements, and sanctions and materiality thresholds where the regulatory text changes faster than AI training cycles.
The output is a prioritised map of where your team's current AI-assisted processes are operating on ground that the research shows to be unreliable, not a generic risk register, but something scoped to your document types and counterparty regulators.
For teams that have an existing AI-use policy, we offer a confidential review against RegLeg's failure-mode catalogue, with a prioritised remediation list rather than a broad gap analysis. We also produce CPD-aligned training material Legal leads can deploy internally, briefing packs and worked examples built around the failure patterns most relevant to listed-issuer counsel, usable for team training, partner sign-off workflows, or regulatory engagement where your firm needs to demonstrate governance around AI tool use.