Executive Summary
Legal teams at Digital Platforms & Marketplaces firms navigating the OECD's 2025 Merger Review Recommendation face a specific and consequential AI failure: when asked to articulate the remedy hierarchy under Section IV.3, AI tools substituted a three-sub-tier ordering drawn from EU merger-control practice, fix-it-first, upfront buyer pool, crown jewel, in place of the Recommendation's actual standalone-business preference within structural remedies.
Across the questions we tested, AI got this wrong in a way that is particularly hard to catch: the imported framework is coherent, internally consistent, and drawn from a real and authoritative source, so a junior reviewing the AI's output has no obvious signal that the answer describes a different jurisdiction's practice rather than the operative OECD text.
For Legal functions that use AI to prepare internal policy frameworks, merger-readiness materials, or regulatory mapping documents ahead of a transaction, this failure means the firm may instruct its deal team on a remedy priority ordering the OECD Recommendation does not actually require, and may do so with false confidence grounded in a misattributed citation.
How AI gets this regulation wrong
The failure pattern we found on this Recommendation is not AI inventing rules wholesale, it is AI accurately citing a real, authoritative source, then silently attributing that source's content to the OECD text when it does not belong there. The table below sets out the specific instance: a well-structured remedy hierarchy that reads as though it comes directly from Section IV.3 but is in fact imported from a different jurisdiction's merger-control practice.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Misattributed | 1 | Finding#1 |
What that means for your team
The risk that materialises here is not an abstract legal inaccuracy, it is a wrong deliverable: a framework document, internal briefing, or deal-team guidance note that encodes a remedy hierarchy the Recommendation does not prescribe. The table below maps that risk to the Legal team's specific exposure under this Recommendation, in the context of a Digital Platforms & Marketplaces firm operating across international jurisdictions.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Wrong deliverable | 1 | Finding#1 |
When this affects your department
Legal teams at Digital Platforms & Marketplaces firms are not passive consumers of the 2025 Merger Review Recommendation, they are the internal authority that determines how the firm positions itself in multi-jurisdictional review processes. When a platform faces a notifiable transaction, the Legal team typically needs to map the operative remedy framework across the jurisdictions in scope, identify where the OECD Recommendation's principles align with or diverge from domestic regimes, and prepare internal guidance for the deal team on what remedies the firm can credibly offer and in what priority order.
AI tools get pulled into this workflow at the scoping and drafting stages: a junior lawyer might use AI to generate a first-pass framework summary, a policy team member might use AI to check the text of Section IV.3 before building a slide deck, or an M&A support team might query AI to understand how the structural/behavioural hierarchy interacts with commitments they have previously offered in EU or US proceedings.
The specific failure identified here strikes at precisely that scoping moment. Section IV.3 establishes a clear two-level priority: structural over behavioural, and within structural, standalone-business divestitures first. AI tools we tested replaced that second-level priority with a timing-based three-sub-tier hierarchy, fix-it-first, upfront buyer pool with trustee backstop, crown jewel, that accurately describes EU merger-control practice but does not appear in the Recommendation. A Legal team building a multi-jurisdictional merger-readiness framework from AI-generated output would embed this misattributed ordering into internal guidance.
The deal team would then enter commitments discussions with a remedy ladder that is coherent and defensible in Brussels but does not reflect what the OECD text actually says, and may not reflect what OECD member-jurisdiction competition authorities expect when they apply the Recommendation's principles domestically.
For a Digital Platforms & Marketplaces firm specifically, the stakes are elevated by the sector's structural characteristics. Platform mergers regularly attract multi-agency review across OECD member jurisdictions, and remedy discussions frequently turn on exactly the kind of structural-remedy granularity that Section IV.3 addresses. If the firm's Legal team has produced internal guidance misstating that hierarchy, guidance that has been signed off and circulated across deal, strategy, and external counsel teams, correcting it mid-review creates delay, internal credibility damage, and a gap between what the firm's internal documentation says and what its external submissions to competition authorities reflect.
The cost is not a regulatory penalty for citing the wrong paragraph; it is the operational and reputational cost of a correctable error that should not have reached deal-team briefing materials.
The findings at a glance
The table below summarises the one finding identified for Legal teams at Digital Platforms & Marketplaces firms under the 2025 Merger Review Recommendation, the question tested, the nature of the AI failure, and the risk classification.
| # | Finding title | Type | Citation ID |
|---|---|---|---|
| 1 | Section IV.3 structural-remedy hierarchy misattributed to EU practice | Hallucination | RLB-F-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q002 |
Aggregate impact
With one finding in this cell, there is no cross-finding pattern to aggregate, but the character of that single failure deserves direct attention from Legal teams, because it is the category of error most likely to pass internal review undetected. The AI did not fabricate a provision. It did not misread a number or invert a threshold. It accurately described a real remedy hierarchy used in a major, well-documented merger-control regime, and it presented that description as though it were the operative text of Section IV.3 of the OECD Recommendation.
The output was coherent, properly structured, and sourced to a genuine body of practice, which is precisely what makes it dangerous. A Legal reviewer who knows EU merger practice well would have no reason to flag the AI's three-sub-tier framework as wrong; it matches their prior experience. The error is only visible to someone who reads Section IV.3 of the Recommendation itself and notices what it does not say.
The systemic risk for a Digital Platforms & Marketplaces Legal team is that this failure mode is invisible to normal quality-control processes. Peer review within the Legal team may not catch it if the reviewers share the same EU practice background. External counsel may not flag it if they are asked to review a completed framework document rather than the underlying regulatory text. And the error propagates: once it enters an internal guidance note or a merger-readiness playbook, subsequent team members treat it as established internal authority.
The misattributed sub-tier hierarchy becomes the firm's working understanding of what Section IV.3 requires, and it stays there until the firm is in an active review and a competition authority's position forces the discrepancy into the open.
For firms operating across multiple OECD member jurisdictions, which is the normal posture for a scaled Digital Platforms & Marketplaces business, the compounding risk is that internal guidance built on a misread of the OECD Recommendation may be used to calibrate expectations across domestic regimes that have incorporated or adapted the Recommendation's principles. A remedy strategy scoped against a wrongly understood hierarchy is not just an internal documentation problem; it is a deal-room problem that surfaces at the worst possible moment.
What your team should do
The default position for Legal teams using AI on this Recommendation should be: AI is not a substitute for reading the primary text. That is not a generic caution about AI, it is a specific instruction for Section IV.3, because the AI failure we identified is only detectable by comparing the AI's output against the Recommendation's actual language. The fix-it-first / upfront buyer pool / crown jewel hierarchy the AI produced is internally plausible and practice-consistent enough that no amount of AI self-review or output polishing will surface the error.
The check has to happen at the source: someone on the Legal team who has read Section IV.3 needs to confirm that what the AI describes is what the text says, before the output enters any internal document.
The practical safeguard is a two-step gate for any AI-assisted work on remedy framework questions under this Recommendation. First, ask the AI to quote the exact text of the provision it is relying on, do not accept a paraphrase. Section IV.3's standalone-business preference is short and unambiguous; an AI that cannot reproduce it verbatim is drawing on something other than the Recommendation.
Second, before any framework document is circulated to a deal team or signed off as internal policy, require a direct comparison against the OECD text by a qualified lawyer, not a re-read of the AI output, but a side-by-side check against OECD/LEGAL/0333. This is a one-time verification step that takes minutes and eliminates the category of error identified here entirely.
AI tools remain genuinely useful in this regulatory space for tasks that do not depend on precise remedy-hierarchy characterisation: mapping which OECD member jurisdictions have enacted implementing legislation, summarising procedural notification requirements, identifying where domestic regimes diverge from the Recommendation's general principles, or drafting background sections of internal briefings that do not make operative claims about Section IV.3's priority ordering. The failure identified here is specific and narrow, it is about the internal sub-tier hierarchy within structural remedies.
Keep AI output out of any document that makes a direct claim about what Section IV.3 requires at that level of specificity, unless the claim has been verified against the primary text.
How RLB Can Help
RegLeg's published Hallucination Research gives your Legal team a concrete pre-flight check before relying on AI output for regulatory analysis. The findings catalogue specific failure modes, misattributed obligations, wrong jurisdictional scope, fabricated enforcement precedents, drawn from the same regulatory texts your team works against. For a Legal function at a digital platforms firm operating across multiple jurisdictions, where a single misread on DSA extraterritoriality or GDPR-adjacent e-commerce liability can cascade into enforcement exposure, this is the kind of ground-truth reference that belongs in your AI-use review process before any output reaches a memo, a board pack, or a regulator.
Beyond the published research, RLB conducts bespoke regulator deep-dives scoped to your firm's specific AI-supported workflows. For Legal teams at digital platforms and marketplaces businesses, the highest-exposure workflows tend to cluster around content-liability horizon-scanning, cross-border data-transfer opinion drafting, and competition-law self-assessment under frameworks like the DMA or emerging digital markets regimes in APAC and LatAm. RLB maps which of those workflows carry the greatest hallucination risk given the current state of AI tools, so your team can calibrate reliance, and document that calibration, rather than discovering gaps through an adverse finding.
For firms that have already deployed AI tools internally, RLB offers a confidential review of your existing AI-use policy against the failure-mode catalogue. That review identifies where your policy's risk controls align with documented failure patterns and where they do not, and produces a prioritised remediation list your Legal and Compliance leads can act on directly. RLB can also develop training material and CPD-aligned content, drawn from real hallucination case examples rather than generic AI-risk frameworks, that your Legal team can use in internal sessions or as part of structured professional development records.