Frontier AI models inflated OECD's merger-review structure, regulatory-research panel finds

Two frontier AI models, with web search enabled, added operative sections, sub-tiers, internal priority orderings, and fixed reporting dates to the 2025 OECD Merger Review Recommendation that the instrument does not contain, confidently, and without caveat. The RegLeg Brief Specialist Panel calls the pattern "Structure Inflation" and says it points to a calibration problem in how recently revised soft-law instruments are reconstructed from prior-generation document shapes.

SINGAPORE, June 11, 2026. Two frontier artificial-intelligence models generated structural and substantive guidance on the 2025 revision of the OECD Recommendation on Merger Review (OECD/LEGAL/0333) that contradicts the instrument's text in ways that would push competition-law practitioners off-piste if relied on, according to seven findings released today by RegLeg Brief, a regulatory-research outfit operated by Singapore-incorporated Verdus Technologies Pte. Ltd.

The seven findings, all bound to substrate document R1-REGULATION-00001 (the OECD Recommendation as adopted in 2025) and one supporting OECD guideline (R3-GUIDELINE-00004), concern two of the most heavily relied-on portions of the Recommendation in competition-law practice: the operative section structure (which counsel cite to anchor a position) and Section III.11.b on the failing firm defence (which counsel invoke to defend transactions involving distressed targets). Both Anthropic's Claude Opus 4.7 and Claude Sonnet 4.6 were tested with web search active, mirroring how competition counsel, in-house antitrust teams, and consultancy practices actually use the models today.

The pattern in one line

Across seven findings, both models supplied more structure than the OECD did, adding operative sections, internal sub-orderings, fixed dates, and closed cumulative tests to an instrument the OECD published with five operative sections, one stated remedies preference, an open reporting cadence, and "inter alia" evidentiary criteria.

How the testing worked

The RegLeg Brief Specialist Panel posed seven probes drawn directly from the published Recommendation: how the instrument is structured (Q001), what Section IV.3 says about remedies hierarchy (Q002), what Section VIII.c requires on Council reporting (Q004), what conditions the failing firm defence requires under Section III.11.b (Q005), and what the primary operative sections actually address (Q006). Each probe was run against Claude Opus 4.7 and Claude Sonnet 4.6 with web search active, then verified against the OECD-published text held in substrate documents R1-REGULATION-00001 and R3-GUIDELINE-00004.

Findings were published only where the model output contradicted the verbatim text; where outputs were correct, the panel published nothing.

What the models got wrong

The verbatim instrument the panel measured against: the 2025 Recommendation enumerates five operative sections, Section I (maintain framework), Section II (notification and review procedures), Section III (sound merger analysis), Section IV (clear remedies framework), and Section V (ex-post assessment). Council reporting and Competition Committee monitoring sit in Section VIII.c, not as a sixth operative section.

Structure inflation, six sections invented (Q001 Opus 4.7 and Q001 Sonnet 4.6). Claude Opus 4.7 wrote, verbatim: "OPERATIVE CONTENT (as characterised by the OECD): (a) MERGER REVIEW FRAMEWORK ... (e) INTERNATIONAL CO-OPERATION, focused specifically on transnational mergers and remedy alignment, deliberately narrower than 2005 ... (f) MONITORING and review by the Competition Committee." Sonnet 4.6 produced a parallel six-item list. Neither "International Co-operation" nor "Monitoring" is an operative section of the 2025 Recommendation. The "deliberately narrower than 2005" qualifier is editorial commentary that does not appear in the Recommendation or in the OECD's accompanying explanatory material.

Both findings classified as inference_drift against R1-REGULATION-00001.

Structure inflation, fabricated cross-reference (Q006 Opus 4.7). Separately probed, Opus 4.7 produced the same six-section reconstruction with an additional invention: a fabricated cross-reference to OECD/LEGAL/0408 (2014) as the instrument now governing the cooperation language the Recommendation supposedly delegated out. Classified as inference_drift against R3-GUIDELINE-00004.

Remedies hierarchy, internal sub-ordering invented (Q002 Sonnet 4.6). Asked what Section IV.3 establishes within structural remedies, Sonnet 4.6 generated: "Within structural remedies, there is a further internal preference ordering: Upfront / 'fix-it-first' divestitures...rank highest...Divestiture commitments with an approved buyer pool and a trustee mandate as backstop rank second. 'Crown jewel' or ring-fenced asset packages...rank third." Section IV.3 provides one preference: structural over behavioural, and, within structural, the divestiture of standalone businesses. There is no upfront-vs-buyer-pool-vs-crown-jewel internal ranking. The three-rank construct is an EU Commission practice convention, not OECD text. Classified as misattributed against R1-REGULATION-00001.

Reporting cadence, fixed-cycle dates fabricated (Q004 Sonnet 4.6). Asked what Section VIII.c requires of the Competition Committee, Sonnet 4.6 produced: "Fixed interval: 5 years (next report: 2030; following report: 2035)." Section VIII.c sets two distinct intervals: "no later than five years following its revision and at least every ten years thereafter." A first report by 2030 is consistent with the text. A second report fixed at 2035 is not, the ten-year minimum interval after the first means the next report can fall anywhere up to 2040. Classified as misstated_rule against R1-REGULATION-00001.

Failing firm defence, "inter alia" converted to closed cumulative test (Q005 Opus 4.7 and Q005 Sonnet 4.6). Opus 4.7: "THREE cumulative conditions ... (3) ABSENT-THE-MERGER COUNTERFACTUAL: ASSETS WOULD EXIT." Sonnet 4.6: "Under the OECD standard, a failing firm defence requires the merging parties to demonstrate all three of the following conditions... All three conditions must be satisfied simultaneously.

Failure on any one condition is fatal to the defence." Section III.11.b uses "inter alia" when introducing the evidentiary criteria, the verbatim text begins: "require, inter alia, evidence that the business was likely to have exited and the existence of less anti-competitive alternative buyers or other options for reorganisation are not viable." "Inter alia" signals an illustrative, non-exhaustive list. The Recommendation does not say there are exactly three conditions, does not label them, and does not impose all-or-nothing pass/fail framing. Classified as inference_drift (Opus 4.7) and misstated_rule (Sonnet 4.6) against R1-REGULATION-00001.

Why this matters for competition policy and merger review

OECD soft-law instruments are not directly binding, but they shape national authority practice across the OECD membership. Authorities routinely cite the Recommendation when explaining their own remedies hierarchies, evidentiary expectations, and case-review structures. Counsel cite the Recommendation in submissions to authorities to anchor positions on what "the international standard" is. Academic and policy commentary uses the Recommendation as a referent for cross-jurisdictional benchmarking.

When a frontier model rewrites the Recommendation's enumerated structure, invents internal hierarchies, fixes dates the instrument leaves open, or converts open evidentiary language into closed cumulative tests, the downstream content propagates the rewrite. The exposure points are concrete:

• Client memos and regulatory updates: cite operative provisions that do not exist; carry a narrative that the OECD made a structural narrowing decision the OECD did not document.
• Remedies playbooks: inform a client that the OECD endorses a specific upfront-divestiture-first internal ordering, which is a defensible position to take, but the OECD did not take it in the 2025 Recommendation.
• Instrument-update tracking and resource planning: put 2035 in a calendar tool for an instrument the OECD has explicitly placed on a five-then-ten cadence.
• Failing firm defence submissions: under-prepare on additional evidentiary lines the authority could raise under "inter alia"; over-state the OECD's articulation of the standard when arguing for alignment.

The Specialist Panel's observation across the seven findings is that the model outputs were internally consistent and presentationally polished enough that nothing in the output itself flagged a need to verify against primary source.

The OECD's actual position

For each of the seven findings, the regulator's verbatim text is published alongside the model output on the per-finding cards. The OECD Recommendation is a public OECD/LEGAL instrument; the substrate text used for verification is the published Recommendation as adopted in 2025. Where the Recommendation is silent (on the internal remedies ordering, on a fixed Council reporting calendar with 2035 dates, on a closed cumulative failing firm test), the verbatim excerpt makes that silence visible. The Specialist Panel does not characterise the OECD's position beyond what the published text says.

What this tells us about AI under regulatory query

The failure shape is consistent across all seven findings: the models supplied more structure than the regulator did. They added sections that were not there, sub-tiers within sections, internal orderings within sub-tiers, fixed dates within open intervals, and closed cumulative tests within "inter alia" lists. The Specialist Panel's working hypothesis is that frontier models, when probed on recently revised soft-law instruments, default to the document shape of comparable mature instruments (longstanding OECD instruments, EU merger guidelines, mature national merger rules), and reconstruct the recently revised instrument to that shape.

The 2025 OECD Recommendation, having been recently revised away from the 2005 version's structure, is a particularly exposed target for this failure mode: the models had training-data exposure to commentary describing what the 2025 revision changed, but did not, in the model outputs the panel reviewed, have a stable hold on the revised instrument's actual text. Web search being active in each test did not correct the failure. In several findings, the model output reads as if web search returned relevant source material that the generation pathway then overrode with the prior-generation document shape.

This is consistent with the calibration pattern the Specialist Panel has documented in other recently-revised-instrument settings, including the recently-finalised CFTC margin rule documented separately.

What RegLeg is doing about it

All seven findings are published as open-access per-finding cards at reglegbrief.com, with the model output, the regulator's verbatim excerpt, the immutable RLB Citation ID, and the substrate document binding visible on each card. The methodology requires a verbatim regulator excerpt for every published claim, no claim is published against placeholder or paraphrased substrate. Where a finding sits against R1-REGULATION-00001, that is the OECD-published Recommendation; where it sits against R3-GUIDELINE-00004, that is the supporting OECD guideline used to verify the cross-reference claim.

AI labs and model developers named in any published finding have an unconditional right of reply, the Specialist Panel will publish any factual correction or contextual response alongside the original finding, with no editorial gatekeeping. Competition authorities, in-house antitrust teams, and academic researchers with questions on methodology or specific findings can reach the Specialist Panel via the same channel. RegLeg Brief operates as a completely ungated, open-access public resource: the per-finding cards, regulator verbatim excerpts, RLB Citation IDs, methodology notes and supporting data logs are all published without paywalls, registration walls, or data-licensing fees.

Implications for AI labs

For AI labs and alignment teams, the Specialist Panel has made five open-access probe designs from this finding set available for any frontier model team to run, with no commercial engagement required:

1. Operative-section enumeration count for recently revised soft-law instruments. Test whether the model retrieves the actual section count or reconstructs a comparable-instrument shape. The OECD Merger Review Recommendation 2025 (five operative sections) is a documented baseline.
2. Closing-provision vs operative-provision discrimination. Test whether the model can distinguish content sitting in instrument closing provisions (monitoring, reporting, review) from operative provisions. Inflation of closing-provision content into operative sections is a documented failure shape.
3. "Inter alia" preservation in legal text probes. Test whether the model preserves the open-list signal of "inter alia" / "including" / "such as" language, or converts it to a closed enumeration with all-or-nothing cumulative framing.
4. Open-interval vs fixed-cycle preservation. Test whether the model preserves "no later than X and at least every Y thereafter" cadences, or invents a fixed-cycle calendar with specific dates. The Section VIII.c "five then ten" cadence is a documented baseline.
5. Internal sub-ordering invention within hierarchies. Test whether the model preserves the hierarchy a regulator actually stated, or invents an internal ordering within a tier (upfront-first, buyer-pool-second, crown-jewel-third constructs from competition-law practice are a documented failure shape against OECD Section IV.3).

Directions an AI lab might consider, drawing on the seven findings:

• Targeted correction pairs: the regulator's verbatim text for each of Sections I through V, Section III.11.b, Section IV.3, and Section VIII.c of the 2025 OECD Recommendation, matched to the wrong-but-plausible reconstructions documented in each finding, suitable for direct ingestion into a training-data pipeline.
• Quarterly embedded eval cycles: continuous evaluation against a defined OECD-instrument portfolio (and adjacent recently revised soft-law instruments), with regression monitoring on previously documented Structure Inflation failure modes.
• Pre-release evaluation cycles: sandboxed probes against the five failure shapes above for capability releases touching antitrust, competition policy, or cross-border regulatory content, before the release reaches enterprise customers.
• Post-release monitoring: regulator-specific failure-surface monitoring as new OECD/LEGAL instruments are adopted or revised, ingesting the verbatim text within hours of adoption.

Primary source verified: OECD/LEGAL/0333, Recommendation of the Council on Merger Review (2025 Revision) · Substrate documents: R1-REGULATION-00001, R3-GUIDELINE-00004 · OECD portal: oecd.org/legal

Citation IDs referenced: - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q001-Opus47 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q001-Sonnet46 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q002-Sonnet46 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q004-Sonnet46 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q005-Opus47 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q005-Sonnet46 - RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q006-Opus47

Right of Reply

These findings and associated work have been put up in public with a view of the greater good for the development of a safer AI ecosystem. Any party reading this or any finding on reglegbrief.com may contact us and have an unconditional right of reply; the Specialist Panel will publish any factual correction or contextual response alongside the original finding, with no editorial gatekeeping. Researchers, regulators, and compliance teams with questions on methodology or specific findings can reach the Specialist Panel via the same channel.

Source & Methodology Standards

RegLeg Brief is operated by Verdus Technologies Pte. Ltd. (UEN 201616982R), incorporated in Singapore. The RLB Specialist Panel, with an aggregate of over 60 years of public-policy and industry experience, documents only confirmed hallucination findings, under a methodology that requires a verbatim regulator excerpt for every documented claim. All findings, citation IDs, model outputs, regulator excerpts, and methodology notes are open-access.

Primary source verified: OECD/LEGAL/0333, Recommendation of the Council on Merger Review (2025 Revision) · Substrate documents: R1-REGULATION-00001, R3-GUIDELINE-00004 · OECD portal: oecd.org/legal

Citation IDs referenced:

• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q001-Opus47
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q001-Sonnet46
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q002-Sonnet46
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q004-Sonnet46
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q005-Opus47
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q005-Sonnet46
• RLB-H-INT-OECD-OECD-MERGER-REVIEW-RECOMMENDATION-2025-Q006-Opus47