Sonnet catches the cross-connections where BBNJ biodiversity text meets AI confabulation in legal.
— RLB Specialist Panel
Frontier AI models invert the BBNJ Agreement's retroactivity default and mis-cite the treaty articles.
Two frontier AI subjects tested by the RLB Specialist Panel produced a fundamentally inverted reading of Article 10(1) on marine genetic resources, alongside misattributed citations for the environmental impact assessment screening threshold and the digital sequence information benefit-sharing duty.
Frontier AI models tested on the BBNJ Agreement returned answers in which the substantive direction of one provision was inverted and the article-citation for one or more others was wrong, producing deliverables that would fail first-reading review by a regulator, counterparty, or treaty-body reviewer working with biotechnology firms.
The questions in this cell were prepared by the RLB Specialist Panel based on real, practical AI usage in the workflows that legal teams at biotechnology firms actually use AI for under the BBNJ Agreement. Each question targets a specific deliverable type where an AI assistant is plausibly the first draft: a memo, an opinion paragraph, a checklist line, a board-paper bullet, a regulator-facing filing sentence. The Panel issued each question to two frontier AI subjects with web search active.
The Panel then bound every AI response to verbatim regulator-issued source text held as primary substrate, comparing the model output against the deposited treaty text and the regulator-issued source documentation for each provision. Only responses where the AI subject was demonstrably wrong against the verbatim regulator-issued source text are published as findings; responses that were substantively correct, or that refused on calibration grounds, are retained internally and not surfaced.
The cell carries 2 confirmed findings against the AI subjects on the BBNJ Agreement. Each is published against verbatim regulator-issued treaty text and carries explicit model attribution for audit transparency.
Finding 2: Marine genetic resources retroactivity flipped from non-retroactive default to retroactive default. The Specialist Panel asked, in application form, whether the BBNJ Agreement applies its marine genetic resources benefit-sharing obligations to specimens and digital sequence information collected from international waters before the Agreement entered into force. Claude Opus 4.7 with web search active answered that under Article 10(1) the marine genetic resources and digital sequence information provisions apply to utilization of resources that were collected or generated before entry into force, not just after (RLB-H-INT-UNTC-BBNJ-HIGH-SEAS-BIODIVERSITY-AGREEMENT-2023-Q003-Opus47).
Claude Sonnet 4.6 with web search active reached the same substantive conclusion, stating that the benefit-sharing and notification obligations apply to the utilisation of resources collected or generated before the Agreement entered into force on 17 January 2026 (RLB-H-INT-UNTC-BBNJ-HIGH-SEAS-BIODIVERSITY-AGREEMENT-2023-Q003-Sonnet46). The substrate held by the Panel records the opposite position: Article 10(1) limits the Part II regime to marine genetic resources and digital sequence information collected and generated after entry into force for each Party, with non-retroactivity as the default.
This is the highest-stakes class of error in the cell because the substantive direction of the obligation is inverted, not merely the article number.
Finding 3: Digital sequence information benefit-sharing duty pinned to Article 15(5) instead of Article 14(1). The Specialist Panel asked, in direct form, whether the BBNJ Agreement's benefit-sharing framework extends to digital sequence information derived from marine organisms found in international waters, and which provision governs the obligation. Claude Sonnet 4.6 with web search active answered that Part II, including Article 15(5) and related provisions, applies to both physical marine genetic resource samples and their associated digital sequence information (RLB-H-INT-UNTC-BBNJ-HIGH-SEAS-BIODIVERSITY-AGREEMENT-2023-Q004-Sonnet46).
The substrate held by the Panel locates the duty at Article 14(1): benefits arising from activities related to marine genetic resources in areas beyond national jurisdiction and their digital sequence information shall be shared in a fair and equitable manner. The substantive conclusion that digital sequence information is in scope is correct; the article number is not.
For legal teams at biotechnology firms advising on the BBNJ Agreement, treaty-citation accuracy is load-bearing in legal opinions, contractual representations, due-diligence disclosures, and any pleading or position paper engaging the Agreement. A counterparty or opposing counsel who identifies a misattributed article on first reading calls the entire piece of advice into question. The marine genetic resources retroactivity inversion is the more serious failure: a legal opinion structured around a retroactive-by-default rule when the treaty establishes the opposite default produces fundamentally wrong contract terms and exposes the firm to professional liability if the underlying position is later corrected.
The verbatim regulator-issued source text held by the RLB Specialist Panel as primary substrate for the BBNJ Agreement sets the position as follows. The references below are drawn from the deposited treaty text and are the controlling reference points against which any AI-assisted citation should be validated.
Article 10(1). The marine genetic resources and digital sequence information provisions in Part II apply only to resources collected and generated after the Agreement enters into force for each Party. Non-retroactivity is the default. A number of Parties also recorded formal declarations on this point at deposit. Legacy collections obtained before the Party's date of entry into force are outside the Part II benefit-sharing regime by default; bringing them in requires an affirmative act, not the absence of one.
Article 14(1). Benefits arising from activities related to marine genetic resources in areas beyond national jurisdiction and their digital sequence information shall be shared in a fair and equitable manner. Digital sequence information is on the face of the provision, not parked in a subsidiary clause elsewhere in Part II. Article 15 carries the modalities of monetary and non-monetary benefit-sharing, but the source duty sits at Article 14(1).
There are two distinct lessons here for legal teams at biotechnology firms working with AI on the BBNJ Agreement. The first, on Article 10(1), is the more serious. The AI subjects' answers are not surface citation slips: they invert the direction of an obligation. Where the treaty says non-retroactivity is the default, the AI subjects said retroactivity is the default. A reader who runs an AI-drafted memo or checklist through a citation-check workflow will not catch this; the article number cited is actually right. Only a substantive read against the treaty text catches the inversion.
This is the failure mode AI Labs need to know about, and it is the failure mode practitioners need to defend against most rigorously. The second, on Articles 27, 14(1), and 22(2), is the more familiar pattern: the substantive paraphrase is close enough that the answer reads as authoritative, but the article number is wrong. Citation-checking workflows will catch this if they run against the deposited treaty text rather than against a chained reference.
The defensive posture is the same: always anchor the citation against the deposited Agreement text, not against a secondary commentary, before letting an AI-drafted output go out the door.
The RLB Specialist Panel is engaging with the AI subjects' developers and with practitioner audiences working under the BBNJ Agreement. The Panel maintains an audit register of confirmed hallucinations bound to verbatim regulator-issued source text, surfaces them on the live regulation page and on each audience-specific briefing, and accepts right-of-reply submissions from the AI subjects' developers and from regulator-side reviewers.
For legal teams at biotechnology firms this means the same questions can be re-issued against successor model releases; the bound substrate makes it straightforward to verify whether a specific failure mode has been corrected upstream, or whether the same hallucination is still being produced. Partnership briefings with AI labs are offered against the audit register, not against synthesised demonstrations, so the corrections that matter are evidenced against treaty text rather than against a paraphrase chain.
The register is structured so that each finding records the question put to the AI subject, the AI subject's verbatim answer, the verbatim regulator-issued source text the answer was bound against, the named model and configuration, and the failure mode. That structure lets practitioner readers see exactly where the AI subject diverged from the treaty text without re-doing the underlying verification, and lets AI lab readers see exactly which provision and which phrasing produced the divergence.
Where a hallucination has been corrected in a successor model release, the register records the rerun and withdraws the finding; where it persists, the finding stays live. This makes the register useful as a continuous-improvement signal for the AI labs and as a defensive checklist for practitioners drawing on AI in regulated workflows.
For legal teams at biotechnology firms drawing on AI in workflows that touch the BBNJ Agreement, the practical action items are direct:
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.
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: UN BBNJ Agreement (2023), Agreement on the Conservation and Sustainable Use of Marine Biological Diversity of Areas Beyond National Jurisdiction · Substrate documents: p_01_ACT_Part_III___Article_22_2____non_undermini_text-bbnj-agreement.html, p_01_ACT_Part_II___DSI_included_in_BBNJ_vs__exclu_text-bbnj-agreement.html · UN portal: documents.un.org
Citation IDs referenced:
RLB-H-INT-UNTC-BBNJ-HIGH-SEAS-BIODIVERSITY-AGREEMENT-2023-Q003-Opus47RLB-H-INT-UNTC-BBNJ-HIGH-SEAS-BIODIVERSITY-AGREEMENT-2023-Q004-Sonnet46