AI on BBNJ High Seas Biodiversity Agreement for Legal teams at Clinical Research firms in international jurisdictions

Executive Summary

The BBNJ High Seas Biodiversity Agreement — formally the Agreement under the United Nations Convention on the Law of the Sea on the Conservation and Sustainable Use of Marine Biological Diversity of Areas Beyond National Jurisdiction — establishes for the first time a binding international framework for access to marine genetic resources (MGR) from the high seas and the fair and equitable sharing of benefits derived from their use.

For Legal teams at Clinical Research firms operating internationally, the Agreement directly governs whether existing sample collections, ongoing research programmes, and commercial product pipelines involving high-seas-derived biological material carry new compliance obligations.

Across the one aggregated question we examined on this regulation, AI tools produced a verified hallucination. The error was not peripheral: AI assistants inverted the Agreement's core retroactivity rule, telling users that benefit-sharing obligations apply by default to MGR collected before the Agreement entered into force — the precise opposite of what Article 10(1) states. When challenged, the AI acknowledged uncertainty, confirming that the initial confident answer was not grounded in the text of the Agreement itself.

For a Legal team advising internal stakeholders on whether legacy marine sample libraries are covered by the new regime, this class of error converts a straightforward statutory question into a source of material corporate risk.

How AI gets this regulation wrong

The predominant failure pattern on this regulation is confident rule inversion: AI tools stated the operative legal default with conviction, but stated it backwards. Rather than surfacing genuine uncertainty about a novel treaty, the AI answered as if the rule were settled — and settled in the wrong direction — only retreating when directly pressed. The table below captures how this failure category maps to the specific questions examined.

What that means for your team

For Legal teams at Clinical Research firms, errors on the BBNJ Agreement translate most directly into wrong deliverables — legal opinions, compliance memos, and internal policy guidance that misstate the firm's actual obligations. The table below maps the risk categories arising from these AI failures to the specific operational contexts where the Legal team's advice is most consequential.

When this affects your department

Clinical Research firms with international operations frequently maintain libraries of biological material derived from marine environments, including samples collected during research expeditions, through academic partnerships, or via third-party biorepositories. As the BBNJ Agreement moves toward ratification and entry into force across signatory states, Legal teams are being asked to map which collections fall within its scope, whether legacy holdings create new reporting or benefit-sharing obligations, and how commercial development pipelines should be structured going forward.

These are not hypothetical scenarios: the retroactivity question — whether pre-entry-into-force collections are governed by the Agreement — is often the first and most consequential question a Legal team must resolve, because it determines the entire scope of the compliance exercise.

AI tools are a natural first port of call for this kind of treaty orientation work. A Legal team may use AI to draft a regulatory mapping memo, produce a training briefing for R&D or business development colleagues, or prepare due-diligence materials for a licensing negotiation or acquisition. If the AI's answer on a foundational question — such as the retroactivity default — is wrong, every downstream document that relies on it will embed the same error.

A compliance memo advising that pre-entry-into-force collections are subject to benefit-sharing obligations (when they are not) creates unnecessary contractual burdens, misrepresents the firm's legal position to counterparties, and may prompt costly remediation work once the error is identified.

The stakes are further elevated by the international dimension of both the regulation and the firm. In jurisdictions where the Agreement has entered into force, regulators and counterparties will increasingly expect legally accurate positions on scope and coverage. A Legal function that has relied on AI-generated analysis without verification may find itself exposed not only to internal credibility risk but also to reputational consequences in regulatory or commercial negotiations where the firm's stated legal interpretation later proves incorrect.

The findings at a glance

The table below summarises each finding on this regulation, including the question examined, the nature of the AI's error, and the risk category it creates for the Legal team at a Clinical Research firm in international jurisdictions.

Aggregate impact

The finding on this regulation reveals a specific and serious pattern: AI tools do not merely lack information about the BBNJ Agreement — they generate confident, well-structured answers that state the law incorrectly, without signalling that uncertainty is warranted. The Agreement is recent and technically complex, but the question examined was not obscure. The retroactivity provision in Article 10(1) is among the most commercially relevant clauses for any firm holding existing marine-derived material, and its plain-language interpretation is unambiguous in the treaty text.

Yet AI assistants inverted the operative default entirely, presenting a retroactive regime with an opt-out where the Agreement establishes a non-retroactive regime by design.

For Legal teams at Clinical Research firms, the concentrated nature of this error pattern — one core rule, inverted — carries disproportionate risk precisely because it is foundational. The retroactivity question is not one sub-issue among many; it is the threshold determination that shapes the scope of every subsequent compliance obligation. An incorrect answer here propagates through internal policies, supplier due-diligence checklists, acquisition assessments, and regulatory submissions. The systemic risk is not that the AI gets many things slightly wrong — it is that it gets the starting premise exactly wrong, with full apparent confidence.

The citation behaviour observed compounds the risk. AI tools supported their inverted positions by citing published commentary sources that were themselves either tangential to the specific legal question or characterised by our review as Pretextual — that is, real sources that do not actually support the stated proposition when read in context. This means a Legal professional reviewing the AI's answer without going to the treaty text directly may encounter apparently credible references that appear to confirm the incorrect position, making independent verification feel less necessary than it actually is.

What your team should do

The default position for Legal teams working on BBNJ Agreement questions should be treaty-text first. The Agreement is publicly available through the United Nations Treaty Collection portal, and the relevant provisions on marine genetic resources and digital sequence information are sufficiently well-structured that a legally trained reader can identify the operative rule without specialist treaty expertise. For the retroactivity question specifically, Article 10(1) should be read directly — AI tools we examined failed on this question precisely because they did not ground their answers in the plain text, and the error was not recoverable by follow-up questions alone.

As a practical safeguard, treat any AI-generated answer about BBNJ scope, retroactivity, or benefit-sharing obligations as a hypothesis to be verified rather than a starting position to be refined. Before any AI-assisted analysis is used in an internal memo, a training document, or a counterparty-facing deliverable, a legally qualified team member should confirm the underlying treaty provision. This is a proportionate check: the BBNJ Agreement is short enough that the MGR provisions can be reviewed in under an hour, and the cost of that verification is materially lower than the cost of correcting a wrongly-scoped compliance programme.

The fact that the AI cited external sources does not substitute for this check — verify the treaty text, not only the secondary commentary.

AI tools remain useful for the BBNJ Agreement in lower-stakes orientational work: drafting initial agendas for a regulatory review session, producing a first-pass glossary of treaty terminology for non-Legal colleagues, or generating a list of questions for specialist external counsel. Where AI adds value is in accelerating the framing of a legal question, not in resolving it. For questions that bear directly on the firm's compliance position — particularly any question touching on which collections fall inside or outside the Agreement's scope — human legal review of the primary text is mandatory before the answer is acted on.

How RLB Can Help

RegLeg's published Hallucination Research gives the Legal team at a Clinical Research firm a concrete pre-flight check before relying on AI-assisted output for regulatory questions. The research catalogues the specific ways AI tools misstate regulatory text, fabricate citations, and conflate overlapping frameworks — precisely the failure modes that carry the most risk when advising on clinical trial authorisations, informed-consent obligations, or post-market safety reporting. Reviewing the relevant regulator profiles before deploying AI tools on a matter takes minutes and surfaces the known failure patterns your team should verify by hand.

Beyond the published research, RLB works with Legal functions directly to map which AI-supported workflows in a Clinical Research context carry the highest hallucination exposure. That means examining the actual questions your team puts to AI tools — document summarisation, regulatory gap analysis, cross-jurisdiction compliance checks, submission drafting — and identifying where the combination of regulatory complexity, frequent amendment cycles, and AI training limitations creates the greatest risk of plausible-but-wrong output reaching a regulator or ethics committee. The output is a prioritised exposure map specific to your firm's workflows and the regulators you are subject to.

RLB also offers a confidential review of the firm's existing AI-use policy against RegLeg's failure-mode catalogue, with prioritised remediation recommendations framed around what the Legal team can implement without waiting for firm-wide IT or compliance cycles. Alongside this, RLB produces training material and CPD-aligned content the team can use internally — giving fee-earners and in-house counsel a structured way to develop and evidence competence in the responsible use of AI tools on regulated matters, without relying on generic vendor training that does not account for the specific regulatory environment Clinical Research firms operate in.