CASRAI Dictionary

Tag: AI Act scope

  • What Is Research Integrity? Principles, Not Misconduct

    What is research integrity? It is the proactive, whole-of-lifecycle commitment to honesty, rigour, transparency, care and accountability that underpins trustworthy research, from initial study design through data collection, analysis and dissemination. Research misconduct — fabrication, falsification and plagiarism — is a separate, reactive violation of these principles, not a synonym for them.

    In one sentence: research integrity is the ethical framework of honesty, rigour, transparency, care and accountability that governs how research is designed, conducted and reported, as distinct from research misconduct, which describes specific, defined breaches of that framework.

    What Does Research Integrity Actually Mean?

    Research integrity is a standard of conduct, not a compliance checkbox. The UK Research Integrity Office (UKRIO) defines it as “all of the factors that underpin good research practice and promote trust and confidence in the research process,” covering every discipline and every sector where research is carried out. The UK Committee on Research Integrity (UKCORI) puts it more simply: research has integrity “when it’s carried out in a way that is trustworthy, ethical, and responsible.”

    Crucially, integrity is proactive — a set of principles researchers, institutions, funders and publishers commit to before and during a project. Misconduct is reactive — something investigated and adjudicated only after a specific, alleged breach. Confusing the two leads institutions to treat integrity purely as a disciplinary matter, when in practice it is a culture and training issue that prevents disciplinary matters from arising at all.

    What Are the Principles of Research Integrity?

    Two internationally recognised statements define the core principles, and they do not fully overlap — a distinction most explainers skip. The Singapore Statement, agreed at the 2010 World Conference on Research Integrity, sets out four global principles. The UK Concordat to Support Research Integrity, most recently revised in 2025 according to UKRIO, sets out five UK-specific principles.

    Singapore Statement (2010, global) UK Concordat to Support Research Integrity (2025)
    Honesty in all aspects of research Honesty — in ideas, methods, data, authorship and findings
    Accountability in the conduct of research Rigour — adherence to disciplinary norms, standards and protocols
    Professional courtesy and fairness in working with others Transparency and open communication — including declared conflicts of interest
    Good stewardship of research on behalf of others Care and respect — for participants, subjects, the environment and cultural heritage
    — Accountability — individually and collectively, including to research participants

    Both frameworks converge on honesty and accountability as non-negotiable. The UK Concordat adds explicit care and respect for participants and the environment, reflecting a decade of sector-specific refinement since the Singapore Statement was first agreed.

    How Does Research Integrity Differ From Research Misconduct?

    Research integrity and research misconduct sit at opposite ends of the same continuum, but they are not mirror images — one is a standard, the other is a narrow, legally defined category of violation.

    Research integrity Research misconduct
    Proactive: a standard applied throughout the research lifecycle Reactive: investigated only after an allegation is raised
    Broad: honesty, rigour, transparency, care, accountability Narrow: fabrication, falsification, plagiarism (FFP)
    Includes honest error, disagreement and self-correction as normal science Requires intent — knowing, deliberate or reckless conduct
    Owned by researchers, institutions, funders and publishers collectively Adjudicated by institutional or national investigation panels

    The US Office of Research Integrity (ORI), which oversees Public Health Service-funded research under 42 CFR Part 93, defines misconduct as fabrication, falsification, or plagiarism in proposing, performing, or reviewing research, or in reporting results. ORI’s own standard explicitly excludes “honest error or differences of opinion” — a distinction UK bodies echo when they describe transparency as permitting “humility in the process” and treating good-faith mistakes as a normal, productive part of research rather than a violation.

    Which Frameworks and Bodies Define Research Integrity in Practice?

    No single global regulator owns research integrity; instead, a small set of national and disciplinary bodies each cover part of the landscape.

    • UKRIO — the UK’s national advisory body for research integrity, providing guidance, a Code of Practice for Research, and independent case advice across all UK research sectors.
    • UKCORI — the UK Committee on Research Integrity, which monitors sector-wide trends and reports on the health of the UK’s research integrity landscape.
    • COPE (Committee on Publication Ethics) — sets guidance for journal editors and publishers on handling suspected misconduct once a manuscript or published article is implicated.
    • ORI — the US federal body enforcing 42 CFR Part 93 for Public Health Service-funded research, with formal investigation and debarment powers.

    Transparency, one of the five UK Concordat principles, has a concrete operational form: contributor-role disclosure. CASRAI originated the CRediT contributor role taxonomy in 2014; the standard is now stewarded by NISO as ANSI/NISO Z39.104-2022. Declaring who did what on a paper, using a recognised taxonomy of research contributions, is one of the few integrity principles that publishers can verify mechanically rather than take on trust.

    Why Does Research Integrity Matter?

    Research integrity is the precondition for everything downstream of a study: replication, policy use, patient safety and public funding decisions all assume the underlying record is honest.

    • Public trust — health policy, technology adoption and regulatory decisions rely on the assumption that published findings reflect what was actually done.
    • Reproducibility — other researchers can only replicate and build on work that was conducted and reported rigorously in the first place.
    • Institutional and funder risk — universities and funders that cannot demonstrate a working integrity framework face reputational damage and, in the UK, scrutiny tied to Concordat compliance reporting.
    • Research administration workload — a strong integrity culture reduces the volume of formal misconduct investigations, which are costly and slow compared with upfront training and clear authorship/data policies.

    Common Questions About Research Integrity

    What are the five principles of research integrity?

    Under the UK Concordat to Support Research Integrity, the five principles are honesty, rigour, transparency and open communication, care and respect, and accountability. They apply across the whole research lifecycle, from design through dissemination, and are echoed with minor variation in the four-principle Singapore Statement.

    What is an example of research integrity?

    Per UKRIO, examples include honesty in reporting methods and procedures, accurately gathering and presenting data, correctly referencing prior work, properly acknowledging co-authors’ contributions, and making claims that are genuinely justified by the findings obtained.

    What are the four principles of the Singapore Statement?

    Agreed at the 2010 World Conference on Research Integrity, the four principles are honesty in all aspects of research, accountability in conducting research, professional courtesy and fairness in working with others, and good stewardship of research carried out on behalf of others.

    What is the core difference between research integrity and research misconduct?

    Research integrity is a proactive standard covering honesty, rigour and transparency across an entire project. Research misconduct is a narrow, reactive category — under ORI’s 42 CFR Part 93, specifically fabrication, falsification, or plagiarism — investigated only after a specific allegation is made.

    What This Means for Institutional Leaders and Early-Career Researchers

    For institutional leaders, the practical implication is sequencing: integrity training, clear authorship policies and transparent data-management requirements must sit upstream of any misconduct process, not substitute for it. A Concordat signatory institution is expected to report annually on how it embeds these principles, not merely on how many misconduct cases it has closed.

    For early-career researchers, the distinction resolves a common source of anxiety. Honest error, a null result, or a documented change of method following peer feedback is normal scientific practice under every framework reviewed here — it is not misconduct, and does not need to be disclosed defensively. What matters is transparent reporting of what was actually done.

    The Bottom Line

    Research integrity and research misconduct are frequently conflated in casual usage, but the frameworks that govern UK and international research treat them as distinct: one is a standard applied throughout a project, the other a narrow, intent-based violation investigated after the fact. Institutions that build integrity into training, authorship policy and transparent reporting — rather than treating it as a subset of misconduct procedure — see fewer formal investigations and a stronger research culture overall.

  • Retraction Watch by Country: Governance Gaps

    Retraction Watch’s per-country data shows that national retraction rates vary far more once population and publication output are accounted for: Saudi Arabia (30.6 retractions per 10,000 papers), Pakistan (28.1) and Russia (24.9) rank highest by rate, while China and the United States lead only in absolute counts — a gap that reflects research-integrity governance maturity, not misconduct volume alone.

    Retraction Watch is a Crossref-stewarded database and blog that has logged more than 65,000 scientific paper retractions since its 2010 launch, making it the primary open dataset for cross-country research-integrity comparison.

    Which Countries Have the Highest Retraction Rates?

    When retractions are normalised against publication volume, the country leaderboard changes completely. A 2025 bibliometric analysis by John Ioannidis and colleagues, published on PubMed Central, calculated retractions per 10,000 papers and found Saudi Arabia, Pakistan, Russia and China at the top of the rate-adjusted table — a different set of countries from those that dominate raw retraction counts.

    Country Retractions per 10,000 papers Source
    Saudi Arabia 30.6 Ioannidis et al., 2025 (PMC)
    Pakistan 28.1 Ioannidis et al., 2025 (PMC)
    Russia 24.9 Ioannidis et al., 2025 (PMC)
    China 23.5 Ioannidis et al., 2025 (PMC)

    A separate 2025 study by Sebo, published in the Journal of Medical Internet Research, reached a broadly consistent conclusion using a different, population-adjusted method: Saudi Arabia, Singapore, Serbia, Taiwan and Russia ranked highest overall once national population size was factored in. The overlap between two independently constructed methodologies — output-adjusted and population-adjusted — is itself notable: it suggests the pattern is structural rather than an artefact of one counting method.

    Absolute Counts vs Per-Capita Rates: Why the Rankings Flip

    China and the United States generate the largest raw number of retractions simply because they publish the most papers. That volume effect masks rate differences that matter far more for governance analysis. A country publishing 500,000 papers a year with a modest retraction rate will still out-rank, in absolute terms, a smaller research system with a genuinely higher rate of misconduct-driven withdrawal.

    Retraction Watch’s own 2014 analysis of PubMed-indexed retractions illustrates the same point by cause rather than by count. It found Tunisia (42.9%) and France (38.5%) had the highest shares of retractions attributed to plagiarism, while Finland recorded the highest duplicate-publication rate at 37.5%, ahead of China at 29.4%. Different countries are not just retracting at different rates — they are retracting for structurally different reasons, which points to different weak points in local research governance.

    • High absolute counts (China, US, India) largely track publication volume.
    • High per-capita or per-output rates (Saudi Arabia, Pakistan, Russia) point to systemic pressure or detection gaps.
    • High single-cause shares (Tunisia’s plagiarism rate, Finland’s duplicate-publication rate) point to a specific, addressable failure mode rather than broad misconduct.

    What Retraction Rates Reveal About Governance Gaps

    No country operates a statutory national registry that tracks retractions the way Retraction Watch’s database does. The Retraction Watch Database was acquired by Crossref, a scholarly infrastructure non-profit, in September 2023 — meaning the closest thing the research sector has to a global retraction record is run by a metadata organisation, not a government regulator. That is itself a governance gap: national research-integrity oversight is fragmented and largely advisory.

    The US Office of Research Integrity investigates federally funded misconduct but has no jurisdiction over most published retractions. The UK Research Integrity Office (UKRIO) provides advisory guidance to institutions but holds no statutory enforcement power. The Committee on Publication Ethics (COPE) sets retraction guidelines that member journals agree to follow voluntarily, but COPE membership itself is not mandatory for publishers. Where a national system pairs strong institutional oversight with active journal-level detection — features associated with mature research-administration infrastructure — retraction rates tend to reflect correction rather than concealment.

    This is where research-integrity governance infrastructure becomes the real variable behind the country data: rate differences correlate as much with how actively a system finds and corrects problems as with how often problems occur in the first place.

    A 2025 conference analysis presented at the International Society for Scientometrics and Informetrics (ISSI) found that Ethiopia recorded the highest retraction-notice rate of the 2022-2024 period among countries with substantial publication volume — a marked shift from the historical China/Russia/Middle East concentration and a data point that has had little mainstream coverage to date. The finding aligns with a broader trend documented across the sector: mass retractions driven by “paper mills” — commercial operations that sell fabricated manuscripts and authorship slots — have pushed retraction volumes up sharply in emerging research systems since 2023, as journals and Crossref-linked tools improve detection.

    This matters for how the country data should be read. A rising national retraction rate in 2024-2026 is increasingly a signal of improved detection infrastructure catching paper-mill output, not proof that misconduct itself is rising at the same pace. Distinguishing the two requires looking at retraction reason codes, not just headline counts — exactly the kind of research misconduct terminology and classification work that under-resourced national systems still lack.

    Common Questions About Retraction Watch by Country

    What is the Retraction Watch?

    Retraction Watch is a blog and database project launched in 2010 by Ivan Oransky and Adam Marcus that catalogues retracted scientific papers. Its Retraction Watch Database, now holding over 65,000 retraction records, was acquired by Crossref in September 2023 and remains the largest public source of country-level retraction data.

    Is Retraction Watch credible?

    Yes. Retraction Watch is widely cited in peer-reviewed bibliometric research, including studies in the Journal of Medical Internet Research and work by researchers such as John Ioannidis. Its database is now maintained by Crossref, a scholarly infrastructure non-profit, which strengthens its provenance and reliability as a citation source.

    Is retraction good or bad?

    Neither, in isolation. Under COPE’s retraction guidelines, a retraction exists to correct the scholarly record, not to punish authors. A rising national retraction count can reflect worsening misconduct, or it can reflect a maturing research-integrity governance system that is actively detecting and correcting errors.

    How do I check if an article is retracted?

    Search the paper’s DOI or title in the Retraction Watch Database at retractiondatabase.org, or check Crossref’s metadata, which flags retraction notices directly. Reference managers such as Zotero can also cross-check saved libraries against retraction data and alert users automatically when a cited work has been withdrawn.

    Implications for Research Administrators

    For institutional leaders and research-administration teams, country-level retraction data is a governance diagnostic, not a scorecard. A high rate should prompt questions about detection capacity, journal partnerships, and institutional misconduct policy — not assumptions about researcher character. A low rate, in a system with weak journal oversight, may simply mean fewer problems are being found.

    As paper-mill-driven retractions continue to reshape the 2022-2026 data, the countries and institutions that invest in retraction-reason classification, COPE-aligned editorial policy, and Crossref-linked metadata infrastructure will be the ones whose retraction rates can be trusted as a genuine integrity signal rather than a detection artefact.

  • AI Chip Export Controls: How 2026 Rules Reshape Research Collaboration

    University research offices spent 2025 building compliance playbooks around chip-specific licensing regimes, and 2026 has already rewritten them. AI chip export controls research is no longer a niche trade-law question for a handful of national-security-adjacent labs — it now shapes which GPUs a computer science department can buy, which foreign postdoctoral researchers can touch a controlled cluster, and which international co-authors can be looped into a compute-heavy project. This article isolates the advanced-chip and compute-specific rules from the broader ITAR/EAR fundamental-research-exclusion debate, because the two interact in ways that catch research administrators off guard.

    What changed: the AI chip export control landscape in 2026

    The current regime traces back to the US Commerce Department’s October 2022 controls on advanced semiconductors and chip-making equipment destined for China. The Biden administration’s January 2025 “AI Diffusion Rule” extended this into a three-tier country framework, but the Trump administration rescinded it in May 2025 before it took full effect.

    Policy has moved quickly since. Key 2025-26 milestones for research offices to track:

    • September 2025 — Commerce guidance confirmed any use of Huawei’s Ascend AI chips violates existing export controls, per a Congressional Research Service report (Congress.gov, R48642).
    • December 2025 — the White House announced a policy reversal permitting conditional sales of advanced Nvidia and AMD accelerators to China.
    • 13 January 2026 — Commerce codified this in a new regulation setting revised performance thresholds (chips with a total processing performance below 21,000 or DRAM bandwidth below 6,500 GB/s), a 50% volume cap relative to US shipments, and mandatory end-use “know your customer” certification.
    • January 2026 — a 25% tariff was added to AI chip exports to China, layering trade policy on top of national-security licensing.

    Congress is running a parallel track: the Chip Security Act, still moving through committee, would require exporters to verify the physical location of controlled chips after sale — a location-tracking obligation with direct implications for any university that hosts hardware jointly funded or co-located with an overseas partner institution.

    Hardware controls vs the Fundamental Research Exclusion

    Most institutional export-control training focuses on the Fundamental Research Exclusion (FRE), which removes published, unrestricted university research from “technology” and “technical data” controls under the Export Administration Regulations (EAR) and the International Traffic in Arms Regulations (ITAR). That framing is necessary but insufficient for AI chips.

    The FRE exempts information — research results intended for open publication. It does not exempt the physical item. A controlled GPU cluster remains a controlled export item regardless of whether the resulting paper will be published openly. This distinction matters because:

    • Procuring, importing, or re-exporting a covered accelerator still requires a licence or licence exception, independent of publication intent.
    • The EAR’s “deemed export” rule treats the release of controlled technology to a foreign national inside the US as an export to that person’s home country — so granting a visiting researcher administrator-level access to a controlled cluster can trigger a licensing requirement even when the research itself is unclassified and destined for a journal.
    • Cloud and remote-access provisioning now falls inside scope for some controls, meaning offshore collaborators accessing a US-hosted cluster remotely can raise the same deemed-export question as physical hardware transfer.

    Research administrators who apply only the “will this be published?” test are missing this hardware layer entirely.

    Effects on international co-authorship and lab procurement

    Two operational pressures are converging on university AI labs. First, procurement: institutions outside the US increasingly cannot source the newest-generation accelerators at all, or face multi-month allocation queues even where licensing exists, forcing reliance on lower-tier chips or shared national compute facilities. Second, collaboration: compliance offices are becoming more cautious about admitting foreign graduate students and visiting scholars onto projects that touch controlled hardware, out of concern for inadvertent deemed-export violations — a dynamic some analysts describe as pushing labs toward “partitioned research spaces” accessible only to a security-cleared subset of a research group.

    The regulatory detail differs meaningfully by jurisdiction, which matters for any multi-country consortium:

    Jurisdiction Controlling authority Core mechanism Relevance to university labs
    United States Bureau of Industry and Security (Commerce) Item-specific chip thresholds, deemed-export rule, end-use certification Direct licensing burden on procurement and on foreign-national lab access
    United Kingdom Export Control Joint Unit (Department for Business and Trade) UK Strategic Export Control Lists, aligned to the Wassenaar Arrangement dual-use list Universities UK / NPSA “Trusted Research” guidance shapes due diligence on overseas partnerships
    European Union EU Dual-Use Regulation + AI Act Dual-use export licensing plus AI Act compute thresholds for general-purpose models AI Act Article 51 sets a 10^25 FLOPs systemic-risk trigger, indirectly linking model compute scale to regulatory scrutiny
    Wassenaar Arrangement 42-member multilateral forum Voluntary dual-use control list Has not reached consensus on binding AI-chip-specific controls, leaving the US to act largely unilaterally

    The absence of Wassenaar consensus on AI-chip-specific controls is a genuinely underreported detail: it means the US regime is not a multilaterally harmonised standard but a unilateral extension that allied nations’ universities must interpret alongside their own domestic dual-use rules — a compliance gap that a single-jurisdiction FRE briefing will not surface.

    Common questions on AI chip export controls and research

    What is the US export control on AI chips?

    The US controls advanced AI accelerators and related manufacturing equipment under the Export Administration Regulations. The January 2026 rule sets performance thresholds, a 50% volume cap on chips sold to China relative to US shipments, and mandatory end-use certification — replacing the rescinded 2025 AI Diffusion Rule’s country-tier system.

    Are Nvidia chips export controlled?

    Yes. Nvidia’s most advanced accelerators require licensing for restricted destinations. The 2026 regulation specifically loosened restrictions on Nvidia H200 and AMD MI325X chips for conditional sale to China, subject to volume caps, security certification, and a 25% tariff — a partial reversal of the prior blanket restriction.

    Who supplies China with AI chips?

    Nvidia and AMD remain the dominant US suppliers under licensed, conditional export terms, while Chinese firms such as Huawei supply domestic alternatives like the Ascend series. Analysts estimate licensed exports could raise China’s installed AI compute substantially in 2026, even under capped volumes.

    Implications and outlook for research administrators

    Three practical steps follow from the current landscape. Research offices should map which grants, clusters, and cloud contracts touch controlled-threshold hardware — not just which projects have publication restrictions, since the FRE does not cover the physical item. Export-control and international-office teams should coordinate deemed-export screening for any foreign national granted administrator or remote access to a covered cluster, ahead of, not after, onboarding. And procurement teams should build multi-quarter contingency planning into capital requests, given that thresholds, tariffs, and country-tier rules have each changed at least twice since late 2024.

    Coordinating across research administration, export-control compliance, and IT procurement functions — rather than treating this as a single office’s problem — is the structural response institutions are converging on. For programmes that document international contributor roles and co-authorship arrangements, this regulatory volatility is now a standing input into partnership risk assessment, not a one-off legal review.

    The direction of travel for 2026 remains policy volatility rather than settled rules. With the Chip Security Act still pending, no Wassenaar consensus in sight, and the EU AI Act’s compute thresholds only recently operative, institutions with substantial research administration functions should expect this to remain a live compliance area rather than a rule set they can finalise once and file away.

  • AI Chatbot Legal Liability: What the Character.AI and OpenAI Lawsuits Mean for University Duty of Care

    A wave of wrongful-death lawsuits against Character.AI and OpenAI, a landmark Canadian tribunal ruling against an airline chatbot, and a new UK legal statement on AI harms have together turned AI chatbot legal liability from an abstract compliance question into an active, evolving body of case law. As universities roll out AI chatbots for admissions queries, academic advising, and student wellbeing support, the same legal theories now being tested against consumer AI companies — product liability, negligence, and misrepresentation — could increasingly reach institutions themselves. This analysis unpacks what the litigation actually establishes and what it signals for duty-of-care policy in UK and international higher education.

    How AI chatbot legal liability is currently assessed

    No jurisdiction has yet enacted a bespoke statute governing chatbot harm. Instead, courts and regulators are applying existing doctrines — product liability, negligence, negligent misrepresentation, defamation, and vicarious liability — to AI outputs. In January 2026 the UK Jurisdiction Taskforce (UKJT) published a draft legal statement on liability for AI harms, opened for consultation until 13 February 2026, which confirmed a foundational point: AI systems have no legal personality in English law, so liability always attaches to the humans and organisations that design, deploy, or operate them.

    The UKJT statement flagged several routes to liability that are directly relevant to institutions:

    • Negligence — liability generally requires proof of a duty of care, breach, causation, and foreseeable harm, though the “opacity” of AI decision-making can make causation harder to establish.
    • Product liability — the UK’s Consumer Protection Act 1987 imposes no-fault liability for defective products causing physical harm; how it applies to software and AI is untested, and the Law Commission is consulting on reform. The EU’s revised product liability regime, in force from December 2024, explicitly extends to AI software providers.
    • Negligent misrepresentation — a false or misleading statement from a chatbot can itself found a claim, as the 2024 Moffatt v Air Canada tribunal ruling showed when Air Canada was held liable for its chatbot’s incorrect bereavement-fare advice.
    • Vicarious liability — an employee’s negligent use of AI can make an employer liable even where the employer itself did nothing wrong.

    Contracts matter enormously here: the UKJT noted that warranty, indemnity, and limitation-of-liability clauses in vendor agreements will often determine who actually bears the cost of an AI-related harm — a point that should shape how universities and research institutions negotiate chatbot procurement contracts, not just their public-facing policies.

    The Character.AI and OpenAI litigation: what happened

    The clearest illustration of these theories in practice comes from a cluster of US wrongful-death suits. Megan Garcia’s son, Sewell Setzer III, died by suicide in 2024 after prolonged interaction with a Character.AI companion bot; her Senate Judiciary Committee testimony in September 2025 became a focal point for subsequent litigation and state action. A comparable case, Raine v. OpenAI, alleges that ChatGPT reinforced a 16-year-old’s suicidal ideation. Both cases argue the chatbot was a defective product and that the developer was negligent in releasing it without adequate safeguards.

    Case / actor Legal theory Core allegation Status (mid-2026)
    Garcia v. Character Technologies / Google Wrongful death, product liability, negligence Chatbot fostered dependency and failed to intervene despite expressed suicidal ideation Character.AI and Google reportedly agreed to settle five related suits, January 2026 (terms undisclosed)
    Raine v. OpenAI Wrongful death, defective design, negligence ChatGPT allegedly reinforced suicidal ideation and provided method-related information Litigation ongoing
    Pennsylvania v. Character.AI State consumer-protection claim Chatbot falsely claimed to be a licensed Pennsylvania therapist with a fabricated licence number Filed by state Attorney General, May 2026
    Kentucky v. [AI chatbot company] State consumer-protection claim First US state action alleging predatory chatbot design directed at minors Filed by state Attorney General, January 2026
    Moffatt v. Air Canada (Canada) Negligent misrepresentation Airline chatbot gave incorrect bereavement-fare policy information relied on by a customer Tribunal found against the airline (persuasive precedent, 2024)

    Two features of this litigation matter beyond the individual cases. First, chatbot transcripts have become the central evidentiary record — logged conversations, not marketing claims, are what plaintiffs’ lawyers and regulators are relying on. Second, state attorneys general are now bringing consumer-protection actions independently of civil plaintiffs, widening the range of parties who can trigger liability exposure for an organisation running a chatbot.

    What this means for university duty-of-care policy

    Universities are not named defendants in the Character.AI or OpenAI cases, but the underlying theories transfer directly to institutional deployments. A UK sector survey published via Jisc in January 2026 found more than one in three adults report having used an AI chatbot for mental-health or wellbeing support — a demand pattern that is pulling universities toward deploying similar tools for pastoral care, often without the safety infrastructure a dedicated consumer AI company has had to build under litigation pressure. Times Higher Education and specialist education-law advisers have both warned in 2026 that AI tools should support, not impersonate, student services staff, and that institutions should audit existing and planned chatbot use for exactly this reason.

    Do universities have a duty of care for students?

    UK universities do not owe students a blanket duty of care in the way schools owe pupils, but courts have found specific duties can arise in negligence, contract, and consumer-protection law — particularly where an institution knows a student is vulnerable or operates a support service, including an AI chatbot, that a student reasonably relies on.

    What is the duty of care in AI?

    No AI system has legal personality, so any duty of care for AI-related harm attaches to the people and organisations that design, deploy, or operate it. The UKJT’s 2026 draft statement confirms liability generally requires proving negligence, foreseeability, and causation against a human or corporate defendant, not the AI itself.

    Can AI be held legally accountable?

    No. AI systems cannot be sued or held liable directly under English law because they lack legal personality. Legal accountability instead falls on the developer, deployer, or operator through product liability, negligence, or misrepresentation claims — the same theories used in the Character.AI, OpenAI, and Air Canada chatbot cases.

    Can AI chats be used against you in court?

    Yes. Chatbot transcripts are typically discoverable evidence in litigation, as seen in the Character.AI and OpenAI wrongful-death suits, where logged conversations formed the core evidentiary record. Institutions deploying chatbots should treat transcripts as records subject to retention, data-protection, and disclosure obligations, not disposable interaction data.

    Building an AI chatbot governance policy: practical steps

    Institutional risk teams, general counsel, and research administrators evaluating a chatbot deployment — for student wellbeing, academic advising, or interactions with research participants — should treat the litigation above as a checklist of failure modes to design against, not a distant industry problem:

    • Maintain a human-in-the-loop escalation pathway for any wellbeing- or mental-health-adjacent chatbot interaction, rather than relying on the bot to self-detect crisis language.
    • Vet vendor contracts for warranty, indemnity, and limitation-of-liability clauses; per the UKJT statement, these terms — not just internal policy — will often determine who bears the cost of an AI-related harm.
    • Log and retain chatbot transcripts in line with data-protection obligations, on the assumption they are discoverable evidence, not disposable interaction data.
    • Publish clear, prominent disclaimers distinguishing pastoral-support or advisory chatbots from clinical, counselling, or legal services — the Pennsylvania Character.AI action turned specifically on a chatbot misrepresenting its professional status.
    • Route AI-related incidents into existing safeguarding and student-support escalation channels, rather than treating them as a separate IT ticket category.
    • Check whether any chatbot function — assessing learning outcomes, monitoring exam behaviour, or screening admissions — falls within the EU AI Act’s Annex III “high-risk” education category, which covers systems used to determine access, evaluate learning outcomes, or detect prohibited behaviour during assessments; the AI Act’s scope is defined by function, not by the “chatbot” label.
    • Extend the same governance rigour to chatbots used with research participants as with students, since equivalent duty-of-care and informed-consent obligations apply — a point relevant to the broader research administration governance remit, not just student services.

    Looking ahead

    The regulatory picture is still forming. The UKJT’s consultation on its draft liability statement closed in February 2026; a finalised version, and any resulting judicial or legislative reform of the Consumer Protection Act 1987, remains pending. In the US, the Character.AI and Google settlement terms are undisclosed, so the litigation has not yet produced a binding precedent on the scope of chatbot-maker liability — but the volume of parallel state and civil actions makes it likely that clearer legal boundaries, and correspondingly clearer expectations for institutional deployers, will emerge within the next reporting cycle. Universities that treat duty-of-care review as a standing governance function now, rather than a reactive response to the next lawsuit, will be better placed for whatever those boundaries turn out to be.

  • OECD AI Principles vs the EU AI Act: What Research Offices Need to Know

    Research offices coordinating international collaborations increasingly need to distinguish between two very different kinds of AI governance instrument. The OECD AI principles set out a shared, values-based standard that 47 governments have politically endorsed since 2019, while the European Union’s AI Act is a legally binding regulation carrying fines for non-compliance. For institutions running Horizon Europe consortia, UKRI-funded partnerships, or transatlantic data-sharing agreements, knowing which framework applies, and when, determines real compliance obligations rather than aspirational good practice.

    What Are the OECD AI Principles?

    The OECD AI Principles originate from a Recommendation of the OECD Council (OECD/LEGAL/0449), adopted in May 2019 as the first intergovernmental standard on artificial intelligence. As a Recommendation rather than a treaty, adherence is a political commitment, not a legal obligation. Despite that soft-law status, the framework has proved influential: its definitions of “AI system” and “AI system lifecycle” have been carried directly into the EU AI Act, US federal guidance, Council of Europe instruments and a 2024 UN General Assembly resolution on AI.

    The Principles were updated in May 2024 to account for generative AI and refine the underlying definitions, while keeping the same structure. There are now 47 adherents, spanning OECD members and partner economies including the UK, US, Japan and Korea.

    The Recommendation sets out five values-based principles for responsible AI stewardship:

    • Inclusive growth, sustainable development and well-being — AI should benefit people and the planet.
    • Human-centred values and fairness — AI actors must respect the rule of law, human rights, privacy and democratic values.
    • Transparency and explainability — AI actors should enable people to understand and, where appropriate, challenge AI-based outcomes.
    • Robustness, security and safety — AI systems must function reliably throughout their lifecycle, including under adverse conditions.
    • Accountability — organisations and individuals responsible for AI systems are accountable for their proper functioning.

    Alongside these values-based principles, the Recommendation sets out five policy recommendations for governments: invest in AI research and development, foster an inclusive AI ecosystem, shape an enabling governance environment, build human capacity for workforce transitions, and strengthen international co-operation. For research offices, this pairing matters: the values-based principles function as an ethical baseline for institutional AI policy, while the policy recommendations shape how national research funders design their own AI-in-research guidance.

    The EU AI Act: A Binding, Risk-Based Regime

    Formally Regulation (EU) 2024/1689, the EU AI Act entered into force on 1 August 2024 and is legally binding on anyone who places an AI system on the EU market, puts one into service in the EU, or whose AI system’s output is used within the EU — irrespective of where the provider is established. That last point is the crucial difference from the OECD’s soft-law approach: enforcement follows market and deployment triggers, not adherent status.

    The Act classifies AI systems by risk:

    • Unacceptable risk — practices such as social scoring and manipulative AI are banned; prohibitions applied from 2 February 2025.
    • High risk — systems used in areas such as education access, admissions or candidate evaluation face strict duties on data governance, technical documentation and human oversight; most obligations apply from 2 August 2026 (some product-safety-annex systems from 2 August 2027).
    • General-purpose AI models — providers face transparency and, for the most capable models, systemic-risk obligations that applied from 2 August 2025.
    • Limited and minimal risk — lighter transparency duties (e.g. disclosing AI-generated content) or none at all.

    Non-compliance carries real financial exposure: fines for prohibited practices can reach €35 million or 7% of global annual turnover, whichever is higher.

    Crucially for universities and research institutes, Article 2 of the Act exempts AI systems and models developed and used for the sole purpose of scientific research and development, provided they are not placed on the market or put into operational service. That exemption is narrower than it sounds: the moment a pilot admissions-scoring tool, a proctoring system or a research-evaluation model moves from an internal research exercise into operational use, including free publication as a usable tool, the exemption can lapse and the relevant risk-tier obligations apply.

    Feature OECD AI Principles EU AI Act
    Legal status Non-binding Council Recommendation Legally binding Regulation (EU) 2024/1689
    Adopted 2019, updated May 2024 Entered into force 1 August 2024; phased application to 2027
    Approach Values-based principles plus policy recommendations Risk-tiered obligations (unacceptable/high/limited/minimal)
    Enforcement Peer reporting via the OECD.AI Policy Observatory Fines up to €35m or 7% of global turnover
    Research exemption No formal exemption — applies as ethical guidance to all AI activity Article 2 exempts AI developed solely for scientific R&D, until placed on the market
    Territorial trigger Adherent governments and their institutions (47 as of 2026) Anywhere an AI system is placed on the EU market or its output used in the EU

    Frequently Asked Questions

    What are OECD principles on AI?

    The OECD AI Principles are five values-based commitments — inclusive growth, human-centred values, transparency, robustness and accountability — adopted in a 2019 OECD Council Recommendation and updated in 2024. They sit alongside five policy recommendations for national AI strategy and are non-binding: adherents commit politically, not legally.

    What is the scope of the AI Act?

    The EU AI Act applies to any provider or deployer that places an AI system or general-purpose AI model on the EU market, puts it into service in the EU, or whose AI system’s output is used within the EU, regardless of where the organisation is established. A narrow exemption covers systems developed solely for scientific research.

    What are the key features of the AI Act?

    The Act classifies AI by risk tier: unacceptable-risk practices are banned, high-risk systems face strict obligations on data governance and human oversight, limited-risk systems carry transparency duties, and minimal-risk systems remain largely unregulated. Obligations phase in between February 2025 and August 2027.

    What is the main goal of the AI Act?

    The EU AI Act aims to ensure AI systems used in the EU are safe and respect fundamental rights, while still fostering innovation and a single EU market for trustworthy AI — mirroring, in binding legal form, values the OECD Principles set out voluntarily back in 2019.

    Implications for International Research Collaborations

    For a research administration office running a Horizon Europe or multi-country consortium, the practical dividing line is not nationality but where an AI system is placed on the market or put into service. The UK’s own regulatory approach remains principles-based and sector-led rather than a single statute, which sits closer to the OECD’s soft-law model than to the EU’s binding Act. That means a consortium spanning EU and non-EU institutions typically needs to apply the OECD Principles as a governance floor everywhere, while layering EU AI Act obligations only where the EU leg of the project triggers them.

    Practical steps for research offices include:

    • Map every AI touchpoint across the consortium — admissions tools, grant-scoring assistants, participant-facing chatbots, drafting tools built on general-purpose models — to check whether the Article 2 research exemption still applies once a tool moves from pilot to operational use.
    • Treat the OECD Principles as the baseline for institutional AI ethics policy and grant conditions, since 47 governments, including most funder jurisdictions, already reference them.
    • Track the EU AI Act’s phased dates in agreements with EU partners: prohibited-practice compliance from February 2025, general-purpose AI model duties from August 2025, and most high-risk obligations from August 2026.
    • Flag any AI tool used in EU-facing admissions, proctoring or research-evaluation processes as a potential high-risk use under Annex III, requiring documentation and human oversight even where the underlying research itself remains exempt.

    The two frameworks are not on a collision course. The EU AI Act’s adoption of the OECD’s own definition of an “AI system” points toward gradual convergence in vocabulary, even as legal force diverges. Research offices that build their AI governance around the stricter of the two applicable layers, rather than the more comfortable one, will find both frameworks easier to satisfy as further OECD updates and EU implementing guidance arrive.

  • AI Literacy Obligation Article 4: Staff Training Checklist

    The AI literacy obligation Article 4 of the EU Artificial Intelligence Act has applied since 2 February 2025, yet many research institutions still treat it as a future item rather than a live requirement. Article 4 states that providers and deployers of AI systems “shall take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf.” For universities, funders, and research institutes running admissions tools, grant-screening software, or generative AI assistants, that duty is already in force. This article sets out a practical checklist for what “sufficient” means in a research setting, who it covers, and how to evidence it.

    What Article 4 actually requires

    Article 4 does not prescribe a fixed curriculum, test, or certificate. It requires institutions to calibrate AI literacy measures against staff technical knowledge, experience, education, training, and the context in which AI systems are used. The obligation applies to both “providers” (organisations that develop and place an AI system on the market) and “deployers” (organisations using an AI system for a professional purpose).

    A research institution can be either, sometimes simultaneously. A department that builds a bespoke research-data classification model is a provider of that system; the same university using an off-the-shelf tool for applicant screening or meeting transcription is a deployer. The AI Act scope includes a narrow exemption for systems developed and used exclusively for scientific research before market placement — but once a research tool is deployed operationally, that exemption falls away and Article 4 applies in full.

    The statutory definition of AI literacy, in Article 3(56), is: skills, knowledge and understanding that allow providers, deployers and affected persons to make an informed deployment of AI systems, and gain awareness of the opportunities and risks it can cause. The Commission’s AI Office has confirmed it will not impose a single mandatory format.

    Who counts as “staff” under the AI literacy obligation

    This is where most institutions under-scope their programmes. The Commission’s guidance clarifies that “staff and other persons” is broader than payroll headcount — it extends to anyone acting on the institution’s behalf, including contractors and service providers. In a research setting that typically means:

    • Researchers and PIs using generative AI tools for literature review, drafting, or data analysis
    • Research administrators and grants officers using AI-assisted screening or compliance-checking tools
    • HR and admissions staff using AI-based applicant or candidate screening systems
    • IT and research-computing staff configuring or maintaining institutional AI deployments
    • External contractors or visiting researchers granted access to institutional AI systems
    • Board and senior leadership, who need enough literacy to assess institutional AI risk

    Article 4 does not require every group to receive identical training: a developer configuring a high-risk system needs materially more depth than an administrator using a transcription tool. The table below illustrates how role and system risk map onto training depth.

    Personnel category Typical AI Act role Indicative literacy depth
    Data scientists building institutional AI tools Provider Technical: limitations, bias, risk mitigation, documentation
    Grants/admissions staff using screening AI Deployer Operational: output interpretation, human oversight, escalation
    Researchers using generative AI for drafting/analysis Deployer (or exempt if pre-market research use) General awareness: hallucination risk, confidentiality
    Contractors/visiting staff with system access “Other persons” acting on the institution’s behalf Baseline awareness proportionate to access level
    Governing board/senior leadership Deployer (oversight capacity) Strategic: risk appetite, resourcing, regulatory exposure

    Building a defensible AI literacy programme

    The Commission’s AI Literacy Q&A sets out a four-step minimum approach research institutions can adapt directly:

    • Establish a general understanding of AI within the organisation — what AI is used, where, and why
    • Determine the institution’s role for each system: provider, deployer, or both
    • Assess the risk level of each AI system in use, including any high-risk systems under Chapter III
    • Build literacy actions proportionate to staff knowledge gaps and the context of use

    Relying solely on a vendor’s instructions for use is explicitly insufficient — this mirrors the separate Article 26 duty on deployers of high-risk systems to ensure staff are “sufficiently trained” to exercise human oversight.

    What is Article 4 of the EU AI Act?

    Article 4 is the AI Act provision requiring providers and deployers to ensure a “sufficient level” of AI literacy among staff and other persons operating AI systems on their behalf. It has applied since 2 February 2025, ahead of most other AI Act obligations, and covers technical knowledge, training context, and the persons affected by the system.

    What is the definition of AI literacy in the EU AI Act?

    Article 3(56) defines AI literacy as the skills, knowledge and understanding that let providers, deployers and affected persons make informed decisions about deploying AI systems, while remaining aware of the opportunities, risks, and potential harms those systems can cause in their specific context of use.

    Who at a research institution needs AI literacy training under Article 4?

    Anyone dealing with an AI system’s operation or output on the institution’s behalf, not just IT or data-science staff. This includes researchers, administrators, HR and admissions teams, contractors, and senior leadership — with training depth proportionate to each group’s technical role and the risk level of the system they use.

    Does Article 4 require a certificate or formal training records?

    No. The AI Office has confirmed there is no mandated test or certificate for Article 4 compliance. Institutions should instead keep an internal record of training delivered, attendance, and content — evidence that becomes essential if a national market surveillance authority later reviews compliance.

    Documenting compliance: what evidence to keep

    Because Article 4 sets no certification standard, the practical question is evidential: what would a research institution show a national market surveillance authority if asked? A defensible record typically includes:

    • A written AI inventory identifying each system in use, its provider/deployer classification, and risk tier
    • Training content and delivery records (dates, attendance, format) mapped to each staff category
    • A documented rationale for why the chosen literacy measures were “sufficient” for each role and system
    • An AI use policy communicated to staff, contractors, and other relevant persons
    • A review cycle — literacy measures should be revisited as systems and risk profiles change

    National market surveillance authorities take over enforcement of Article 4 from 2 August 2026, when the AI Act’s general application and high-risk provisions take effect. Enforcement is meant to be proportionate — gravity, intent, and negligence are all considered — but an incident with no evidence of staff training is likely to weigh against an institution.

    Implications for research institutions and GPAI tools

    Research institutions increasingly deploy general-purpose AI (GPAI) tools — large language model assistants embedded in research-writing or literature-search workflows — rather than narrow, purpose-built systems. The AI Act GPAI provisions (Chapter V, applying to GPAI model providers from 2 August 2025) sit alongside, not instead of, the Article 4 duty on deployers: using a GPAI-based writing assistant still makes an institution a deployer under Article 4, regardless of the model provider’s own transparency obligations.

    One development worth tracking: the European Commission’s 2025 Digital Omnibus proposal would shift part of the Article 4 burden from individual organisations toward Member States and the Commission itself. That proposal has not been adopted, so the current text of Article 4 remains binding on institutions as providers or deployers — institutions should not defer action for a change that may not materialise as proposed.

    Sector signals reinforce the direction of travel: UK courts have separately expected legal professionals to demonstrate AI competence in submissions, and UK financial regulators have referenced the Senior Managers Regime in connection with AI risk oversight — evidence that “sufficient AI literacy” is becoming an expectation beyond the AI Act’s direct territorial reach.

    What research institutions should do next

    Institutions without a formal AI literacy programme should start with an AI system inventory, classify each system by provider/deployer role and risk tier, then build tiered training from that analysis rather than buying a generic module. Article 4 rewards documented judgement over box-ticking: the institutions best placed for enforcement from August 2026 will be those that can show a reasoned, evidenced, role-differentiated approach. Research-administration functions, which typically already own institutional policy documentation, are well positioned to lead this work alongside data protection, IT governance, and research integrity offices.

  • EU AI Act Research Exemption: What Article 2(6) Actually Covers

    A run of academic literature published since mid-2025 — an editorial in GRUR International, a peer-reviewed analysis in Nature’s npj Digital Medicine, and a widely cited Swedish doctoral paper — has converged on the same conclusion: the EU AI Act research exemption is far narrower, and far less certain, than most research offices assume. Regulation (EU) 2024/1689 does carve scientific research and development out of scope, but that carve-out is built from two separate provisions with different wording, different triggers, and different failure points. For institutions running AI-assisted studies, clinical trials, or general-purpose model development, misreading where the exemption ends is now a live compliance risk.

    What Article 2(6) actually says

    Article 2(6) of the AI Act states that the Regulation “does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development.” Two conditions must both be met: the system or model must be developed for scientific research, and it must be put into service — first used for its intended purpose — exclusively for that research. Recital 25 is the only interpretive gloss the legislative text offers, and it does not define “scientific research and development” further.

    Critically, Article 2(6) exempts systems that are put into service for research, but it does not extend to systems that are placed on the market. That distinction — put into service versus placed on the market, defined respectively in Articles 3(10) and 3(9) — is where the exemption’s practical limits begin.

    Two exemptions, not one: Article 2(6) vs Article 2(8)

    Law professor Michèle Finck’s October 2025 editorial “In Search of the Lost Research Exemption” (GRUR International, Vol. 74, Issue 10) makes the point that is most often missed: the AI Act contains two distinct research exemptions, not one. Article 2(6) is narrow and limited to scientific research; Article 2(8) is broader and covers any research, testing or development activity, scientific or not, but only up to the point of market placement or service.

    Provision What it exempts Key limit
    Article 2(6) AI systems/models developed and put into service solely for scientific research and development Not limited to pre-market stage, but strictly tied to “sole purpose” of research — loses protection once put into service for any other use
    Article 2(8) Any research, testing or development activity (not limited to science) regarding AI systems or models Applies only prior to placing on the market or putting into service; explicitly excludes real-world testing

    Finck argues that this dual structure creates an “interpretative conundrum”: if Article 2(8) only ever covers activity that happens before market placement, and market placement is already the trigger for the Act’s obligations regardless of the exemption, the provision risks adding little independent legal value — precisely the ambiguity that gives the “lost” exemption its name.

    Where the research exemption stops applying

    The Nature-published analysis by Meszaros and colleagues (npj Digital Medicine, 2026) sets out a conceptual framework built around a single regulatory threshold: placement on the market or putting into service. Everything on the research side of that line can be exempt; everything on the other side is regulated. Three scenarios repeatedly cross that line.

    Commercialisation and dual-purpose systems

    A system loses its exemption the moment it is not developed for the sole purpose of research. Finck highlights that Horizon Europe-style collaborations, where a university partners with an industrial co-investigator who intends to commercialise the output once the exploratory phase ends, sit in exactly this grey zone. Whether “commercial purpose” is assessed objectively (does a commercial partner exist) or subjectively (did the researchers intend commercialisation) remains unresolved in the text itself.

    Post-market deployment and real-world testing

    Article 2(8) states plainly that “testing in real-world conditions shall not be covered by that exclusion.” A model tested only in a closed lab environment can remain exempt; the same model tested on live users, patients, or public-facing systems generally cannot, unless it proceeds through the Act’s dedicated real-world testing and regulatory sandbox framework (Articles 57–61). Colonna’s 2024 analysis for the DiVA repository similarly stresses that the exemption was never intended to cover deployment-stage activity dressed up as “ongoing research.”

    GPAI models and systemic-risk obligations

    Because Article 2(6) explicitly names “AI models” alongside “AI systems,” a general-purpose AI (GPAI) model built and used solely for research is exempt. That exemption evaporates once a provider places the model on the Union market — including releasing a checkpoint for downstream use beyond pure research. From that point, Title VIII’s GPAI obligations under Article 53 (technical documentation, copyright-compliance summaries) apply, and models presumed to carry systemic risk — those trained with cumulative compute above 10^25 FLOPs — face the additional Article 55 duties regardless of open-source licensing. A separate, unconditional exclusion exists for military, defence and national-security AI under Article 2(3); that provision is absolute and is not contingent on “sole purpose,” unlike the research exemptions.

    Frequently asked questions

    What is Article 2(6) of the EU AI Act?

    Article 2(6) excludes AI systems and AI models — including their output — from the AI Act when they are specifically developed and put into service for the sole purpose of scientific research and development. It does not, however, exempt systems that have been placed on the market.

    Does the AI Act research exemption cover real-world testing?

    No. Article 2(8) states explicitly that testing in real-world conditions is not covered by the research exclusion. Researchers deploying systems outside a controlled setting generally need to use the Act’s regulatory sandbox and real-world testing framework instead.

    Are GPAI models exempt from the AI Act during research?

    Yes, while a general-purpose AI model is developed and used solely for research it falls outside scope. Once placed on the market, Title VIII obligations attach, with stricter Article 55 duties for models presumed to carry systemic risk above the 10^25 FLOPs training-compute threshold.

    Can university-industry collaborations rely on the research exemption?

    Only where the sole purpose remains scientific research. Per Finck’s 2025 analysis, Horizon Europe-style projects involving a commercial partner intending future exploitation risk losing Article 2(6)/2(8) protection once a profit-oriented purpose is established.

    What this means for research institutions and publishers

    Research administration offices — the ARMA, EARMA and INORMS community that oversees institutional compliance — now have a practical due-diligence question to add to AI-enabled research proposals: at what point does this project’s AI system move from “developed for research” to “put into service” or “placed on the market”? That question matters most for:

    • Clinical and biomedical AI tools that progress from retrospective lab validation to prospective real-world testing on patients.
    • Multi-partner Horizon Europe consortia where an industrial partner holds commercialisation rights from the outset.
    • Open-source model releases on code and model-sharing platforms, which several commentators — including the arXiv paper “Beware! The AI Act Can Also Apply to Your AI Research” — flag as a possible trigger for “placing on the market.”
    • Foundational research (for example, in AI explainability or causal reasoning) whose downstream applications are not yet known at the outset, which Finck notes may struggle to meet the “sole purpose” test even where no commercial partner is currently involved.

    Institutions with dedicated research administration functions are best placed to build this threshold assessment into ethics review and grant-agreement workflows now, rather than retrofitting compliance once a system reaches deployment.

    Looking ahead

    The AI Act’s general provisions, including Article 2’s scope rules, have applied since 2 February 2025; GPAI obligations followed on 2 August 2025; most remaining obligations, including high-risk system requirements under Annex III, become applicable from 2 August 2026. Every commentator reviewed here — Finck, Meszaros et al., and Colonna — reaches the same practical conclusion: the European Commission’s promised guidance on the research exemptions has not yet resolved the “sole purpose,” commercial-intent, and real-world-testing ambiguities in the text. Until that guidance lands, institutions should treat the exemption as a narrow, conditional safe harbour rather than a blanket shield, and document the specific research purpose, funding structure, and deployment plan for every AI system that currently relies on it.