CASRAI Dictionary

Tag: general purpose ai model

  • Foundation Model vs General-Purpose AI Systems Under the EU AI Act

    A foundation model is a large-scale AI model trained on broad data that can be adapted to many downstream tasks; a general-purpose AI system is the deployed product built on top of it. The foundation model vs general-purpose AI distinction is not academic under EU Regulation 2024/1689 (the AI Act): “general-purpose AI model” obligations attach to the model itself under Chapter V, while “AI system” obligations attach to a deployed application under Title III and depend entirely on how that application is used. For a university lab that fine-tunes an open-weight model into a research tool and releases it publicly, which category applies — and when — determines a materially different compliance workload.

    The AI Act defines a general-purpose AI model (Article 3(63)) as one that “displays significant generality” and can “competently perform a wide range of distinct tasks”, regardless of how it is placed on the market. A general-purpose AI system (Article 3(66)) is an AI system based on such a model that can serve a variety of purposes, whether used directly or integrated into other systems. The two terms are frequently conflated, but the Act’s compliance architecture depends on keeping them separate.

    What is a foundation model? A working definition

    A foundation model is an AI model trained on large, broad datasets — often using self-supervised learning — that generalises to a wide range of downstream tasks rather than being built for one narrow purpose. Large language models are the most cited example, but the concept also covers text-to-image, text-to-video, and multimodal models.

    The term originated in AI research (Stanford’s Center for Research on Foundation Models coined it in 2021) and is not itself a defined legal term in the AI Act. The Act’s operative category, general-purpose AI model, captures the same phenomenon in enforceable language, and in practice most models researchers call “foundation models” meet that legal definition.

    AI Act definitions: general-purpose AI model vs AI system

    The AI Act runs two separate regulatory tracks that a research team must understand independently, because passing one does not exempt a project from the other.

    The general-purpose AI model track (Chapter V, Articles 51–56) regulates the model layer. Under Article 53(1), providers must draw up technical documentation, inform downstream integrators, publish a training-content summary, and maintain a copyright policy. The Commission’s AI Office guidance is explicit that these “provisions… apply to the model itself, regardless of what is or will be its ultimate use.” Models trained with over 1025 FLOP are presumed to pose systemic risk under Article 51(1)(a), triggering evaluation and incident-reporting duties under Article 55.

    The AI system track (Title III, Articles 6–49) regulates the application layer. Obligations here — including the high-risk conformity-assessment regime under Annex III — depend entirely on context of use: what the system does, who it affects, and which domain it operates in. A system built on a fully documented foundation model can still be high-risk; one built on an under-documented model can still sit outside Title III if its use case is low-risk.

    Crucially, Article 2(8) excludes “any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service” from the Act’s scope. This is the single most important provision for university labs: internal experimentation is largely unregulated, but that exemption ends the moment a tool is placed on the market — including released publicly under an open licence.

    When does fine-tuning create a new “provider”?

    Recital 97 clarifies that a model is still “placed on the market” if its provider integrates it into their own AI system, unless use is purely internal, third parties’ rights are unaffected, and the model carries no systemic risk. The Commission’s guidelines on scope (Section 3.2) add an indicative criterion: a downstream entity that fine-tunes an existing model becomes provider of a new general-purpose AI model once the modification’s training compute exceeds one-third of the original model’s training compute. Below that threshold, the original provider keeps Article 53 responsibility, and the lab’s own duties attach mainly at the system layer.

    Why the distinction matters for research software

    University labs increasingly build research tools — literature-screening assistants, peer-review triage systems, integrity checkers, grant-matching recommenders — on foundation models. Getting the model/system distinction wrong creates two risks: over-compliance (treating a low-risk tool as needing Annex III conformity assessment) or under-compliance (releasing a public tool without required Article 53 documentation).

    Two provisions reduce the burden for academic and open-source releases. First, Article 53(2) exempts providers from the Article 53(1)(a) and (b) documentation duties if the model is released open-source with public weights, architecture, and usage information — though this never extends to systemic-risk models. Second, Recital 109 requires that provider obligations “take due account of the size of the provider,” with simplified routes for SMEs and start-ups, a category many university spin-outs fall into.

    Neither exemption touches the AI system track. A tool that screens grant applicants or assesses researcher performance may still fall under Annex III’s high-risk categories regardless of how the underlying model is licensed.

    Worked examples for university labs

    The scenarios below show how the two tracks diverge depending on what is built and how it is released.

    Scenario General-purpose AI model duties AI system duties
    Fine-tunes an open-weight model for internal literature screening only, never released Not applicable — Article 2(8) research exemption while internal Not applicable — no placing on the market
    Releases the fine-tuned model publicly, open licence, weights published; fine-tuning compute under one-third of base compute Base-model provider keeps Article 53 duties; lab is not a new provider (one-third-compute criterion) Must assess use case under Title III; low-risk screening unlikely to trigger Annex III
    Substantially retrains a model (compute over one-third of base) and releases it as a research-assessment tool used in hiring-adjacent decisions Lab becomes provider of a new general-purpose AI model; full Article 53 duties apply Likely high-risk under Annex III (employment use); conformity assessment and human oversight required from 2 August 2026

    Frequently asked questions

    Are foundation models general purpose?

    Most, but not all. A model must display significant generality and competently perform a wide range of distinct tasks to meet the AI Act’s Article 3(63) definition; models fine-tuned to be narrowly specialised for one domain can fall outside it even if built on a general-purpose base.

    Is a chatbot built on a foundation model itself a foundation model?

    No. A chatbot is a deployed AI system; the language model behind it is the general-purpose AI model. The chatbot’s obligations run through Title III based on use case, while the model’s obligations run through Chapter V regardless of that use case.

    What is the difference between a foundation model and an AI agent?

    A foundation model is the underlying trained model; an AI agent is a system that uses one or more models to pursue goals with some autonomy, such as planning or calling tools. Under the AI Act, an agent built on a foundation model is assessed as an AI system, with obligations set by context of use, not by which model powers it.

    Compliance implications and what comes next

    For research offices evaluating an AI-powered tool before release, four checks separate compliant projects from exposed ones:

    • Confirm whether the tool is still covered by the Article 2(8) research exemption, or whether release plans amount to “placing on the market.”
    • Calculate whether fine-tuning compute crosses the one-third-of-original-compute threshold that can make the lab a new model provider.
    • Check the Article 53(2) open-source exemption conditions — weights, architecture, and usage information must all be public.
    • Assess the deployed system separately against Title III and Annex III; model-layer compliance never substitutes for system-layer risk classification.

    The compliance calendar is tightening. Prohibited AI practice rules have applied since 2 February 2025, general-purpose AI model obligations since 2 August 2025, and high-risk AI system obligations under Annex III take effect from 2 August 2026 — weeks away for any lab planning to release research software touching employment, education, or other Annex III domains. Labs that map tools against both tracks before release, not after, avoid finding out the hard way through a market-surveillance enquiry. Institutional research administration offices are increasingly the first checkpoint for that mapping.

  • General-Purpose AI Model Rules for Universities

    A general-purpose AI model is defined in Article 3(63) of the EU AI Act as an AI model that displays significant generality, can competently perform a wide range of distinct tasks, and can be integrated into varied downstream systems, regardless of how it is placed on the market. A university prototype never placed on the market or put into service falls outside these duties under the Article 2(8) research exemption; the moment that model, or a fine-tuned derivative, is published or supplied to a third party, Article 53 obligations on technical documentation, downstream information, copyright policy, and a training-content summary can attach.

    A general purpose ai model is not defined by branding — it is defined by a compute-and-capability test in the Act’s recitals and the Commission’s guidelines. That test matters for universities, because the same model can sit inside the research exemption one day and carry a regulated “provider” obligation the next, depending on what happens to it.

    What is a general-purpose AI model under the AI Act?

    Under Article 3(63) of the AI Act, a general-purpose AI model is one that “displays significant generality and is capable of competently performing a wide range of distinct tasks… and that can be integrated into a variety of downstream systems or applications.” Recital 98 anchors this operationally: models trained with at least a billion parameters using large-scale self-supervision are presumed to meet the generality bar.

    The Commission’s guidelines on the scope of GPAI obligations add a practical marker: an indicative threshold of 10^23 FLOP of training compute, combined with the ability to generate language, images, or video, points to GPAI status. A narrow classifier trained on one labelled dataset for one task — say, detecting a specific cell morphology in a pathology dataset — meets neither bar and is not a GPAI model; it is a narrow AI system governed, if at all, by the Act’s risk-based system rules in Chapters II–IV, not the model-layer rules in Chapter V.

    Foundation model vs general-purpose AI vs narrow research tool: the definitional test

    “Foundation model” is not a term the AI Act defines — it originates from the 2021 Stanford Center for Research on Foundation Models literature. What most people call a foundation model is simply a species of general-purpose AI model; the Act creates no separate legal category for it and regulates instead at the model layer (Chapter V) and the system layer (Chapters II–IV).

    A narrow research tool is not a defined legal term either. It survives through Article 2(8), which states the AI Act “does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service.” The three-way distinction that matters for universities is:

    Category Legal basis Obligation trigger
    General-purpose AI model Article 3(63), Recital 98 Placed on the market or put into service (any use case, any modality)
    “Foundation model” (industry term) No AI Act definition — a subset of GPAI in practice Same as GPAI: placement on the market
    Narrow research tool Article 2(8) exemption Never triggered while confined to pre-market research, testing or development

    The trigger is not architecture, size, or pedigree — it is whether the model has been placed on the market, defined in Article 3(9) as first supply for distribution or use in the Union, paid or free. A lab training a model for internal experiments stays inside Article 2(8); publishing its weights publicly, or licensing it to a partner, places it on the market and can trigger Article 53.

    What are the Article 53 obligations for GPAI providers?

    Article 53(1) of the AI Act sets four core obligations for any provider of a general-purpose AI model placed on the EU market, regardless of size or academic status:

    • Technical documentation (Article 53(1)(a)) — training/testing processes and evaluation results, available on request to the AI Office and national authorities.
    • Information for downstream providers (Article 53(1)(b)) — capabilities, limitations, and foreseeable misuse, so downstream builders can meet their own duties.
    • A copyright policy (Article 53(1)(c)) — how the provider complies with EU copyright law and respects rights reservations in training data.
    • A public training-content summary (Article 53(1)(d)) — using the Commission’s template, describing what content trained the model.

    These obligations have applied since 2 August 2025 under Article 113(b), with transitional rules for pre-existing models under Article 111(3). A heavier tier applies only to GPAI models posing systemic risk — an indicative 10^25 FLOP training-compute threshold under Article 51(1)(a)-(2), costing tens of millions of euros to reach per the Commission — triggering adversarial evaluation, risk mitigation, incident reporting, and cybersecurity duties under Article 55. Almost no university-trained model approaches 10^25 FLOP, so this tier rarely applies academically; base Article 53 duties can still apply well below it.

    How does the research exemption apply to universities that fine-tune models?

    Universities interact with the GPAI rules from two directions, and conflating them is the single most common compliance error. First, a university training its own large model purely for internal research stays inside Article 2(8) for as long as the model is neither placed on the market nor put into service — a threshold not crossed the moment a paper is published, but more likely triggered by supplying the model for use or distribution.

    Second, and less understood, is what happens when a university fine-tunes an existing GPAI model from a commercial lab. Recital 109 and the Commission’s guidelines indicate fine-tuning can make the modifier a “provider” of a new model — but only for the modification, not the underlying base model. The indicative criterion: if fine-tuning compute exceeds roughly one-third of the original model’s training compute, the university takes on provider obligations for that modification, typically discharged by supplementing existing documentation with what changed.

    Two further carve-outs matter for university open-model releases:

    • The open-source exemption (Article 53(2)) removes the documentation and downstream-information duties (53(1)(a)-(b)) for models released free and open-source with public parameters, weights, architecture and usage information — not the copyright-policy or training-summary duties, and never for a systemic-risk model.
    • Downstream system duties are separate. A university embedding a GPAI model in a chatbot, decision-support tool, or admissions-screening application must separately assess that system against the Act’s risk-based rules in Chapters II–IV, whether or not it is also a GPAI “provider.”

    Answer-first Q&A

    Is general purpose AI the same as generative AI?

    No. Generative AI describes output type — models producing text, images, audio or video — while general-purpose AI is a legal classification under Article 3(63) based on generality and integration potential. Most large generative models qualify as GPAI, per Recital 99, but the two terms describe different properties and are not interchangeable in the regulation.

    What is an example of general-purpose AI?

    Large language and multimodal models generating text, images, or code across many tasks — commonly called foundation models — are the Act’s typical example under Recital 99. A university-trained model only qualifies once it meets Recital 98’s generality and compute markers, not merely because it shares similar architecture.

    Is ChatGPT general-purpose AI?

    Yes. Widely used commercial assistants built on large generative models meet the Article 3(63) generality test and are treated as general-purpose AI models, subject to Article 53 obligations from their provider. This differs from a university’s narrow, single-task research classifier, which would not meet the same test.

    Implications for universities releasing open models

    The practical stakes are documentation discipline, not automatic prohibition. A university planning to publish a fine-tuned or original model should map, before release, whether Article 2(8) still applies, whether the one-third fine-tuning threshold has been crossed, and whether Article 53(2)’s open-source exemption covers the intended licence. Institutions publishing under permissive open licences with full weights and architecture disclosure can shed the heaviest documentation duties while still owing a copyright policy and training-content summary — obligations echoing the transparency practices research administrators already apply to funder mandates such as UKRI’s and Horizon Europe’s.

    As the AI Office refines its guidelines and the General-Purpose AI Code of Practice — assessed as adequate by the Commission and AI Board in 2025 — becomes the default route to demonstrating compliance, universities treating model-release governance as a standing institutional process, not a one-off legal review, will be best placed to keep publishing openly without falling foul of Article 53.