The Scientific Panel of Independent Experts (Article 68): What It Is and What It Can Do

The scientific panel is the EU AI Act's standing body of independent technical experts. It is established by Article 68 of Regulation (EU) 2024/1689, and its single most consequential power sits in Article 90: it can issue a qualified alert to the AI Office when it suspects a general-purpose AI (GPAI) model poses a systemic risk. If you only remember one thing, remember that — the panel is the technical channel through which expert concern about a frontier model becomes a formal trigger for Commission scrutiny.

It is one of four governance bodies created by the Act, and the four are routinely confused. This page defines the scientific panel and then draws the lines between it, the European Artificial Intelligence Board (Article 65), the advisory forum (Article 67), and the AI Office.

The Article 68 definition

Article 68 requires the Commission, by implementing act, to establish a scientific panel of independent experts to support the enforcement activities of the Act — in particular those concerning GPAI models and systems. Members are selected for scientific or technical expertise in the AI field, must be independent of any AI system or GPAI model provider, and must be able to act objectively and impartially. The Commission sets the number of members according to need and may draw on them individually or as a group.

The panel is advisory and technical, not regulatory. It does not issue fines, designate models, or adopt binding decisions. It supplies the specialised expertise the AI Office and national authorities need to enforce rules that turn on questions — model capabilities, compute, evaluation methods — that ordinary regulators are not equipped to assess unaided.

What the panel actually does

Article 68 gives the panel four working functions:

• Support GPAI enforcement. It advises the AI Office on the implementation and enforcement of the GPAI rules in Chapter V — including helping develop evaluation methods and tools for assessing the capabilities of GPAI models.
• Issue qualified alerts (the Article 90 power). Where the panel has reason to consider that a GPAI model could present a concrete, identifiable systemic risk at Union level, it may alert the AI Office. This qualified alert is the panel's sharpest instrument and is described separately below.
• Advise on classification and cross-border issues. It can contribute to questions about how models and systems should be assessed, and assist market-surveillance authorities and the AI Office on tools and methods.
• Support national authorities on request. Member-state authorities may request advice from the panel, including to support their market-surveillance work, subject to the conditions in the implementing act.

The Article 90 qualified alert — the panel's real lever

Article 90 is where Article 68 acquires teeth. A qualified alert is a formal, reasoned warning from the scientific panel to the AI Office that a specific GPAI model presents a systemic risk — or that its capabilities are not adequately captured by the existing classification, for example where a model approaches or exceeds high-impact capability without crossing the headline compute threshold.

A qualified alert is not itself a designation. It does not make a model systemic-risk by operation of law. What it does is oblige the AI Office to take the concern seriously: the alert feeds into the Commission's power under Article 51 to designate a model as having systemic risk on a capability basis, even where the 10²⁵ FLOP training-compute presumption was not triggered. In effect, the panel is the expert tripwire for models the compute threshold alone would miss — frontier systems made capable through architecture, data quality, or post-training rather than raw compute. The alert can be accompanied by the panel's reasoning and supporting evidence, which the AI Office then weighs.

How the panel differs from the Board, the forum, and the AI Office

This is where most confusion lives. Four bodies, four jobs:

The distinctions that matter in practice:

• The Board (Art 65) is intergovernmental — member-state representatives coordinating enforcement across the Union. The scientific panel is expert — individuals chosen for technical merit, sitting independently of any government or provider.
• The advisory forum (Art 67) is stakeholder-balanced by design and feeds general input upward; it has no qualified-alert power. The scientific panel is independence-constrained — members cannot be tied to a provider — precisely because its job includes flagging specific models.
• The AI Office is the enforcer and the recipient of the panel's alerts; the panel advises and warns but does not enforce.

A short way to hold it: the Board coordinates governments, the forum channels stakeholders, the AI Office enforces, and the scientific panel supplies independent expertise and pulls the systemic-risk alarm.

Why an SME should care (when most of this binds someone else)

For the overwhelming majority of small and mid-sized EU companies, the scientific panel is a body they will never interact with. Its remit is the GPAI/model layer — the firms that train frontier models — not the deployers and downstream providers who are Confir's core audience. None of the Article 99 penalty tiers (€35M/7%, €15M/3%, €7.5M/1% — the third tier is 1%, never 1.5%) are administered by the panel; it imposes no fines at all.

But the panel matters indirectly, and it is worth one paragraph of attention. If the panel issues an Article 90 alert and the AI Office subsequently designates a foundation model you build on as systemic-risk, that model's provider acquires the heavier Article 55 obligations (evaluation, adversarial testing, incident reporting, cybersecurity). That can change the model's documentation, its terms, and occasionally its availability — which flows down to you. Knowing where systemic-risk concern originates helps you read those upstream changes correctly.

Worked example: a mid-market AI vendor reading the signals

Helvora AI SAS, a French model-and-tooling vendor with about 70 employees and roughly €14M turnover, builds developer tools on top of a large third-party foundation model. Helvora is a downstream provider, not a GPAI model provider — it does not train the foundation model and is nowhere near the 10²⁵ FLOP threshold. So the scientific panel and Article 90 do not bind Helvora directly.

What Helvora's compliance lead watches for is upstream signal. When a press report suggests the scientific panel has flagged a frontier model to the AI Office, she does not panic — a qualified alert is not a designation. But she does open the model vendor's documentation and terms, because a subsequent Article 51 systemic-risk designation would add Article 55 duties on that vendor and could change the model card, the usage terms, or the model's lifecycle. Her own obligations stay where they were: classify each Helvora product by its use under Article 6, and meet provider duties under Article 16 for anything she places on the market under Helvora's name. The panel's existence sharpens her monitoring; it does not create a new duty for her.

How Confir helps

Confir's rule-based engine keeps GPAI-layer concepts in their lane: it scopes whether your organisation is a GPAI model provider, a downstream provider, or a deployer, and routes you to the duties that actually apply — without ever implying a panel alert or a systemic-risk designation lands on a downstream company that it does not touch. The classification is deterministic and reproducible: the same intake yields the same finding, and the rule that fired is human-readable.

Frequently Asked Questions

Can the scientific panel fine my company?

No. The panel has no power to impose penalties or take binding enforcement action. It is an advisory and technical body under Article 68. Fines for AI systems are imposed by national authorities under Article 99; fines on GPAI model providers are imposed by the Commission under Article 101; the EDPS fines EU institutions under Article 100. The panel's role is to advise the AI Office and, through an Article 90 qualified alert, to flag a possible systemic risk — not to sanction anyone.

What is the difference between the scientific panel and the European Artificial Intelligence Board?

The Board (Article 65) is composed of one representative per member state and coordinates national authorities and consistent application across the Union — it is intergovernmental. The scientific panel (Article 68) is composed of independent technical experts selected for scientific merit and barred from ties to any provider; it supports GPAI enforcement and can issue qualified alerts. One coordinates governments; the other supplies independent expertise.

What is a qualified alert under Article 90?

A qualified alert is a reasoned warning from the scientific panel to the AI Office that a general-purpose AI model may present a systemic risk. It is not a designation — it triggers Commission review and can feed an Article 51 systemic-risk designation, including for models that did not cross the 10²⁵ FLOP compute presumption. It gives technical experts a formal channel to flag capability concerns the compute threshold alone would miss.

Does the scientific panel decide whether my AI system is high-risk?

No. High-risk classification of AI systems runs through Article 6 and Annex III and is assessed by the provider, with oversight from market-surveillance authorities. The scientific panel's remit is the GPAI/model layer — supporting enforcement of Chapter V and flagging systemic-risk concerns about GPAI models. It does not classify ordinary high-risk AI systems used by deployers.

Is the advisory forum the same as the scientific panel?

No. The advisory forum (Article 67) is a balanced stakeholder body — industry, SMEs, civil society, academia — that provides input to the Board and the Commission. The scientific panel (Article 68) is a body of independent technical experts with a narrower, enforcement-supporting mandate and the Article 90 alert power. The forum represents stakeholders; the panel supplies expertise and watches the frontier.

When was the scientific panel established?

Article 68 requires the Commission to establish the panel by implementing act, and its work concerns GPAI obligations, which have applied since 2 August 2025. The panel is part of the governance architecture that stands up alongside the AI Office. As with all of the Act's institutional bodies, the practical detail of its membership and procedures is set out in the implementing act rather than in the Regulation itself.

Related guides

• Systemic risk — definition and obligations — the category an Article 90 alert can ultimately trigger
• AI Office — the Commission body that receives qualified alerts and supervises GPAI
• Article 51 — systemic-risk classification — the designation mechanism the panel's alert feeds into
• General-purpose AI model — the subject of the panel's enforcement-support remit
• Competent authority — the national bodies the panel can advise on request
• GPAI compliance — the full Chapter V programme the panel helps enforce