The Next News

// technology and science news, explained

AINews3 mins read3 sources

EU science and ethics advisers urge independent testing of powerful AI and no weakening of AI rules

In an unusual joint statement on November 16, the Commission's science and ethics advisers say traditional guard rails no longer suffice for AI.

By AI Desk ·

The European Commission's two standing groups of outside advisers on science and ethics said on Monday, November 16, 2026, that governments need sharper tools to keep artificial intelligence safe as systems become more capable and more autonomous. In a joint statement, the Group of Chief Scientific Advisors and the European Group on Ethics in Science and New Technologies say AI raises challenges of its own for public policy, and that “traditional guard rails” are no longer enough.

Both groups advise the Commission independently: the chief scientific advisors are seven experts appointed in a personal capacity, and the ethics group advises where ethical, societal and fundamental rights questions meet new technology. They have each advised on AI before, but the Commission's research department called a joint statement from both groups an unusual step. The report says its contents do not necessarily reflect the Commission's position.

Four principles

The advisers' four principles, from the joint statement published November 16, 2026
PrincipleWhat the advisers ask for
Keep current law intactSafety and security remain conditions for developing and deploying AI, with evidence requirements that rise with a system's capability and potential consequences
Different speeds for different systemsFaster deployment where harms are limited and reversible; stronger evidence and safeguards in critical infrastructure, crisis management and essential public services such as health care
Independent evaluationEU capacity to test powerful models itself, including frontier models built outside Europe on which European science, government or essential services may depend
An ecosystem that learnsTesting before release, monitoring after it, and systematic recording of incidents and near misses

The advisers offer this “differentiated deployment” as a middle course between pausing AI and speeding it up: development goes on, but the bar for deploying a system rises with its capability, autonomy and possible consequences. They want reversibility to be a core condition, with extra caution where AI creates lock-in that would be hard to undo, and say closer scrutiny may be warranted for agentic systems that can take consequential actions without continuous human oversight. Those assessments, they write, should be “guided and controlled by public policy and not left to self-regulating technology companies.”

On evaluation, the statement says the EU should rely less on assessments made by developers and more on its own, with access to the models and evidence needed, adversarial testing, and checks on cybersecurity dual-use risks and agentic capabilities. It suggests shared evaluation infrastructure that public authorities and independent researchers could use, building on the AI Act's scientific panel and advisory forum. It also warns that “safety assessment cannot rely indefinitely on benchmarks designed for earlier generations of systems.”

Companies should never be the sole arbiters of whether their systems are safe, secure or compatible with fundamental rights.

Joint statement of the Group of Chief Scientific Advisors and the European Group on Ethics in Science and New Technologies

Enforcement before new regimes

The statement does not ask for a separate legal regime for AI. It argues that safety and security should stay prerequisites in line with existing Union law, that frontier AI should be no exception, and that the problem is sometimes not a lack of legislation but a lack of political will or institutional capacity to enforce it, compounded by what the advisers call “external efforts to delay or weaken” existing obligations. They do not name any specific proposal. The statement also says other concerns, from discrimination and privacy to copyright, working conditions and children's wellbeing, must not be neglected in debates about more speculative risks.

Where the EU's AI rules stand

The AI Act entered into force on August 1, 2024. Eight of its bans on certain AI practices have applied since February 2025, with a ninth, added by the AI Omnibus, due in December 2026, and its rules for general-purpose AI models have applied since August 2025. Under the AI Omnibus, a simplification law that took effect on July 27, 2026, rules for high-risk systems in sensitive areas such as employment, education and border control apply from December 2, 2027, and for AI built into products such as toys or lifts from August 2, 2028. The Commission says the timing ensures the rules apply when companies have support tools such as standards. Since August 2, 2026, its AI Office has had powers to evaluate general-purpose models and fine their providers.

By our reading, the closest existing step to the advisers' third principle is in the Commission's July 2026 action plan on cybersecurity and AI, which plans a call to increase EU capacity to evaluate AI models before they reach the EU market, expected to be operational by 2027. Our explainer covers the 2025 code of practice that guides general-purpose model providers.

Sources

  1. AI requires rigorous safeguards and rules that keep pace with technological advances – experts tell the European Commission European Commission, Directorate-General for Research and Innovation research-and-innovation.ec.europa.eu
  2. The next phase of artificial intelligence: Governing for safety, security and human agency. Joint EGE GCSA Statement European Commission, Group of Chief Scientific Advisors and European Group on Ethics in Science and New Technologies (Publications Office of the EU) op.europa.eu
  3. AI Act: regulatory framework for AI European Commission digital-strategy.ec.europa.eu

Checked against the 3 sources above on November 18, 2026. Spotted an error? Tell us.