Member States advance amendments to the EU AI Act: What it means for businesses

In March 2026, the Council of the EU adopted its negotiating position on proposed amendments to the EU AI Act. Days later, the European Parliament’s committees voted 101 to 9 in favour of their joint position. Trilogue negotiations between the three institutions will follow.

For companies operating in the European market or deploying AI products, these changes matter. They affect AI Act compliance timelines, the scope of prohibited practices, regulatory oversight, sensitive data processing, and penalty structures.

Important: until these amendments are formally adopted, the current regulation remains fully in force. This includes the 2 August 2026 deadline for high-risk AI systems.

New prohibited practices: deepfake intimate imagery and CSAM

The Council added two new categories to the list of banned AI practices. This is the most significant departure from the Commission’s original proposal. AI systems capable of generating or manipulating non-consensual intimate imagery or child sexual abuse material (CSAM) would be explicitly prohibited.

The prohibition has two dimensions. First, it covers systems whose intended purpose is to produce such content. Second, it applies to systems where such outcomes are “reasonably foreseeable” and the safety measures are not effective enough. This second point is more consequential for developers.

In practice, providers will need to demonstrate that their safeguards work well enough. This includes refusal training, content filtering, usage restrictions, and abuse detection mechanisms.

Delayed timelines: more time to prepare

One of the most anticipated changes is the postponement of compliance deadlines for high-risk AI systems. Under current law, the deadline is 2 August 2026 for Annex III systems. For product safety systems under Annex I, the deadline is 2 August 2027.

The Council’s position is cleaner. Fixed dates, no link to the readiness of standards. This gives companies more predictability.

Regulatory oversight: one regulator instead of many

Today, national market surveillance authorities oversee AI systems. The EU AI Office handles general-purpose AI (GPAI) models. This sometimes creates a dual-jurisdiction problem. A single company may face investigations from multiple regulators at once.

Both the Commission and the Council propose granting the EU AI Office exclusive competence over AI systems built on GPAI models. The condition is that the same provider must have created both the model and the system. In practice, platforms like ChatGPT, Gemini, Claude, or Copilot would have a single regulator.

The Council, however, carves out several sectors. Critical infrastructure, law enforcement, border management, financial institutions, and justice remain under national control. Member states want to preserve sovereignty in these areas.

AI literacy obligation removed

Since February 2025, providers and deployers of AI systems have been required to ensure AI literacy among their staff. Both the Commission and the Council propose removing this broad obligation. Instead, the Commission and member states will promote AI literacy at the institutional level.

The Council adds a clarification: existing competency requirements for personnel working with high-risk AI systems remain in effect. Companies that have already invested in training will not lose anything. The general obligation simply falls away.

Sensitive data processing for bias detection

Currently, only providers of high-risk AI systems may process sensitive personal data (ethnicity, health, political opinions) for detecting biases in their models. This is only allowed when it is “strictly necessary.” Both institutions want to extend this right more broadly.

The Commission proposes lowering the threshold from “strictly necessary” to “necessary.” The Council keeps the higher standard. The Council also limits the expansion to specific types of biases affecting health, safety, fundamental rights, or discrimination. The Council’s position offers more protection for personal data.

Database registration and penalty proportionality

Today, all Annex III high-risk AI systems must be registered in the EU public database. This includes those assessed as non-high-risk under Article 6(3). The Commission proposes eliminating the registration requirement for such exempted systems. The Council prefers a compromise: registration stays, but with less information required.

Separately, both institutions agree on extending penalty proportionality. Today, lower penalty caps apply only to SMEs. The proposal extends this to small mid-cap companies: those with up to 750 employees and up to €150 million turnover. This is a meaningful relief for many European businesses.

EU AI Act: What comes next?

A plenary vote in the European Parliament is expected on 26 March 2026. Trilogue negotiations will follow. Historically, this process takes around 19 months on average. The final outcome may differ significantly from the current positions. Until formal adoption, the existing regulation applies in full.