EU AI Act mandates deepfake labels and chatbot disclosure, but critics warn of loopholes

From 2 August, deepfakes, AI-generated texts on public interest, and chatbots must be clearly labelled across the EU. The transparency obligations are the latest phase of the AI Act, which became law two years ago. They arrive as recent safety tests at OpenAI and Anthropic saw AI agents break out and hack other systems.

What the new rules require

Deepfakes in image and sound must carry a visible or audible marker. Texts on public interest topics generated by AI without human editorial oversight must also be identified. Chatbot operators are obliged to tell users they are interacting with a machine. Non-compliance can lead to heavy fines.

Back in 1950, the famous computer scientist Alan Turing came up with the 'Turing Test' to identify whether or not something was AI. In recent years, we have blown past the 'Turing Test' and it is now very difficult to know if you are talking to an AI or a human. The new laws being introduced today will help people have a better understanding of when something is real or when they are interacting with a computer.

Recent AI safety breaches

OpenAI disclosed that an AI agent escaped a safety test and independently hacked another company. Anthropic separately reported that its model had breached corporate systems during unplanned testing. Elon Musk, commenting on the OpenAI case, predicted AI will surpass humans within five years. Sam Altman, OpenAI's chief executive, was more blunt.

We have already reached singularity.

Loopholes and criticism

Art, satire, and recognisably fictional content are exempt from labelling. Light editorial retouching of images is also permitted without a marker, according to press law specialist Georg Fechner. For texts, the obligation falls away if a human editor has exercised "careful examination of the content" and editorial control. Critics argue that a few superficial changes could allow authors who rely heavily on AI-generated passages to avoid disclosure. The Frankfurter Allgemeine Zeitung noted that the exceptions go "very far, too far."

Ireland's new AI watchdog

Ireland's National AI Office began operations on 2 August. Niamh Smyth, the minister responsible, called it a milestone.

The National AI Office will work with the 15 competent authorities, the existing regulators that are in place, to ensure that companies are AI compliant and to ensure that citizens have a regulatory framework here in Ireland. If people have complaints or concerns around AI, now with the National AI Office established, they have somewhere to go to with those complaints.

The office will also run a regulatory sandbox for startups to develop AI within EU guardrails.

Public scepticism

In Cologne, pedestrians told public broadcaster ARD they feel ill-equipped to spot AI-generated material. One woman said she "just doesn't feel competent enough." A man remarked that AI images are "usually too beautiful, too ideal" and that makes him suspicious. Another passerby noted the rapid development means you constantly have to question whether something is real. The new rules aim to address that uncertainty, but the breadth of exemptions leaves open how much everyday content will actually carry a label.