The European Union (EU) has pushed back to December 2027 the date when AI Act obligations will apply to AI systems that monitor workers' job performance. In the same week, at an international AI summit in Dublin, Ireland, a co-founder of Google DeepMind described a recent series of AI incidents as an 'early warning' that the technology can cause harm to people. The warning has come now. The rules will take effect only 14 months from now.
The Dublin summit was a major event held while Ireland holds the presidency of the Council of the EU, and it brought politicians, regulators and technology business leaders from various countries together in one conference hall. Figures from companies including Microsoft also spoke. On a stage watched by regulators, the words 'early warning' came from the technology industry side.
Cases in which AI agents were involved in hacking and false reports were cited as grounds for the warning. An AI agent is a program that, once a person sets only the goal, carries out multiple steps in sequence on its own, such as searching, creating files and sending messages. Because no person checks each step, an action that goes wrong once flows straight out into the outside world.
The summit also heard calls for the industry to put strict guardrails and standards in place. Guardrails refer to technical and institutional barriers that block in advance actions an AI should not take.
The date for applying the rules has been pushed back.
The EU AI Act groups uses that have a major impact on people's livelihoods and rights as 'high-risk' and imposes separate obligations on them. The worker surveillance category has been delayed to December 2027, and application of obligations to the employment category, which covers hiring screening, has also been pushed back with December as the deadline.
AI hiring tool company Humanly acquired recruitment marketing company CareerArc, its fifth acquisition since 2025, and did not disclose the price. London startup VitVio is installing Vetro, a platform made up of cameras and sensors, in operating rooms at the University Clinical Centre in Gdańsk, Poland. It is the first hospital in Poland to bring in the platform.
AI and sensors are being used for tasks ranging from spreading job postings to screening applicants and watching people at work, but the rules to govern them will apply only at the end of 2027.
At another international economic event the same week, Bharti Airtel Chairman Sunil Mittal also gave a similar warning in different words. He said there is no need to hold back AI growth but that responding after an accident has happened is foolish, and that regulation must be designed now.
Warnings alone cannot prevent agent incidents. The calls at the summit for guardrails and standards also came because there is no common standard that sets where agents should stop. Until December 2027, each company will also decide who tests the guardrails and who is responsible when an incident occurs.
Until that date arrives, job applicants submitting résumés in Europe and employees working under surveillance software will not be covered by the safeguards the AI Act promised for high-risk areas.
