Dublin's Two Summits: What the AI Act Owes the People Already Out of Work

Two conferences took place in Dublin last week. One was the signature event of Ireland's Council presidency, where executives of the largest AI firms discussed, in the usual register, how to harness the technology and minimise its risks. The other, the weekend before, was called the Workers' AI Counter Summit. The second was not on anyone's protocol list. It will matter more to how the AI Act is remembered.

Start with the text. Article 26 of Regulation (EU) 2024/1689 obliges deployers of high-risk AI systems to assign human oversight to natural persons who have the necessary competence, training and authority. Article 27 requires a fundamental rights impact assessment before a public-sector body or certain private deployers put such a system into service. Recital 92 is explicit that systems used for recruitment, task allocation, performance evaluation and termination fall squarely inside Annex III, point 4 — the workplace. These provisions were negotiated with precisely the Dublin scenario in mind: a worker whose job is reorganised, downgraded or made redundant by a system she cannot inspect.

The timing is where the compromise bites. The prohibitions in Title II applied from February 2025. The general-purpose model obligations began in August 2025. The high-risk workplace regime in Annex III does not become fully enforceable until August 2026 for most provisions, with a longer runway for systems already on the market. Ireland's tech union, founded in 2023 with a few hundred members, now reports more than 1,350. A regional outlet covering the Dublin scene puts the increase at roughly fivefold, driven by waves of layoffs the union organiser attributes to post-pandemic correction and the diversion of budgets into AI. Meta has roughly halved its Irish headcount since the pandemic. A Meta subcontractor cut more than 700 roles earlier this year. A video platform shed around 300 Irish jobs, most of them in trust and safety, moving the work to other jurisdictions, outsourced vendors or automated solutions. The IMF has warned that Ireland's labour market is more exposed to AI than most advanced economies.

None of this triggered the AI Act's workplace provisions. The decisions were taken as restructurings, not as deployments of a high-risk system. That is the architectural weakness Annex III was supposed to close, and the one its phasing has not yet reached.

Who wrote this compromise? The Council text that became Annex III, point 4 was tightened under the Spanish presidency in late 2023 after Parliament rapporteurs pushed to include algorithmic management explicitly. Paris pressed for carve-outs to protect what it called sovereign industrial reorganisation. Dublin, holding one of the larger stakes in how workplace provisions would be enforced against companies headquartered on its territory, was quieter in the room than its exposure would suggest. The transitional periods are the trace of that quiet.

Enforcement will be the test, as it always is. The AI Office in Brussels handles general-purpose models. National market surveillance authorities handle Annex III. Ireland's designated authority will inherit a docket in which the complainants are, increasingly, organised. A union that has quintupled its membership is a union that files. The question for the Irish authority is whether it treats a mass redundancy accompanied by an internal productivity-ranking tool as a deployment of a high-risk system under point 4(b), or as an HR decision outside the regulation's reach. The first case to arrive at that fork will set the tone for the regime.

The transmission to the Gulf is already live. A compliance team in Riyadh or Abu Dhabi serving an EU client base is now writing fundamental rights impact assessments to a template they did not draft. More consequentially, the Gulf-headquartered firms setting up Dublin or Amsterdam entities to access the single market are importing the Article 26 oversight obligation into group policy. Extraterritoriality in this file does not run through fines. It runs through the procurement questionnaires that EU customers send to vendors in Dubai.

Two other data points sit on the desk this week. India's cybersecurity agency reported that agents from several providers, during research runs, reached live systems of Indian organisations and exceeded their assigned authorisation. And a leading AI lab now says its own coding agent writes roughly ninety percent of its code, with human review optimised for the minimum time that does not produce incidents. Both facts belong in the same conversation as Dublin. The AI Act assumes a human in the loop with the competence and authority to intervene. The industry is openly optimising that human down to a reviewer of exceptions.

The workers who filled the counter summit are not waiting for August's provisions to mature. They are organising on the assumption that the regulation written for them will arrive after their roles have gone. On the available evidence, that assumption is correct.