The European Union has announced that the text of its long-awaited AI Act has been agreed, following a marathon 36-hour debate between the European Commission, European Council and European Parliament. The law is intended to regulate artificial intelligence in a similar way to finance and healthcare, by protecting EU citizens and ensuring that countries, companies and military contractors act responsibly. But will it work?
Part of the problem facing the EU has been that AI is increasingly a moving target. The law has been under discussion since April 2021, long before the launch of ChatGPT, Google Bard and the clutch of powerful text and image-generating models that have sprung up in the past two years. In fact, the president of the European Commission, Ursula von der Leyen, promised to urgently bring in AI legislation when she was elected in 2019.
AI was, and is, developing so fast that regulation is playing catch-up. And member states had been debating where the balance should lay between fostering innovation in the industry 鈥� which is likely to be highly lucrative 鈥� and protecting citizens.
Advertisement
at digital rights campaign group Access Now points out that all that has been agreed is a high-level outline of what will be in the new legislation; the technical details and exact wording will be hashed out in a series of meetings with experts in coming months.
The current draft includes requirements for transparency in AI used in 鈥渉igh-risk鈥� applications, like job recruitment and education, where bias can have huge impacts on people鈥檚 lives. It also bans using AI to recognise emotions of people at work or in education, and outlaws firms scraping the internet for images and building massive face-recognition databases. But a ban on real-time face recognition didn鈥檛 make it into the text as originally expected.
However, Leufer says that the exact wording will determine the outcome. For instance, there is a ban on using AI in predictive policing 鈥� trying to forecast who will commit crimes, when and where. Yet this ban currently only applies when AI is doing the whole task, not if there is a human involved in any part of the process, which Leufer says will in reality always be the case. There are also exemptions for national security and military use of AI.
鈥淭here is this desire from some parts of the negotiating ensemble to agree on things that look good on paper, give them a little soundbite, but to put in all the little loopholes and exceptions in the definitions,鈥� says Leufer.
He warns that companies have also got off relatively lightly in terms of safety testing and with just 鈥渇airly basic transparency requirements鈥� for models. And while even that degree of transparency may be unpalatable for such a secretive industry, it will do little to delay them rolling out commercial models.
We can expect further tweaks before the Act becomes law. While it is edging towards a final draft, it will still need to go to the European Parliament for a vote early next year. If it passes that, then it could come into effect in 2025, but delays are also possible.
While there are no official, immediate impacts for companies, there are already signs that technology giants are treading carefully 鈥� or at least giving the appearance of doing so, either to curry favour or flex their muscles as part of lobbying efforts.
Google launched its new Gemini AI model last week in the US and more than 170 other countries, but notably excluded the UK and the European Union. A Google spokesperson said that the company was 鈥渨orking with local policies and regulators to make sure that we鈥檙e abiding by local laws and other such things before we launch in other areas鈥�. It is worth noting that the original launch of its Bard chatbot earlier this year was also delayed in Europe, but it was made available there relatively quickly after.
However, the real impact on companies will only be apparent years from now, when we see if regulators have the budgets and expertise to investigate, judge and enforce the new rules. Companies will be liable for , but this will be a last resort.聽As Kris Shrishak at the Irish Council for Civil Liberties : 鈥淲ithout strong enforcement, this deal will have no meaning.鈥�
