EU lawmakers 'hold breath' on eve of AI vote

"I'm holding my breath until the last second," said Romanian liberal MEP Dragoș Tudorache, earlier this week.

Tudorache, along with Italian socialist lawmaker Brando Benifei, are the parliament's lead negotiators on the Act.

First proposed by the European Commission in 2021, the bill comes at a time when general purpose AI like Chat GPT have exploded onto the market.

The Council, representing member states, reached its negotiating position last December.

Now the parliament's committee is set to vote on a draft put forward by Tudorache and Benifei before it goes to the plenary in June.

"This is the kind of negotiation where no one can leave the room happy or unhappy," said Tudorache. "The key is to retain control and human oversight and trustworthiness," he said.

Tudorache described it as the "magic touch", whereby the future powered by AI is one where humans are still in control.

"If we don't get that control right, then you might have this dystopian view that are being expressed by many," he said.

The new rulebook includes placing an outright ban on certain practices, while listing others as high-risk.

Unlike the parliament, the council has stopped short on slapping an outright ban on predictive policing, emotion recognition, and facial recognition in public spaces as part of a security-centric mix.

Not everyone is keen on some of these bans, including within the parliament's Renew Europe, the liberal faction where Tudorache is seated.

With some exceptions, the centre-right EPP is not happy with the bans either, while the Greens and socialists are in favour. Those tricky questions and positions are likely to play out at a committee level vote on 11 May.

So far, civil society has broadly welcomed the parliament's draft given its beefed up rights safeguards.

"The parliament has really listened to what civil society has asked, and has made some quite big changes," said Daniel Leufer, a senior policy analyst at Access Now, an NGO that defends digital rights.

This includes transparency obligations and fundamental rights impact assessments for users, he said.

Such transparency is needed to prevent scandals from erupting. Recently, an investigation by the Amsterdam-based LightHouse Reports, revealed that Spain had been using secretive AI to crack down on fraud and reduce public spending on its sick leave benefits.

But at the same time, Leufer worries about loopholes that could derail the regulation when it comes to classifying high risks.

"It's very important that the high risk classification process doesn't contain any loopholes that would allow unscrupulous providers to exempt themselves without consequences and without public oversight," he said.

Despite that possibility, the parliament's extended list of bans is also being praised.

Sarah Chander, an AI expert at the Brussels-based European Digital Rights (EDRi), said it would send a clear message to governments and AI developers that "some uses of AI are just too harmful to be allowed."

But this still falls short when it comes to protecting migrants from AI harm, including where AI is used to facilitate pushbacks, she says.

The proposed ban on predictive policing needs improvement to cover location-based predictive policing, essentially a form of automated racial profiling, she said.

Amnesty International made similar comments in a letter sent to MEPs earlier this month.

"The draft AI Act does not guarantee people on the move the same level of protections against AI-induced harm that are afforded to European citizens," stated the letter.