The first draft was so unreliable that AI was said to be the fifth element after water, earth, fire and air. Fortunately, after more than 18 months of work, the European Parliament has reached a position on Artificial Intelligence that the majority defines as "serious and balanced". A document of more than 70 pages that has received the support of the main political groups in Europe and is the first stone for the great challenge of regulating AI.
We have spoken with the three Spanish MEPs who are members of the AIDA commission to understand what is Europe's vision of Artificial Intelligence. Because despite the fact that the approved text is the same, the debate on this technology has many edges, it is still open and it is what will mark the future AI Act. These are the reflections on Artificial Intelligence of Pilar del Castillo (PPE), Ibán García ( S&D) and Susana Solís (Renew).
Why Europe now attaches importance to Artificial Intelligence
"We have been using this technology for many decades. But what is new is its massive implementation in all processes. If we do not regulate it, many issues will generate a series of risks and, in some cases, unfortunately, consequences. Taking into account the power intrinsic that this technology has, it is time for Europe to go together", replies Ibán García, Socialist MEP and member of the Committee on Legal Affairs.
"Anything that allows AI to be applied under guaranteed conditions and does not affect fundamental rights interests us all," says Pilar del Castillo, popular MEP co-president of the Parliamentary Intergroup on Artificial Intelligence and former Minister of Education, Culture and Sport.
For Susana Solís, liberal MEP, engineer and member of the European Industry, SMEs, AI and Digital Policy intergroups: "this report is a roadmap; a wake-up call. A warning that Europe is falling behind but that it is still We are on time". Despite starting from an unfavorable position, the different MEPs are optimistic and defend that "more has been done in two years than was done in the last decade". And as an example, Solís gives the RGPD: "The Data Protection Regulation took 7 years. Now we have much more experience and everything is being done faster."
Europe wants to repeat with AI the success of the Data Protection Regulation. But they know they don't have that much time (that one took from 2012 to 2018).
The response of MEPs to whether a specific law is really necessary is unanimous, despite the fact that they remember that there are still fringes to define it. "Parliament now has a comprehensive report from different angles where opportunities and risks are addressed. The regulatory framework, which will be the AI Act, is what is being discussed now," explains Pilar. From the Commission they made a law proposal, now the European Parliament (elected by the citizens) has voted what should be the main lines and the red lines.
The strategy to rival the US and China
Currently, the public-private investment of the European Union in AI is around 12-14,000 million euros per year. "They believe that between 20 and 25,000 million euros are necessary to be champions in AI here. It is a figure that would keep us in the race, but the distance is great," says Susana. "To talk face to face with the US and China we need to have a Digital Single Market and right now we don't have it."
"AI feeds on data, but then there are other issues. Supercomputers and cloud services, of which there are still none in Europe. There are projects like Gaia-X, but there we are clearly delayed. Europe lacks that infrastructure and this is what is being tackled for the first time with Horizon Europe and the Digital Europe program, with a budget of 7,000 million euros, among other things to be able to obtain these supercomputers and this necessary infrastructure to investigate in AI", comments Pilar.
"I believe that important steps have been taken from the institutions, because until recently there was nothing defined. In the countries there is a growing awareness, but I believe that greater and highly focused efforts must be made. In infrastructure there is a clear deficit , but if we compare it with 5 years ago, we would not have had this discussion," Pilar summarizes about how much the landscape has changed.
A few years ago the situation was completely different, with an investment that did not even reach half. In these years, Europe has understood that it must take an important turn, but they are also quite aware. "Even so, we won't be at the level of China," warns Ibán.
The reality is harsh. Only 8 of the 100 largest companies in AI are European. But the situation in the private sphere is so worrying that none of the three MEPs could tell me on the fly a leading European company in artificial intelligence.
"Europe has some differentiating elements that may help. First our network of universities and then what is known as the 'Brussels Effect', which we already saw with the GDPR. We anticipate that the market will end up looking quite similar to ours, with our obligations from the point of view of from an ethical point of view," continues Ibán. "As much as some threaten, they are not going to leave the European Union. We are a market of 500 million people. Leaving it for not adapting to minimum standards is not worth it."
The intention is to move towards a basic standardization of AI worldwide, something in line with what is advocated by the OECD and UNESCO. "Your proposal goes in a similar direction," says Ibán. "It breathes a lot of our culture, which can be the basis of an international AI convention. If the US wants to apply a certain technology, it must respect a series of rules such as explaining where the data has come from or how AI works. We hope that in the end even China explains to us how its algorithms work and its companies that want to be in Europe adapt to the rules".
Ibán explains some of the contacts between Europe and the rest of the powers. With Think Tanks in Washington and even occasionally with China, interested in seeing what their companies would need to adapt to Europe. In fact, Ibán points out that "at the moment in the US they have nothing and China has made regulations. But the idea is not to compete against these countries, but to compete together."
The investment will be difficult to match, but the intention is to try to lead the bases of an international AI convention. One that even US and Chinese companies have to meet. Europe hopes to be a large enough market to force them to adapt their method of working.
To attract talent, establishing clear legislation is considered a positive point. "We have done it with Data Protection. If we do it well, and this is that it does not stop innovation, I believe that Europe can set the standards in AI. In the US at the moment they are not doing anything beyond allowing it", explains Susana .
"Europe has had quite an impact. In the US they start from a different political culture. Systems that have been built in a different way," explains Pilar. "We have to find a convergence. Common spaces. The United States is not so guaranteeing, but there is no doubt that we have to look for connection bridges because we share many values. It is a competitor, but also someone with whom we have many possibilities of a beneficial relationship. ".
Being agile vs protecting rights: how this balance has been achieved
Regulating AI is challenging by its very definition. "We are talking about a technology capable of self-modifying and gradually increasing its power and efficiency by itself," explains Ibán. Still, MEPs are happy to have struck a balance. A text that, on the one hand, encourages the uses of AI and, on the other, highlights the risks of certain uses.
"The text is truly balanced. There is still a long way to go and it will be resolved in the AI Act. There are many issues where there is still a distance, but we have had more than 1,000 amendments and we have reached a compromise among all. An enormous reform Nothing to do with the initial draft where the five elements of the earth were discussed, "explains Ibán ironically.
The criticism from the socialist side is that the initial draft presented by Axel Voss did not give enough importance to the risks and the impact on fundamental rights that AI can have. This is how he explains it: "it was very focused on geopolitical issues and a conflict with Data Protection was even inferred."
"The majority position has also been reflected in the voting, with 495 votes in favor, 102 abstentions and only 34 votes against," summarizes Pilar. "There have been conversations with many sectors. Agriculture, transport, etc. I believe that all of them are sectors in which AI has a role with clear benefits. Let's think about health with remote treatment or agriculture, which changes completely with the possibility to have better and more sustainable products", reflects Pilar on the positive effects.
The final position of the Parliament does not include any prohibition and states that it must be an agile legislation, but it does record the risks in sectors such as health or the world of work.
"The risk approach is what allows us to deal with what concerns us in practice. It is the same approach that the European Parliament adopted on Ethics in AI," says Ibán. The report reviews 6 cases where AI can play a delicate role: health, energy efficiency, foreign policy, competitiveness, the labor market and the future of democracy.
Lilian Edwards, Professor of Law and Innovation at Newcastle University.
There are dozens of examples: from algorithms that discriminate by sex or race when selecting candidates for interviews to algorithms that could affect patients. But also more extreme cases such as algorithms that force you to smile at work or algorithms that offer the best parking spaces to those who have a better car.
The proposal of the European Commission is to establish four levels of risk, with associated obligations depending on each level. In the case of Parliament's report, some ideas have been reflected while other demands have been left without sufficient support.
Human intervention has a prominent role, but there is debate about how it will be implemented. "It is still not considered how these audits will be carried out. From the AI Act, one could reach the conclusion that it privatizes the system. A body, an agency and a coordinating body at European level have been proposed, but for me it is insufficient." Along similar lines, Susana explains that "it is not clear how the European supervisor will work so that there are not 27 different interpretations for each country."
The categorization into risk levels is quite accepted, but there are doubts about the functioning of the European Supervisor and the level of requirements necessary for AI systems with lower risk.
"The balance is the most difficult when there are so many committees legislating," says Susana. "One of the points where this dilemma will be seen will be in the definition of Artificial Intelligence." Another point is when defining what are considered high-risk systems. "80% of AIs are not high risk. That's where we don't have to put too many hurdles." And Susana explains to us a problem related to one of the commissions to which she belongs. "The issue of certifying green algorithms. I would love to, but we don't have tools as such. We can require responsible energy use statements or that the most efficient algorithm be used, but we are not ready to put a green algorithm label. No We have to legislate beyond what is necessary.
Who will decide at what level each AI usage goes remains to be seen. "It will be defined by a series of sectors, but it will have to be seen in the final version of the AI Act. It is expected to be a list that is updated every so often. Then it will be necessary to see that there are no duplications. For example, in health Some instruments are already regulated by the Law on Medical Devices. In those of high risk, it will be necessary to avoid creating duplications so as not to conflict with other laws".
"There are many issues to contemplate," continues Susana. "It is not the same for Finland as it is for Spain, with its autonomous communities. In a decentralized country like ours, how are you going to create a central body that handles all that amount of data? What incentives do we give hospitals to feed the database? In a centralized country it is easier to audit".
"Another matter is that, for example, the AI Act does not yet contemplate the distinction between user and end-user, for example between doctor and patient," says Susana, in relation to the fact that the risk of AI is not the same for some and for others. others. The hospital, in this example, would be responsible, but the description of the AI system would take a different approach for each.
The great debate: what to do with biometric identification
It was the tightest amendment in the vote, although in the end it did not go ahead. The European Parliament finally opted for the non-prohibition of biometric identification systems, but there is a very powerful current that advocates it. Ibán acknowledges that within his group there are different opinions and this was reflected in the vote. In his case, he expresses that "I am not in favor of the per se prohibition of any technology", a position also shared by Pilar and Susana.
"I'm not in favor of banning any technology per se."
"Facial recognition has many uses. The Commission already classifies some as unacceptable. For example, AI cannot be applied for real-time biometric recognition in public areas. Ban is already contemplated there," explains Pilar. "We are still in a very early phase and the exceptions in the regulation are being negotiated. But there is already talk that it is not allowed except in cases where there is judicial authorization and it is used for reasons of imminent danger."
Renew's position is equivalent. "The systems of 'Social Scoring' or massive surveillance in real time should be prohibited in public places, except in certain cases where there is a judicial order limited in time or to pursue a terrorist attack," Susana explains. "In the final version of the AI Act I think there are going to be very few exceptions where it is authorized."
This issue is perhaps the clearest to visualize the debate between MEPs more inclined to regulate and those who advocate establishing a direct ban. Societies in defense of digital rights such as EDRi advocated a direct ban, but it has been decided not to include it.
"There are those who speak of a mere executive order when it is an emergency situation and there are those who advocate that it not be used in any circumstance," Ibán responds, describing that the approach of exceptions before the general prohibition allows better defining under what conditions will be used and thus "eliminate any other possibilities".
For the groups advocating a ban, the European Parliament has not been brave enough. Similarly, the text does not contain a direct prohibition of lethal autonomous systems with AI. Yes, there is a recommendation for its regulation, but the door is left open for the AI Act not to regulate the military use of AI.
Pilar summarizes in a simple way why the European Parliament has not taken a position on autonomous weapons with AI: "there would be no consensus".
"AI in the military sector is already being used, the thing is to see where and how it is being used," explains Susana. Regarding the military sector, the Liberal MEP points out that "many sectors wanted to be excluded. In the case of the military, I believe that it should be left out, in terms of its characteristics due to the defense and sovereignty of the EU. In certain sectors we have to put Our European values are in front of us. We have seen it with energy. Here it is the same. Our great strength is our values, our great advantage is to do it differently. Can we stay behind in the arms race? It is possible, but above all we have to do it with our values From there we can build a strong industry."
Will even small businesses need to explain how their algorithms work?
The debate revolves around the traceability and explainability of the algorithms, but "later in practice it is much more difficult to carry out," admits Ibán. "A developer will tell you that even they don't know how it got there." Faced with the technical challenge for a small company to explain how all the algorithms they use work, the socialist MEP explains that "we are not going to demand what cannot be done", but he does point out that "we cannot lower the demands one gram, because we are talking about compliance with fundamental rights".
The solution they propose is audits for institutions, fines for those who do not meet the requirements based on income and a powerful Artificial Intelligence Agency, with the capacity to provide assistance to small and medium-sized companies. "Using the excuse that more is spent on a new institution, some argue that it is not necessary. But if an algorithm can generate discrimination and affects me as a citizen, I do not care if Google does it or whoever. Protection must be the same," says Ibán.
Pilar has a different position and advocates for certain exemptions: "You have to pay attention to small companies, with different requirements. Make sure they don't have that additional burden of compliance."
"We have to do it so that you don't create too much bureaucracy for small ones," says Susana, who describes that for large companies, the explainability of algorithms is not such a big challenge, since many of them have fairly exhaustive control. "It does not mean that the exact code is going to be requested, but rather to explain how the algorithm works and what data it has been fed with."
One of the examples is the algorithm of applications such as food delivery at home. These algorithms can pose discrimination or a problem for riders. With the future legislation, these workers must be able to know what factors are being taken into account, be it for example the delivery time, their availability, the weather with which they worked or the feedback of the users.
In this sense, Spain is one of the pioneering countries in demanding more transparency with algorithms in cases like this, since the "Rider Law" already includes this precept. The AI Act would entail the expansion of this transparency to more sectors and better regulations. Susana also recalls that beyond the AI Strategy, Spain has presented itself as a pilot country to see how these algorithms could be certified.
Equivalent to Data Protection, small businesses will be challenged to provide explanations of how their algorithms work. It is not yet clear what level of demand they will have or what facilities Europe will offer them to achieve it.
In the European Parliament, the meetings of politicians with the different lobbies are quite controlled. They recognize that there have been many contacts because practically all sectors feel affected. "The legislation that is going to affect them the most is the DSA/DMA, but there is also interest in this legislation on AI," says Susana.
"There are those who believe that we fell short and others that anything goes. In the case of the Big Tech there is everything, there are even those who proposed a very strict regimen, but everything has a trick, because when the big ones propose a strict regimen what they propose In practice, it is legislation that prevents the emergence of new competitors, since they could not carry out all the regulations that could hypothetically be put on AI," explains Ibán.
Copyright for the AI has been set aside
If we talk about Artificial Intelligence, one of the latest projects that has surprised us the most is DALL-E 2, with its amazing creations. Asking the various MEPs about Intellectual Property and copyright, they tell us that this issue has not been dealt with at the moment in the European Parliament.
"There are debates with rethinking copyright, but it is not solved. There are doubts about to what degree AI intervention would be taken into account, what degree of use is permissible and on whom the rights would fall; if the programmer, the users or the company…", reflects Ibán.
"It is an open field. For the time being, the framework of the European Union Directive on copyright is still in force," says Pilar.
"Parliament's report is a roadmap. There is still nothing, but we should review it. AI is going to open many fronts," describes Susana, who gives us NFTs as an example and how within the so-called metaverses they can be sold and /or replicate digital objects of recognized brands.
At the moment, Europe has not opened the copyright and AI melon, although from the European Union Intellectual Property Office (EUIPO) it is already one of the most debated topics.
As of 2023, citizens will have to know if an AI is in charge
The mental framework is to have a regulation for AI that lasts until at least 2030, despite everyone recognizing that this technology evolves at breakneck speed and it will be very difficult to think what the associated needs will be.
Once the European Parliament has established what the main lines should be (no prohibition, risks in sectors such as health or work, promotion of investment in AI…), it is time for the different commissions to continue working to define the AI Act , which will be the legal text. Looking ahead to autumn, Parliament should vote on this regulation again, with an eye toward the beginning of 2023 for final approval.
As with the DSA and DMA, the AI Act is intended to be a Regulation. That is, it will be directly applicable after a fixed period (usually between one and two years) and therefore no transposition will be necessary in each country. With this, it is also sought that there is not too much casuistry in each country that could generate additional conflicts.
"Citizens must understand that AI is something of the now," says Susana. "We must know that when we go to a job interview or request an appointment for an operation, the first selection may be an algorithm. As a citizen I want to know if the one who made a decision has been an algorithm and if it has had negative consequences."