The bloc has failed to develop general purpose artificial intelligence systems that are increasingly the backbone for products like chatbots or automated emails. Without them, the EU can’t enforce its vision of ethical AI, a new report warns
EU companies are becoming increasingly reliant on US and Chinese companies like Meta, Google, Microsoft and Baidu, to access general purpose AI systems that underpin other AI tools, a dependency that could stop Brussels setting global standards for the technology.
No EU countries have yet managed to develop a general purpose AI system, and will struggle to do so, due to the sheer amount of money, data and computation they require to build, according to a new report.
“Instead, European companies will likely rely on systems that are developed elsewhere,” warns Emerging Non-European Monopolies in the Global AI Market, a report by the Future of Life Institute, which tries to steer the development of new technologies.
“Given that the utility of these AI systems largely depends on how many resources are invested and that the biggest players are in the best position to attract foreign capital, these dependencies are likely to grow over time,” the report says.
General purpose AI systems, also called foundation models, provide the basis for applications such as AI chatbots, automatic translation and advertisement generation. They can generate not only text answers to prompts, but increasingly create video, images and audio too.
EU companies are already using one of the most prominent systems – the San Francisco based GPT-3 – to build their own products, such as a service that turns a sequence of words into an email, according to the report.
If the EU doesn’t host any of its own general purpose AI systems, it may struggle to regulate AI as a whole through its proposed AI Act, said Risto Uuk, a policy researcher at the institute and author of the report.
Brussels hopes to set global AI ethical standards through the act, billed as the first such major piece of legislation in the world. It stands in contrast to a relatively laissez faire, voluntary approach that dominates in the US.
But there have been multiple warnings that the act fails to cover general purpose AI systems, instead leaving liability for problems like bias and toxicity on the shoulders of EU companies that develop specific applications on top.
Then there’s the economic cost of relying on US and Chinese companies, Uuk noted. GTP-3 charges per word generated, leaving EU companies that build on top of the model with a potentially sizable bill that could eat into profits.
It’s not just the tech giants of the US and China that have created general purpose AI models. Last year saw the release of HyperCLOVA, billed as the South Korean answer to GTP-3. Shortly afterwards, an Israeli company, AI21 Labs, released its system, Jurassic-1, while Google-owned DeepMind, based in the UK, released a system this year called Gopher.
The report found just two EU-based projects attempting to build similar AI systems: the German start-up Aleph Alpha, and the Large European AI Models project, backed by a number of German firms and research institutions.
“Small European companies are unlikely to catch up with their global competitors anytime soon,” the report says.
This is in part because of the vast cost of training a general purpose AI model. Uuk’s report puts the price tag at between $9 – $35 million.
And this cost could keep on rising, said Uuk. He pointed to estimates from the US chipmaker Nvidia that in five years’ time, companies could invest up to a billion dollars to train a single system, because the rewards for having one on the leading edge are so lucrative.