The decision by the EU to regulate ChatGPT by classifying it as a search engine will seem logical to those who use it as such. But what if you use it as a friend? Or as a therapist? Or to discuss the upcoming elections?
These are some of the many new questions raised by the advent of generative AI for European regulators — who, according to experts, have not yet found satisfactory answers.
On Monday, the European executive classified ChatGPT among the 'very large online search engines' under the Digital Services Act (DSA), thereby imposing a series of transparency and risk mitigation obligations similar to those applied to traditional sector leaders such as Google and Bing (from Microsoft), which have more than 45 million monthly users. ChatGPT's parent company, OpenAI, faces fines of up to 6% of its global annual turnover if it fails to comply with these rules.
However, this classification is restrictive and applies only to the parts of the tool that function as a search engine, meaning that conversations in which the chatbot contributes its own input do not appear to be covered. Under the DSA, the Commission could have chosen to classify ChatGPT in the category of 'very large online platforms,' which generally applies to social networks and e-commerce platforms and entails different obligations.
Yet neither of these categories reflects the full range of possible uses of chatbots.
'ChatGPT is much more than a simple search engine, and there are also risks related to the chatbot itself that escape the strictest obligations [of the regulation],' pointed out Christel Schaldemose, a Danish socialist MEP, who was one of the main negotiators of this legislation.
She highlighted risks to children, 'such as emotional dependency and manipulative or addictive design,' and urged the Commission to clarify to what extent these risks are taken into account by the current regulation.
ChatGPT and its competitors are at the center of a global controversy, as they have been linked to the tragic suicides of teenagers, notably that of Adam Raine, a 16-year-old Californian, whose parents have launched legal action against the company.
Use cases of AI-based chatbots for therapy or companionship are not marginal: according to a recent study by the insurance sector, as many as 60% of adults worldwide use them for therapeutic purposes.
A hybrid space
Current legislation on digital services, adopted in 2022, did not anticipate the rise of chatbots. João Pedro Quintais, a law lecturer at the University of Amsterdam, describes ChatGPT as a 'hybrid' technology that combines the functions of a search engine and an online platform, as well as other functions closer to those of a publisher of its own content.
The classification as a 'search engine' could limit the Commission's ability to control how ChatGPT manages risks related not only to adolescent mental health, but also to election integrity or illicit content.
Asking ChatGPT to name the candidates in a local election should now fall within the scope of the DSA. By contrast, a conversation between a user and a chatbot about which candidate to vote for, during which false information could be spread, might not be covered.
It took the Commission slightly less than a year to finalize the classification of this chatbot, its teams having struggled to determine under which category to place it. According to experts, without having consulted the full text enacting this classification announced this week, it is difficult to know exactly what OpenAI's obligations will be.
The classification as a 'platform' would have imposed additional content moderation obligations on OpenAI under the DSA, but it would also have allowed the company to escape a significant portion of its responsibilities. The DSA, like other legislation worldwide on platforms, is built around the principle of 'safe harbor,' whereby companies are not responsible for content present on their platforms, as that content is uploaded by users. But the advent of chatbots raises the question of whether a conversation between a person and a machine can be considered 'user-generated content'.
Models and risks
While regulatory attention to ChatGPT as a consumer application is recent, the Commission is closely monitoring the artificial intelligence models that underlie it via its AI Act.
OpenAI, the parent company of ChatGPT, as well as companies such as Anthropic and Gemini (from Google) are developing what are known as general-purpose AI models, capable of performing a wide variety of tasks that may present 'potential systemic risks'.
Since last August, under the AI Act, these companies have been obliged to 'assess and mitigate' the risks presented by these models. The Commission began enforcing this legislation in late August, when it grilled a series of AI-specialized companies on their safety procedures.
Previously, a group of experts had identified four risks: the development of nuclear or biological weapons, loss of control over the models, AI models out of control that begin to engage in hacking, or models engaging in large-scale manipulation.
But as frightening as they are, these systemic risks do not take into account a whole range of potential problems, both for users and for society as a whole.
'[These guidelines] are more focused on what are called existential risks than on risks to fundamental rights,' raised Daniel Leufer, policy lead on emerging technologies at AccessNow.
He argued that with the classification of ChatGPT under the DSA, there is a chance 'to get into the details of design choices and to consider ChatGPT more for what it really is, namely a product'.
