Hallucinations: EU Commission takes Microsoft Bing to task over AI dangers

Microsoft has received a pink slip from Brussels: Based on the Digital Services Act (DSA), the EU Commission is demanding information about specific risks that could arise from functions for generative artificial intelligence (AI) in the Bing search engine. In particular, the Brussels-based government institution is looking at the ChatGPT-based bot Copilot and the image generator Image Creator from Microsoft Designer. The Commission sent a list of questions to the software company on March 14, but it did not respond to the satisfaction of the executive body. It has therefore now set Microsoft a deadline of May 27 to respond with the official request.

Concerns about deepfakes and manipulation

With the legally binding request for information, the Commission is also asking Bing to provide internal documents and data that the operator has not yet disclosed. This step is based on the suspicion that Bing may have misled users due to risks associated with generative AI such as so-called hallucinations, the viral spread of deepfakes and the automated manipulation of services. This would be a clear violation of the DSA.

Over a year ago, the Commission classified Bing as a huge search engine within the meaning of the Platform Act with more than 45 million users in the EU. Relevant operators must carry out risk assessments and minimize identified risks, for example to democracy, public safety, fundamental rights and the protection of minors. Generative AI is one of the risks identified by the government institution in its guidelines on the integrity of electoral processes. It is particularly concerned about the upcoming European elections at the beginning of June.

Copilot has often been caught out for errors in election topics

If Bing does not respond within the deadline, the Commission can impose fines of up to one percent of the provider's total annual income or global turnover, as well as regular fines of up to five percent of the average daily income. A similarly high penalty may be imposed if the operator provides false, incomplete or misleading information. The Commission points out that a request for information is an initial investigative act that does not anticipate possible further steps. A decision will be made because of the assessment of the responses and, if necessary, formal proceedings for DSA infringements will be initiated. Such investigations are already underway against Meta, TikTok and X. The Microsoft subsidiary LinkedIn and AliExpress also have to answer questionnaires.

Clara Helming from the civil society organization AlgorithmWatch welcomes the action against Bing. "Our research shows that Copilot often gives incorrect answers to basic election-related questions. In a joint study with AI Forensics on the elections in Bavaria, Hesse and Switzerland in 2023, we found" that a third of relevant information contained errors. For example, the chatbot had distorted poll results and even invented scandals about politicians. Helming is irritated: "Although we informed Microsoft about the problem, subsequent tests showed no improvement."

The difficulties identified are system-related and "are due to the probabilistic nature of large language models", Salvatore Romano from AI Forensics explains. If chatbots cannot provide accuracy, "they should avoid answering tricky questions". A Microsoft spokesperson asserted: "We have cooperated fully with the European Commission under the voluntary information request and remain committed to responding to their questions and disclosing more about our approach to digital security and DSA compliance."

(bme)