Chinese artificial intelligence firm DeepSeek has agreed to launch the first country-specific version of its namesake chatbot, while tackling the problem of “hallucinations” in its AI models to meet the requirements of regulators in Italy.
DeepSeek assented to the package of commitments after a months-long negotiation with the Italian Competition Authority, known as the AGCM, which had earlier accused the Hangzhou-based start-up of not sufficiently warning users in the country about the risk of its chatbot hallucinating – producing incorrect or misleading information.
The AGCM announced on Monday that its investigation into DeepSeek had ended after the company committed to “making its disclosures about the risk of hallucinations more transparent, intelligible and immediate”.
In particular, the regulator praised DeepSeek for committing to lower the hallucination rate of its AI models through technical fixes, calling it “commendable”.
“[DeepSeek] has stated that the phenomenon of AI model hallucinations is a global challenge that cannot be entirely eliminated,” the AGCM said in a notice.
While hallucinations occur in every generative AI service, DeepSeek’s effort to comply with Italian authorities – which are some of the most active in Europe to enforce AI regulations – would appear to augur well for the company’s future business expansion.
Researchers at ChatGPT operator OpenAI, for example, have blamed existing model training practices for incentivising the large language models that power these chatbots to make guesses at answers rather than acknowledge uncertainty.
In recent months, Italian regulators have penalised US tech giants, including Facebook parent Meta Platforms and Google, for anticompetitive and data privacy violations.
The AGCM, whose remit includes consumer protection, launched its investigation into DeepSeek’s hallucinations last June.
According to a notice, DeepSeek had engaged in several rounds of interaction with the AGCM. These included submitting proposed remedial actions on September 15, September 22 and November 21.
Those measures included showing warnings about hallucinations and information about terms and conditions in Italian, when a user’s internet protocol address is in Italy or when prompted in Italian. Such information is currently only available in English and Chinese.
A check by the Post on Wednesday showed that DeepSeek’s website, when the language is set to Italian, already features a longer hallucination warning below the dialogue window than the one currently on the English website.
The company also proposed to organise workshops for staff members “to ensure full awareness within its relevant business units” of Italian consumer law.
“The workshops will certainly increase ... the necessary knowledge within DeepSeek regarding [Italian] consumer law and its practical implementation,” wrote Diego Rigatti, Amedeo Della Croce and Carlotta Mosca, the intellectual property and competition lawyers from London-headquartered Fieldfisher, which represented the Chinese company.
DeepSeek pledged to submit a report to the AGCM within 120 days about the progress of its commitments, as the regulator warned that the investigation may be reopened and a fine of up to €10mil (US$11.7mil) may be imposed should the required changes not be made within that period.
“The changes to the user interface and terms and conditions are relatively easy to implement, while the technical improvements are harder to guarantee,” said Fang Liang, the head of safety and governance in China at Beijing-based AI safety consultancy Concordia AI.
DeepSeek’s chatbot was pulled from Italian app stores in January last year, just days after the company’s release of its R1 model, when the country’s personal data watchdog, known as the Garante, raised concerns about the firm’s data handling practices.
An unresolved question remains on whether DeepSeek’s services fall under the scope of the European Union’s major legislation, the Digital Services Act (DSA), which would subject it to even greater scrutiny from Brussels.
In its notice, the AGCM expressed its doubts after another Italian regulatory body, the communications watchdog Agcom, shared its assessment that DeepSeek’s services constituted a search engine that was subject to the DSA.
“The qualifications of DeepSeek as a ‘search engine’ ... under the DSA require further clarification,” the AGCM said. – South China Morning Post
