DeepSeek tunes its AI for Italy after hallucination probe

Chinese artificial intelligence company DeepSeek has agreed to launch its first exclusively national version of its chatbot, specifically adapted to Italian national needs, provided it can meet regulatory stipulations.

The announcement arrives as the Italian watchdog, AGCM, is becoming more rigorous in regulating artificial intelligence (AI) in an attempt to eliminate hallucinations.

Italy is one of the most stringent countries in the EU with AI regulation, as exemplified by its regular investigations into companies like Meta, often resulting in fines, as well as clamping down on Google for sports streaming piracy.

More subtle problems arise, however, such as the interpretation of the term “search engine.” That used to encompass a singular platform such as Google or Yahoo, but now AI chatbots can scrape the web from a multitude of angles, often cobbling together their own fabricated narratives.

The AGCM acknowledged that: “[DeepSeek] has stated that the phenomenon of AI model hallucinations is a global challenge that cannot be entirely eliminated.”

In turn, Deepseek has pledged to reduce hallucinations, which has been praised as “commendable” by the regulations. It remains unclear how this will be measured, though the company has organized a program of workshops to get its staff up to speed with Italian law.

Deepseek is to submit a report to the AGCM to set their commitments in stone. Noncompliance could result in a €10 million ($11.7 million) fine.

When dissecting the various segments of their new obligations, Concordia AI spokesperson Fang Liang observed:

The changes to the user interface and terms and conditions are relatively easy to implement, while the technical improvements are harder to guarantee.

Hallucinations are common

Hallucinations affect all generative AI, not just DeepSeek, and researchers at industry leaders like OpenAI lament current training practices that encourage guesswork, as opposed to conceding uncertainty.

The AGCM has swiftly asserted that Deepseek needs to be “making its disclosures about the risk of hallucinations more transparent, intelligible and immediate.”

The commitments also mark a potential path back into Italy after DeepSeek’s chatbot was removed from app stores in January last year, following data-handling concerns raised by the watchdog. A return will depend on whether regulators are satisfied with its transparency measures, and on whether the service is ultimately classified as a search engine under the EU’s Digital Services Act.

Unlock more exclusive Cybernews content on YouTube.