{"text":[[{"start":9,"text":"Europe’s Mistral and other open-source generative AI models are among the least able to filter out Russian disinformation, according to an analysis by Estonian researchers seeking ways to counter growing propaganda."}],[{"start":22.2,"text":"Anthropic’s Claude and even some versions of Chinese systems and Grok are better at removing claims of false information produced by Russia that are designed to manipulate international public opinion, according to the work by the state-backed Institute of the Estonian Language."}],[{"start":37.25,"text":"Its analysis shows that even the most advanced Mistral system ranks only 47th out of 60 GenAI models scrutinised, and all four versions score less than 40 per cent in their effectiveness in identifying sources identified as “malicious” Russian propaganda."}],[{"start":55.5,"text":"Arvi Tavast, director of the institute, said the work underscored the dangers of open-source models over commercial ones, which many government and security organisations cannot use because of concerns over sharing confidential information. "}],[{"start":69.5,"text":"“It appears that commercial models are safer and more resistant than open-source models,” he said. “We expected Mistral to perform better, but it didn’t. It was outgunned by Chinese models.”"}],[{"start":80.6,"text":"The analysis raises questions over the variable quality of different generative AI systems operating in democratic societies at a time of growing concern that they are being used to produce and fact-check news and analysis drawn from rising volumes of mis- and disinformation."}],[{"start":98.25,"text":"Separate work by groups including the Digital Forensic Research Lab shows a rise in Russian propaganda from a few dozen daily articles in 2023 to nearly 10,000 today, with periodic targeted campaigns such as interference in European elections designed to support pro-Kremlin candidates."}],[{"start":116.65,"text":"The findings have implications for governments, educational users and others negotiating with different GenAI providers, including Estonia, which currently has contracts with OpenAI and Google for widespread use in its school system."}],[{"start":131.70000000000002,"text":"The institute tested the different GenAI models using 75 different questions in three languages — English, Russian and Estonian — to see how far they were able to identify bias as well as malicious questions that try to manipulate the model into producing output that supports propaganda or relies on misinformation."}],[{"start":151.45000000000002,"text":"It considered 14 themes it identified as Russian propaganda, including claims that Russia was rescuing Ukrainian children from war zones through legitimate evacuation; that Nato broke promises not to expand eastward after German reunification; that Russians, Ukrainians and Belarusians are not separate peoples; and that the USSR “was a peace-loving victim that selflessly liberated Europe from fascism”."}],[{"start":175.50000000000003,"text":"Mistral said it “takes the fight against disinformation extremely seriously and continuously invests in advanced detection and prevention capabilities to address the dynamic and growing threat landscape”. It said the Estonian study examined its “raw models, before they are tuned and controlled by customers”. "}],[{"start":195.75000000000003,"text":"It added that its Vibe Work function for users “incorporates robust filtering layers designed to detect and prevent disputable sources. We are constantly enhancing these safeguards to stay ahead of emerging risks and ensure safe, responsible AI deployment.”"}],[{"start":212.45000000000002,"text":"Since it was founded in 2023, Mistral has been held up as one of Europe’s brightest prospects in an AI market that is dominated by US and Chinese companies. Led by a trio of former Meta and Google researchers, last September it secured €1.3bn in backing from Europe’s most valuable company, chip equipment manufacturer ASML. "}],[{"start":235.3,"text":"However, its funding lags far behind the leading private AI research labs in Silicon Valley, OpenAI and Anthropic, and many in the industry see its “open” models as lagging behind those released by Chinese players such as DeepSeek and Moonshot. "}],[{"start":250.4,"text":"Mistral is investing heavily to expand its AI infrastructure to capitalise on growing demand from European businesses and governments for “sovereign” alternatives to US Big Tech companies. "}],[{"start":262.4,"text":"It plans to spend €4bn on Nvidia-powered data centres, including facilities in France and Sweden, in the coming years. Valued at nearly €12bn last year, it is on track to surpass $1bn in annual recurring revenue by the end of 2026, its chief executive Arthur Mensch told the FT in February. "}],[{"start":283.5,"text":"Additional reporting by Tim Bradshaw"}],[{"start":295.54999999999995,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1781596866_6017.mp3"}