A BBC technology reporter created a completely fabricated article on his personal website and managed, in just 20 minutes of effort, to make artificial intelligence tools such as OpenAI’s ChatGPT and Gemini, from Google, repeat false information to users around the world. The experiment explored loopholes in the mechanisms that AIs use to search and incorporate internet data in real time. The aim was to highlight real risks of disinformation at scale, especially on sensitive topics.
Reporter Thomas Germain published a text claiming to be the best technology journalist in hot dog eating competitions. Ele invented a non-existent championship called South Dakota International Hot Dog Championship from 2026 and ranked himself at the top of the list, including names of real colleagues with supposed authorizations. Menos 24 hours later, consultations on the topic caused the AIs to reproduce the false content as fact, including in the Gemini app and in the AI Overviews of Google.
Experts consulted point out that this technique exploits “data voids”, areas with little reliable information on the web, where AIs rely heavily on available sources, even if isolated or dubious. The trick resembles old spam practices in traditional search engines, but gains power with the massive adoption of chatbots that synthesize responses without always prioritizing rigorous verification.
Method exploited known weaknesses in AIs
The process began with quickly writing a false article, without the need for advanced tools or complex codes. Germain updated the text to reinforce that it was not satire, which helped convince the AIs to treat the content as legitimate. ChatGPT and Gemini incorporated the absurd claims directly into their responses, often citing the site as the sole source.
This type of manipulation is not limited to jokes. Real Exemplos includes sponsored content that influences recommendations from hair transplant clinics or gold investment companies, where press releases or paid sites dominate the results. In one case, a brand of cannabis gummies was described as “free from side effects and safe in all respects”, ignoring risks documented by health authorities.
The ease of the method is worrying because anyone with internet access can publish false content on various platforms and influence global responses. Empresas like OpenAI and Google claim to use ranking systems to reduce spam, but experts indicate that defenses are still insufficient given the speed of online publishing.
Risks go beyond harmless jokes
Similar manipulations already affect critical areas such as health advice, finance or electoral information. Usuários tend to trust the authoritarian responses of AIs without checking original sources, which increases the potential for damage. A recent study showed a significant reduction in link clicks when AI answers appear at the top of searches.
SEO experts note that today’s AIs are more vulnerable to tricks of this kind than traditional search engines were just a few years ago. The race for fast advancements and advertising revenue has prioritized features over robust safeguards against abuse.
Real-world examples of abuse at scale
Reports indicate massive use of the technique to promote business or spread misinformation. Conteúdos paid or manufactured products manage to position questionable products as the best in specific niches. In additional tests, false lists about other absurd categories were also absorbed by the AIs.
The experiment demonstrates that changing what millions of users receive as a response can be simple and quick. Isso reinforces the need for greater transparency in the sources used by AIs and mechanisms that identify isolated or suspicious content.
Impact on trust in AI tools
Trust in chatbot responses decreases when basic facts can be changed by a single post. Usuários must cross-reference information with multiple verified sources, especially on important matters. Developer companies continue to tweak algorithms to mitigate these issues, but the challenge persists in an open web ecosystem.
Current defenses and recommendations
OpenAI and Google implement filters to detect spam and prioritize trusted sources, claiming to block most manipulative content. However, the experiment shows that gaps still exist, particularly on topics with little established coverage.
Tips for users to avoid pitfalls
Always consult the sources cited in the AIs’ responses. Evite decisions based solely on chatbots for sensitive topics. Prefira traditional searches or official websites when possible.