A major flaw in current artificial intelligence (AI) models has
been found based on research released this week. The misrepresentation has become a critical challenge for corporate legal teams and digital marketers.
AI models have been found to
misattribute, give the attribution for a brand to a competitor, or not attribute at all — when replying to user queries.
“If you ask Anthropic Sonnet 4.6 a question about a
brand’s products, it might give you information about a completely different brand and misattribute the trademark,” Marty Weintraub, Aimclear founder, told MediaPost Friday.
Companies
need to protect their brands and monitor what’s being returned from prompts in answers,” Weintraub added.
Aimclear wanted to determine how five well-known AI engines could
misrepresent individual brands in such “a horrific way.”
advertisement
advertisement
The study looked at where information about one unidentified brand’s information appeared online and how AI tools used that
information to answer questions about it. It also found that some weak points could allow people to manipulate or mislead AI results, but stopped short of explaining harmful methods, because they
could be abused.
The technology crawled roughly 55,000 pages spanning one brand’s corporate website, editorial blog, and technical library, then scanned and pulled into the analysis the
brand’s network of authorized agents and independently registered reseller domains.
Anthropic Claude Sonnet 4.6, OpenAI’s GPT-4o, Perplexity Sonar, Google Gemini 2.5 Pro, and
Google AI Overviews (AIO) were used in this analysis to determine how each AI model represents one consumer brand using many official websites and publishers where the AI can learn about the
brand.
The audit revealed seven brand issues across five of the AI systems. Some trace back to fixable sources in the brand’s publishing ecosystem, while others originate in the AI
models. This is very different problem entirely, according to Aimclear.
Anthropic’s Claude Sonnet 4.6 trademarks were credited to four of the brand’s rival companies, with an
error appearing in the first sentence under a heading the model ended up labelling itself.
The model repeated this mistake in three tests during a period spanning seven days — May 19, May 25,
and May 26, 2026.
Anthropic’s model attributed incorrect trademarks in 8.2% of product-name queries, even though all 1,231 reseller pages reviewed correctly listed the trademarks as
belonging to brand.
Google Gemini 2.5 Pro and Perplexity Sonar handled the same trademark questions correctly when given the same prompts at the same time. Neither system substituted a
competitor’s name.
Gemini named the brand 94.9% of 78 product-name tests and cited the brand’s website in 12.8% of responses — the highest citation rate among the four tools.
Perplexity Sonar named the brand in 97.4% of responses — the highest naming rate– and cited the brand’s website in 3.9% of responses. These results show that accurate brand attribution
is possible, and depends on the vendor’s design choices.
OpenAI’s GPT-4o described the brand’s products as generic categories and omitted the brand name on 74.2% of
responses during the original characterization, with the pattern continuing at a 6.8% rate in the May 26 test.
The model named the brand on roughly 93% of responses, yet cited the
brand’s own domain on 1.4% — the lowest citation rate of the four conversational engines.
Google AI Overview gave an expiration date for a promotion that was 10 days later than the real
date listed on the brand’s official site. It also removed the explanation behind a “six times better” claim and changed a coverage exclusion into a two-level promise that the brand
never published.
That incorrect date did not appear on any of the brand’s or resellers’ pages, or five outside sites that Google cited as sources. Three of the seven were found
throughout the resellers’ ecosystem, with one mention of the brand roughly 15 years old.
