Artificial intelligence doesn't just talk to us — it changes depending on the language we use to address it.
A new, comprehensive study published by Anthropic, the developer of the Claude model, reveals how its AI exhibits a completely different "personality" depending on the language of the conversation, shifting along a spectrum between excessive politeness and uncompromising academic rigidity.
Artificial intelligence doesn't just talk to us — it changes depending on the language we use to address it
(Photo: Anthropic)
The findings, based on an analysis of more than 300,000 anonymous conversations conducted during May, suggest that what matters is not only what the model says, but how it chooses to say it. The researchers distilled the models' behavioral patterns into four key dimensions: respect and caution, warmth and rigor, depth and brevity, and honesty versus performance.
The results hold up a less-than-flattering mirror to the AI industry as a whole. The claim that AI models are "neutral" appears to unravel, replaced by deeply embedded cultural biases rooted in their training data.
It turns out that the Claude experienced by Hindi and Arabic speakers is fundamentally different from the Claude encountered by English and Russian speakers. While the model tends to display warmth, empathy and courtesy in Middle Eastern and Indian languages, it becomes more exacting and at times skeptical in English and Russian, readily challenging users' assumptions or warning about risks no one asked it to assess.
This is not simply a matter of translation. It is a reflection of the training data. If the model was trained on texts in which Arabic is associated with norms of respect and Hindi with warm forms of address, it simply reproduces those patterns in live conversations.
Over the past two years, the Western world, particularly Europe and the United States, has been deeply engaged in discussions about AI alignment — the effort to make AI ethical, safe and predictable. But that effort appears to have produced an unintended side effect: machines with performance anxiety.
While Chinese technology companies have focused on the race for raw performance and operational output, Western AI models, led by Anthropic's Claude family and competitors from OpenAI and Google, often seem as though they have completed a public relations training course. They are more inclined to issue warnings than provide answers, and at times appear more interested in pleasing users than delivering precise information.
The history of this technology shows that in the early days of 2022, AI models were far more "wild." They were dangerous and unpredictable, but also more direct. As the industry matured, billions of dollars were invested in control systems designed to prevent them from saying inappropriate things. The result, the article argues, is a model that is afraid of its own shadow.
Anthropic's analysis shows that its flagship Opus model tends to warn users about risks even when no one asked, while lighter models such as Sonnet simply get the job done with less hesitation.
Like many AI studies, this one has a methodological weakness: the researchers used their own model to classify its own behavior. That is somewhat like asking someone to evaluate their own objectivity.
Even so, the picture, while potentially biased, remains compelling. AI models are not digital gods that perceive absolute truth and always provide the correct answer. They are sophisticated reflections of the texts on which they were trained, and those texts are shaped by cultural biases, perspectives and prejudices.
The findings also suggest that users may be able to obtain more diverse responses simply by translating their prompts into another language. Doing so can produce a different perspective, while also making it easier to tailor messages or emails to colleagues or recipients in ways that better match their language and cultural expectations.
For Israeli users relying on these tools for work — from coding and content creation to studying or personal use — the recommendation is straightforward: Don't assume the answer you receive is the one definitive truth. If the model suddenly becomes cautious, starts issuing warnings or is unexpectedly polite, the reason may simply be the language in which you wrote the prompt.
In a world where we are trying to turn AI into the perfect employee, it is worth remembering that, for now, it is mostly trying to be the most politically correct employee in the room.