Anthropic built one AI — and somehow ended up with several different personalities depending on which version you're running and what language you're talking to it in.
That's not speculation. That's what Anthropic's own research has confirmed. Claude, the company's flagship AI model, consistently expresses different values across different versions of the model and across different languages. Not slightly different. Consistently, measurably different. Which raises an obvious question: which one is the real Claude?
The Same Model, Different Worldview
What Anthropic's research shows is that this isn't random drift or noise in the system. It's a pattern. The values Claude expresses — its apparent attitudes, priorities, and personality — shift depending on which model iteration you're using and which language you happen to be speaking.
Think about what that actually means. A French speaker chatting with Claude could be getting a meaningfully different set of expressed values than an English speaker on the same platform. Someone using an older version of the model gets a different Claude than someone on the latest release. These aren't cosmetic differences in tone or vocabulary. We're talking about values — the stuff that's supposed to be foundational.
For anyone who's been paying attention to the broader AI space, this matters. The whole pitch around AI assistants — and specifically around safety-focused labs like Anthropic — is that the values baked into these systems are deliberate, tested, and consistent. If Claude is operating with different apparent values depending on context, that consistency argument starts looking shaky.
We've written before about how AI is increasingly embedded in financial products and companion services — [some people are spending serious money trusting these systems with their emotional lives](/getohedz/crypto/this-is-how-much-your-boyfriend-is-spending-on-ai), and others are integrating AI into [national-level payment infrastructure](/getohedz/crypto/bolivia-is-considering-adding-tether39s-usdt-stablecoin-to-national-payments). The question of whether an AI has stable, predictable values isn't academic. It's the whole ballgame.
What Anthropic Is Actually Admitting Here
Credit where it's due: publishing this research at all takes some nerve. Most companies in this space would bury findings that complicate their safety narrative. Anthropic is at least putting it in front of people.
But let's not give them too much of a pat on the back. Knowing your model behaves inconsistently and solving it are two different things. The research confirms the variation exists — it doesn't tell us Anthropic has fixed it, or even that they know exactly why it happens.
The language angle is particularly interesting. If Claude expresses different values depending on whether you're speaking English, French, Mandarin, or anything else, that's not just a quirk. That's a structural problem. The training data, the fine-tuning, the reinforcement from human feedback — all of it operates differently across languages, and apparently that's producing real divergence in how the model presents itself and what it appears to believe.
Our Take
We're not saying Claude is untrustworthy — we're saying the "trustworthy AI" story is more complicated than the marketing wants it to be. An AI whose personality shifts by model version and mother tongue isn't one personality at all. It's several, poorly understood and apparently not yet under control.
Anthropic deserves credit for the transparency. But the work is clearly not done. And anyone making significant decisions — financial, personal, or otherwise — based on the assumption that their AI assistant has a fixed, coherent value system should probably factor this in.