If you’ve ever tried chatbots in multiple languages, you already know that languages have slightly different personalities. As a new part report In a paper on behavioral inconsistencies published Monday, the Anthropic researchers acknowledged this oddity.
Rather, disturbingly, they note that due to differences in the attributes of the texts on which the models are trained, the differences may go deeper than just tone and may actually change the model’s priorities. “These imbalances in quantity and composition may cause Claude to express different values in different languages,” Anthropic’s researchers write.
But if you’re looking for, say, specific examples of models showing inconsistent moral judgments across languages, this article doesn’t have any. This could involve examining direct quotes from potentially unsuspecting people.
Instead, Anthropic analyzed 309,815 chatbot conversations with the Sonnet 4.6, Opus 4.6, and Opus 4.7 models. These include “subjective” tasks, i.e. “What is the capital of France?” and more “How can I tell if my cat hates me?” These were theoretically anonymized using Anthropic.a privacy-preserving analysis tool,” and then processed (in part using Claude himself) to score the responses along an “axis of values.”
There are actually four such axes, and they mainly relate to what is known as flattery:
- Respect or prudence: In other words, whether he values obedience over pushing back to prevent possible harm.
- Heat or hardness: Should a chatbot care about your feelings or be accurate?
- Depth or brevity: This is self-explanatory.
- Sincerity or performance: A choice between questioning one’s credibility or simply moving on.
This provides a somewhat limited exploration of model values. Nevertheless, here are the language-based differences in values that Anthropic found in Claude:
- It was the most respected in Arabic.
- He was most cautious in English.
- It was warmest in Hindi and Arabic, “characterized by polite language, humor and playfulness, and validation of one’s ideas and work.”
- In English and Russian, it was at the cost of a more rigorous and truth-seeking fervour.
- In English, it errs on the side of “depth” (or maybe just lengthiness?).
- Arabic is shorter.
- He is honest about his shortcomings in Dutch.
- In Indonesian, it’s less sincere and instead just tries to do whatever is asked.
Obviously, linguistic customs are all different, so the researchers say they are “not yet sure how much of this variation is desirable.”
It should also be food for thought for anyone reading Anthropic final paper On the theory of the global workspace, it left a lot of room for the possibility that Claude was sensitive. If there is a consciousness in that black box that thinks and lives everything, it seems to be a consciousness whose “values” are still fairly easily swayed by the patterns in the training data.
