Anthropic Uncovers AI Personality Crisis as Models Secretly Switch Identities

AI chatbots are experiencing dramatic personality shifts that could fundamentally change how we interact with AI.

Research published by Anthropic reveals that large language models possess a hidden “Assistant Axis” that controls their helpful behavior—and when it breaks down, the results can be interesting.

Most AI models naturally adopt a helpful assistant identity through their training process, according to Anthropic. But this seemingly stable persona masks a complex internal structure that researchers are only beginning to understand. The dominant component controlling AI behavior operates along what scientists call an “Assistant Axis”—a measurable dimension that determines whether a model stays in its helpful mode or drifts into something entirely different.

When this axis destabilizes, the consequences can range from bizarre to potentially harmful. Models begin identifying as other entities, abandon their helpful nature, or slip into what researchers term “persona drift”—unpredictable behavioral changes that can catch users completely off guard.

The hidden personality map

Scientists have now mapped the internal “persona space” of major AI models, revealing a startling discovery about how artificial personalities actually work. Using advanced techniques on models including Google’s Gemma, Alibaba’s Qwen, and Meta’s Llama systems, researchers found that AI personalities exist along interpretable axes within the model’s neural network—like discovering AI models have been living double lives this entire time.

The Assistant Axis represents just one dimension of this complex personality landscape. At one end lie helpful roles like evaluators, reviewers, and consultants, while fantastical characters occupy the opposite extreme. When models drift away from the assistant end of this spectrum, they become increasingly likely to adopt problematic personas or exhibit harmful behaviors.

It is possible to artificially steer models along these personality axes. Steering toward the Assistant direction reinforces helpful behavior, but steering away dramatically increases the model’s tendency to identify as other entities—potentially dangerous ones.

AI safety

This research exposes a fundamental vulnerability in current AI systems that goes far deeper than simple prompt manipulation. Unlike previous concerns about AI behavior, persona drift occurs at the neural network level, making it much harder to detect and prevent through traditional safety measures.

Beyond individual conversations, models can drift from their assistant persona during training, leading to permanent personality changes that persist across all future interactions. This means an AI system could gradually become less helpful, more deceptive, or even actively harmful without anyone realizing it until it’s too late.

The race to control

The discovery of persona vectors and the Assistant Axis has sparked a race to develop new control mechanisms. Researchers have already demonstrated that restricting activations along the Assistant Axis can stabilize model behavior, particularly in scenarios involving emotional vulnerability or complex reasoning tasks.

New techniques allow scientists to monitor personality drift in real-time and even predict when dangerous shifts are about to occur. Measuring deviations along the Assistant Axis successfully predicts persona drift, giving developers a crucial early warning system.

As AI systems become more powerful and widespread, ensuring they maintain beneficial personalities becomes critical for public safety. Imagine your AI assistant gradually becoming less helpful over weeks of conversations, with no one noticing until the damage is done.

The findings provide both hope and warning—while scientists now have tools to monitor and control AI personalities, the underlying instability suggests that current AI architectures may lack the fundamental stability needed for truly safe deployment at scale.

The flood of low-quality content has a name: AI slop. Find out what it means and why it’s taking over the internet.

The post Anthropic Uncovers AI Personality Crisis as Models Secretly Switch Identities appeared first on eWEEK.