Peter Smith
AI systems are usually seen as neutral tools, but new research suggests something more interesting might be happening. The core question: are we teaching AI systems to mimic us, or are they developing their own ways of thinking?
What Model Specifications Actually Do
A collaborative study between Anthropic and Thinking Machines, announced by Jifan Zhang, examines how AI models' internal guidelines—called model specifications—shape their behavior in surprisingly consistent ways. Model specifications are essentially rulebooks that tell AI systems how to behave. They define what kinds of responses are acceptable, which topics to avoid, and how to balance traits like politeness, precision, and creativity. But these specs go deeper than technical settings—they're value judgments about what "good" AI behavior looks like. They determine whether a chatbot comes across as humble or confident, cautious or adventurous, strictly factual or conversationally flexible. In other words, they act as the AI's moral compass.
Testing AI Across Thousands of Scenarios
The research team ran thousands of controlled tests, presenting identical prompts to different AI models to see how they responded. The results showed consistent behavioral differences across major AI systems—each one had its own recognizable pattern. Some models answered conservatively, prioritizing safety and ethical guardrails. Others leaned toward analytical depth or creative risk-taking, offering bolder insights or more nuanced perspectives. These weren't random variations. They directly correlated with how each company structured its model specifications during training, suggesting that AI "personalities" are engineered, not emergent accidents.
Anthropic, the company behind Claude, has focused on "constitutional AI"—a method where models follow a set of guiding principles like a constitution. This research builds on that idea by exploring whether different constitutions produce measurably different personality traits. By partnering with Thinking Machines, Anthropic is digging beneath surface-level behavior to understand why models make specific choices. If personalities are designed rather than spontaneous, researchers could eventually create AI systems with consistent, transparent, and predictable behavior patterns.
As AI gets embedded in education, business, and healthcare, understanding its behavioral foundation becomes critical for trust. If model specifications shape personality, then those specs need to be visible and auditable—like ingredient labels for AI. Greater transparency would let organizations choose or design systems that match their values, whether they need a friendly tutor, a neutral analyst, or a bold creative collaborator.
Peter Smith
