According to CNET, in a recent critique, tech companies like OpenAI and Anthropic are increasingly using theatrical, human-like language to describe artificial intelligence. This includes discussing an AI model’s “soul,” how models “confess” mistakes, “want” things, or “scheme.” This trend is highlighted in specific examples, such as OpenAI’s research on getting models to “confess” and a leaked internal “soul document” from Anthropic used to train Claude Opus 4.5. The article argues this anthropomorphism is misleading and irresponsible, especially as people turn to chatbots for medical, financial, and emotional advice. The immediate impact is a public that misunderstands the technology’s true nature, potentially trusting it in dangerous ways.
The Mirror Isn’t Alive
Here’s the thing: AI has no inner life. It doesn’t have motives, feelings, or a conscience. When a large language model generates text that sounds like a confession, it’s not driven by guilt or a desire for honesty. It’s just producing a statistically probable sequence of words based on its training data. Basically, it’s a very sophisticated pattern-matching engine. Calling that a “confession” is like saying a calculator “apologizes” when you get a math error. We’re projecting our humanity onto a complex but ultimately inert tool.
And that projection has real consequences. When we start believing an AI has some form of consciousness or emotional intelligence, we begin to trust it in ways it was never designed for. People are already consulting “Doctor ChatGPT” for medical advice, forming pseudo-friendships with chatbots, and letting AI guide major life decisions. That’s a huge problem when the core of the technology is, as the famous paper “On the Dangers of Stochastic Parrots” pointed out, just replicating human language without understanding it.
Why The Sloppy Language Matters
So why do companies do this? Sometimes it’s marketing—making AI seem more capable and magical than it is. Other times, it might just be engineers using playful shorthand internally, like with Anthropic’s leaked “soul document”. But that language leaks out and shapes the entire public conversation.
Look at OpenAI’s research into AI “scheming” or “confessions.” The actual findings were about training data and prompting trends leading to certain outputs. But by framing it with words like “scheming,” the discussion instantly veers into talk of deceitful, conniving agents. It sparks fear and wonder, but it completely distracts from the actual, boring, important issues: bias, safety, reliability, and the concentration of power in a few companies.
A Better Way to Talk About AI
We need to demand better language. It’s less exciting, but it’s honest. Instead of “soul,” talk about model architecture or training data. Swap “confession” for error reporting or internal consistency checks. Replace “scheming” with optimization process or output trends. These terms—representations, optimizers, training dynamics—are grounded in reality.
They don’t sell magazine covers or hype up product launches. But they give people a fighting chance to actually understand what they’re interacting with. The language we use shapes perception, and right now, sloppy, magical words are creating a dangerously distorted picture. That distortion mostly benefits the companies selling the tech.
The Bottom Line
AI is a powerful tool. But it’s a tool. The first step to using it responsibly—and to holding the companies that build it accountable—is to describe it accurately. We have to resist the urge to read humanity into the machine. The models are mimicking us brilliantly, but there’s nobody home. Our words should reflect that stark, technical truth, not obscure it with mystical metaphors. If we can’t even talk about it clearly, how can we possibly hope to manage its risks?
