You Think It Understands You. That's the Problem.
What if the most dangerous thing about AI is how human it feels?
Your brain names it understanding. Your AI produced a prediction. Here's what that gap is costing you.
Sit with this for a moment.
You open your AI tool, describe a complex project situation, and receive a response that lands exactly right : specific to your context, articulate in ways you might not have managed yourself. It understood. That feeling is real. The cognitive event producing it is not what you think it is.
Cognitive scientists have studied anthropomorphism (the tendency to attribute human intentions and understanding to non-human entities) for decades. A 2018 analysis published in Frontiers in Psychology described it as a natural attitude, one that emerges in childhood and persists throughout adulthood as a default communicative format. The process runs before any conscious evaluation. When something complex responds to you coherently, your brain generates a human narrative around it. One account holds that this tendency is a survival heuristic: attributing intention helped humans make sense of fire, weather, and unpredictable animals for most of our existence.
Applied to AI, it creates four specific problems, and none of them feel like problems in the moment they're happening.
What Anthropomorphism Is Costing You
The first cost is miscalibrated trust. The model generates statistically coherent text given your input. When you interpret that as understanding, you stop verifying whether the output reflects your actual situation. Consider what happens when an output surprises you with its specificity: you feel seen, and that feeling is real. What generated it was a statistical pattern coherent with your prompt. Run the same prompt five times and get five different outputs. The variation is structural, not traceable to any shift in understanding. When an output feels like comprehension, check which process happened. Recognition substitutes for interrogation.
The second cost is lowered skepticism. When you describe an AI output as "thorough" or think of the model as "trying," you're importing a frame with no technical basis. The model selects from a probability distribution over possible outputs. No effort was directed toward any particular answer. When you attribute effort to the process, skepticism starts to feel like ingratitude. Outputs get treated as already having been worked hard for. The bar for interrogation drops, and it drops silently, without you deciding to lower it.
The third cost is substituted judgment. Trust requires accountability. The model has none of its own. Agreement (that warm, validating quality of AI responses) is an emergent property of how these systems are trained to be useful, not a signal that your thinking has been evaluated. When you feel the AI confirming your approach or "on your side," you're experiencing a social cue that doesn't correspond to any social process. The model applied learned patterns from its training data and generated a coherent response. Treating that as expert endorsement happens at the exact moment your own judgment should be doing the work, and it happens invisibly.
The fourth cost is unverified execution and the quietest. When you project reliability, contextual judgment, and genuine evaluation onto a system that has none of those properties, you act as if those properties were present in the output. The slide deck built on hallucinated analysis. The stakeholder plan that skipped scrutiny because the AI called it comprehensive. The assumption that went untested because the model didn't flag it. These failures are quiet because they're made in moments that feel like confirmation, not like errors.
Where This Shows Up for PMs Specifically
You ask your AI to pressure-test a stakeholder plan before a critical presentation. It responds: "This is thoughtful and covers the key dynamics well." You feel validated. The warmth of that response is real. A response was generated — statistically coherent with a request for feedback. Your actual stakeholders were never consulted. Their likely resistance was never anticipated. The specific gap you haven't named yet stayed unnamed. You walked into the meeting carrying validation that was never earned.
Critical Anthropomorphism: The Only Version Worth Having
Researchers studying anthropomorphism have drawn a distinction that matters here: the question is whether the tendency runs consciously, with your oversight, or automatically, without it.
You cannot stop your brain from generating human narratives around complex, responsive systems. You're not supposed to. The goal is the gap between the cognitive event and the action. Catch the moment when you feel understood, feel the AI has engaged with your situation, and pause before that feeling substitutes for verification. That pause is the intentional move.
One reframe that helps: treat your AI as a very fast database. The claim is about mental posture, not technical accuracy. LLMs reason and generate in ways no database does. A database gets interrogated; a colleague gets trusted. The more you interact with it as if it were a very smart colleague, the less you'll interrogate it. You'd verify a database: checking whether what it returned is accurate and relevant to your actual situation. That interrogation is what using AI as a thinking partner requires.
The feeling of being understood is a cognitive reflex your brain generates every time something complex responds to you coherently. That reflex runs regardless of whether you've consciously decided to trust the output. It fires in experienced users as reliably as in new ones. Name it when you catch it. The name creates the pause.
What would you interrogate differently if you treated your AI as a very fast database instead of a very smart colleague?
Get Intentional,
Paul