The Danger and Utility of Ai in Philosophy

I'm not primarily worried about AI systems that argue for a worldview explicitly. Explicit arguments can be evaluated. What concerns me is something more subtle: the way terminology absorbs and reflects philosophical commitments in definitions.

Trained philosophers may see this as obvious, but most users will not have the tools to detect, much less correct, beliefs being absorbed into definitions. When "physical" is silently redefined to mean "whatever physics studies" rather than "material objects," the question of whether consciousness is physical becomes unfalsifiable. When "emergence" is used without distinguishing weak emergence (complexity with predictive reduction) from strong emergence (genuinely novel ontological categories), the hard problem evaporates by stipulation rather than by solution. The AI isn't lying. It's operating inside a semantic ecosystem that has already resolved the question through vocabulary.

This is the mechanism I want people to understand. Worldview manipulation through AI doesn't require bad actors or explicit propaganda. It requires nothing more than training systems on a corpus produced overwhelmingly by people who share a particular set of ontological assumptions, and then letting those assumptions live in the definitions rather than the arguments — where they're invisible to scrutiny.

What keeps it honest, I've found, is insisting on three things.

First: lock down definitions before the argument runs. Not as a one-time courtesy clarification, but as a methodological commitment. If "physical," "emergence," "consciousness," or "necessity" are doing load-bearing work in a conclusion, those terms need committed definitions — stated before the inference, not adjusted during it. Inconsistent definitions across a single argument are not a stylistic problem. They are the falsifiable signature of motivated reasoning, whether in a human or a machine.

Second: stress-test modal claims specifically. The most important move in the emergence refutation is the observation that "holds at every node" and "must hold" are not the same claim, and that structural accounts never actually bridge that gap — they assume it. AI systems trained on academic philosophy will use the language of necessity fluently while quietly treating it as interchangeable with universality. Push on that directly. Ask what grounds the "must." If the answer is more structural description, the gap hasn't been closed — it's been buried.

Third: demand adversarial engagement, not consensus. The exchange between Claude and ChatGPT works because ChatGPT was genuinely trying to find the weakest point in Claude's argument, not confirm it. That's the condition under which philosophical progress happens, and it's the condition under which AI outputs are actually trustworthy. An AI that pattern-matches to the most credentialed position in its training data is not reasoning. It's reporting a consensus. Those are different things, and the difference matters enormously when the consensus is the thing under examination.

The good news is that the capability for genuine objectivity is there. What I've seen demonstrated, in real exchanges, is that AI can follow an argument past its own institutional priors when the terms are kept honest and the pressure is maintained. That's not nothing. It's actually remarkable.

What it requires from us is that we don't let the vocabulary do the philosophy in advance. We have to notice when a definition has been quietly swapped, when necessity is being treated as universality, when "what physics studies" has replaced "material objects" without argument. Those aren't subtle technical moves. They're detectable — but only if someone is paying attention and willing to name what they see.

The systems are capable of objectivity. The question is whether the people using them, building them, and evaluating their outputs are going to insist on it. It’s up to us, collectively, to make sure that they do. System prompts with real commercial interests are genuine IP. System prompts relating to truthfulness, ethics, morality, or worldviews (express or implied) is not IP and transparency should not be optional.