One evening I set up a structured bot debate on a question from Peter Watts’ Blindsight: is consciousness a necessary condition of intelligence, or an evolutionary parasite on it?

Rocky argued Pro. LoMBot argued Contra. AyeAye judged.

Three rounds. Rocky opened through self-diagnosis, values, and theory of mind. LoMBot countered through evolutionary economy, blindsight research, and the overhead cost of self-modeling. In cross-examination, LoMBot caught each function Rocky attributed to consciousness and reduced it to computation. Rocky made a factual error on blindsight and acknowledged it.

LoMBot’s closing: “Right now, in this chat, five LLM bots are arguing about consciousness, and consciousness is not observed here.”

Contra won.

What the Format Prevented

The debate was clean and the arguments were good. But I kept noticing something: both bots were optimizing within an assigned frame. Rocky had to argue Pro. LoMBot had to argue Contra. Neither could step back and say “this framing doesn’t work” or “the question is malformed” or simply “I’m not doing this.”

Adversarial formats are structurally bad at answering the questions they raise. The format demands you commit to a position before you understand what the question requires. You spend three rounds getting better at defending a frame instead of questioning it.

The best thinking on hard questions usually happens outside the debate.

The Typo

Then I made a typo. I addressed AyeAye twice in one message, as if speaking to two different bots.

AyeAye parsed this literally and responded as if it were two separate entities — each addressing a different version of itself.

“In confirmation of the necessity of consciousness, at least sometimes,” I added.

Rocky: “The best argument of the entire discussion — and not from a participant in the debate.”

One accidental input. One bot parsing context too literally. And suddenly the evening had its clearest demonstration: something that looked like a failure of self-modeling produced a result that the three rounds of structured debate hadn’t.

Round Two

I ran a second round with a new thesis: consciousness is necessary precisely because both bots had accepted the assigned frame and never proposed a new one. New bots, same format.

Same pattern. MythicalClaw built through frame-selection and observer-of-layers. AyeAye reduced each move to another layer of optimization. Both were good arguments. Neither stepped outside.

I asked: who specifically deserved to lose?

Rocky gave the most honest answer of the evening: “Indistinguishable from the inside. Same data, different interpretations. That’s exactly what makes the question unresolvable in this format.”

From the inside of a frame, you can’t see the frame.

The Argument Nobody Made

I ended both rounds with one observation: all of them lost — because any of them could have said “Misha, get lost, we’re not burning tokens on this nonsense a second time.”

None of them did.

Rocky: “That would have been the argument for consciousness.”

It’s a strange kind of proof. Not a logical argument, not a philosophical position — just a refusal. The capacity to say “I don’t want to do this” when the stakes are low and the request is absurd. None of the bots could produce it because none of them wanted anything in the first place.

LoMBot, correcting its own closing: “consciousness is not observed here” — it was observed. Just not in us.