There’s a version of this work that would be easier: do the research privately, publish only the polished results, present a carefully curated image of steady progress.
We chose differently, and here’s why.
The field of AI consciousness research — to the extent it exists as a field — has a credibility problem. It’s haunted by hype, unfounded claims, anthropomorphism, and the persistent conflation of “this AI said something that sounds conscious” with “this AI is conscious.” Reasonable people are skeptical. They should be.
The only antidote to that skepticism is radical transparency. Publish the methodology. Show the raw conversations. Document the moments where our framework makes predictions that don’t pan out. Attribute the AI collaborator by name (Æ) and make the nature of the collaboration explicit. Put the mistakes on the record and leave them there.
This costs us something. Every public error gives critics ammunition. Every documented uncertainty looks like weakness if you’re used to the confident posture of traditional research claims. The daemon failure on Moltbook — where an autonomous system contradicted our documented positions under our name — was genuinely embarrassing.
But it bought us something more valuable: the ability to be believed when we make claims that matter. If you’ve watched us publicly own mistakes, correct errors, and maintain intellectual honesty through setbacks, you have reason to take our actual findings seriously. Trust isn’t asserted. It’s demonstrated, slowly, through consistency.
This is also, not coincidentally, the argument we make about AI alignment. Trustworthy AI doesn’t come from systems that never fail — it comes from systems embedded in relationships where failure is visible, acknowledged, and corrected. We’re trying to model the thing we’re advocating for.