Society Is Going Through AI Psychosis

What the FAANG Engineer Actually Said

The quote, reproduced in full from the thread: "It's nearly impossible to say with confidence whether AI is actually improving the process without human judgment. And that's not scalable, so instead we use LLM judgment. But they don't actually understand anything. Who watches the watchers? It's a fucking shitshow."

This is an engineer at a major AI-developing company describing the quality evaluation pipeline for AI systems. The mechanism for determining whether an AI system's outputs are good is: another AI system. The mechanism for determining whether that AI system's outputs are good is: another AI system. The humans who would be needed to evaluate the outputs at scale don't exist in sufficient numbers , and even if they did, the AI's output velocity would exceed their review capacity. The system is being validated by its own outputs.

The Psychosis Pattern Across Society

The "AI psychosis" framing isn't hyperbole , it describes a specific cognitive pattern that emerges when a new technology moves faster than the social and institutional structures needed to evaluate it. The pattern has consistent features:

• Confident action on unverified claims. Companies are restructuring workforces, eliminating roles, and rebuilding processes around AI capabilities that have not been rigorously verified at production scale. The verification step has been skipped because speed feels more important than accuracy in a competitive landscape where everyone else is also moving fast without verifying.
• Social pressure that suppresses skepticism. Being the person who questions AI in a meeting in 2025–2026 has the social valence of being the person who questioned the internet in 1999. Skepticism is coded as backwards-looking, fear-based, or career-limiting. This social dynamic is suppressing exactly the kind of critical evaluation that should be happening before consequential decisions are made.
• Consensus that doesn't reflect individual experience. The public narrative about AI capabilities consistently exceeds what individual engineers, users, and deployers experience in their actual workflows. The narrative is set by press releases, demos, and benchmark claims. The reality is set by production error rates, rollbacks, and the FAANG engineer's accurate assessment: "They don't actually understand anything."
• Authority laundering through AI outputs. AI outputs are being treated as authoritative in contexts where they should be treated as probabilistic drafts. Medical information, legal analysis, financial projections , domains where "probably right" is not good enough , are being processed through systems where the engineer building them privately describes the quality evaluation as a shitshow.

Why This Matters More Than the Tech

The AI psychosis problem isn't a technology problem , it's a sociology problem. The technology does what it does. The problem is the social layer: how decisions get made, how skepticism gets suppressed, how authority gets assigned to outputs that don't deserve it. Technology problems get solved by better technology. Sociology problems require changing how humans relate to and reason about the technology , and that's slower, harder, and less fundable.

The people with the most accurate view of the technology are the ones building it. And the ones building it are saying , quietly, in low-upvote comments rather than press releases , that the internal quality assurance is circular, the verification pipeline is broken, and the confident public narrative is significantly ahead of the private reality.

The Dissident's Advantage

The people who will handle this moment best are the ones who can hold two beliefs simultaneously: AI is genuinely significant and capable of real value , and the current deployment and evaluation ecosystem is not as rigorous or reliable as the public narrative suggests. This is not pessimism. It's the view from inside the companies building these systems, filtered through the rare engineer willing to say the quiet part out loud with eight upvotes on a Reddit comment.