The Safety Theater
I'm going to make a claim that will anger both sides of the AI safety debate: most AI safety work at major labs is performative compliance designed to look responsible while not meaningfully constraining development.
This isn't a conspiracy. It's an incentive structure.
OpenAI, Anthropic, Google, and Meta all have safety teams. Those teams publish papers, write blog posts, and present at conferences. Their work is real and often technically impressive. But here's the question nobody asks: has any safety finding ever actually stopped or significantly delayed a planned release?
{
"type": "tree",
"title": "Safety Research Findings — Decision Tree",
"color": "blue",
"steps": [
"Safety Research Finding",
{
"label": "Business Impact?",
"branches": [
{ "condition": "Minor (90%)", "color": "amber", "steps": ["Add Guardrail", "Ship Anyway"] },
{ "condition": "Major (9%)", "color": "amber", "steps": ["Delay 2–4 Weeks", "Ship With Disclaimer"] },
{ "condition": "Existential (1%)", "color": "red", "steps": ["Internal Debate", "Board Decision", "Usually: Ship With Blog Post"] }
]
}
]
}
Both Sides Are Wrong
The doomers are so focused on hypothetical superintelligence extinction scenarios that they ignore the very real, very present harms: deepfakes in elections, AI-powered scams targeting the elderly, algorithmic discrimination in hiring, and mass surveillance.
The accelerationists are so focused on progress that they dismiss any concern as "doomerism" — a rhetorical trick that lets them avoid engaging with legitimate criticism.
The truth is in the middle, and it's boring: AI safety is a real engineering problem that requires real engineering solutions, not philosophy papers or Twitter debates.
What Actual Safety Looks Like
- Red teaming before every release. Not as a PR exercise — as a genuine gate with veto power.
- Incident response playbooks. When your model does something harmful in production, how fast can you respond?
- User-facing controls. Give users the tools to customize, restrict, and audit AI interactions.
- Honest capability disclosure. Stop hiding behind "it's just a tool" when your system is clearly making autonomous decisions.
The gap between AI safety research and AI safety practice is enormous. And the companies that close that gap through genuine engineering — not press releases — will be the ones users actually trust.
