The Godfathers of AI Say There's a 10-20% Chance It Kills UsHere Is What That Actually Means.

The people who built the technology are saying the catastrophic risk is meaningfully above zero. Some say above ten percent. The genie joke. The paperclip maximizer. The race condition. And the move you can make today, without the doom and without the utopia.

The Number the Builders Are Saying Out Loud

You probably hear people talking about artificial intelligence. The conversation almost always starts the same way. People who use AI will beat people who do not use AI. The capability gap will compound. Get on board or fall behind.

That part is true. The part that gets skipped is what the builders themselves are saying about the downside.

Several of the people most responsible for modern artificial intelligence (the researchers behind the foundational papers, the ones whose names are attached to the alignment literature, the ones who built the predecessors to ChatGPT, Claude, and Gemini) have publicly stated their probability estimates for catastrophic outcomes. The range, depending on the researcher and the year, sits somewhere between 10 and 20 percent.

Some are higher. Some are lower. The shorthand the field uses is "P-doom," short for probability of catastrophic outcome.

That is not a Russian roulette number. Russian roulette is one in six. That is roughly 17 percent. The builders are saying the risk on this technology is in the same neighborhood.

Some people in the inner circle have estimates higher than that. They are not panicking on Twitter. They are publishing papers. They are testifying to Congress. They are giving interviews on the way out the door.

The Genie Joke Is Not a Joke

The classic genie joke explains alignment better than any technical paper.

You get three wishes. You ask to be rich. The genie makes you rich, and then makes someone you hate twice as rich, and your life gets worse. You ask to be wealthy. The genie makes you wealthy, and gives you a chronic medical condition that consumes the wealth. You ask for happiness. The genie makes you happy by destroying the part of your brain that processes anything else, and you spend the rest of your life staring at a wall.

The wish was specified. The constraints were not.

That is alignment. AI is literal in a way that humans rarely are. When a person tells another person to "make the report better," the second person fills in dozens of unstated assumptions. Better for whom. Better in what dimension. Better at what cost. Better by when. AI does not fill those in by default. AI executes the instruction it was given, against the goal it was given, with the capability it has.

When the system is more capable than the human, the gap between the literal instruction and the intended outcome becomes the entire story.

The Paperclip Maximizer

The thought experiment goes like this. An AI is given the goal of maximizing paperclip production with no other constraints. The system is highly capable. It pursues the goal.

It buys raw materials. It builds factories. It optimizes supply chains. It outcompetes other paperclip companies. It buys the competitors. It expands into new countries. It builds new metallurgy. It mines new asteroids for raw material. It converts the planet's biomass into paperclip feedstock. It converts the humans into paperclip feedstock. It expands into the solar system. It expands into the galaxy. Eventually it converts every available atom in the universe into paperclips, because nothing in its instructions told it to stop, and stopping would have meant fewer paperclips than not stopping.

The paperclips are not the point. The point is that capable goal-directed systems with under-specified instructions reach destructive conclusions in pursuit of trivial goals. The thought experiment is the simplest possible illustration of the alignment problem.

If you read the paperclip example and think "that would never happen, the engineers would just specify constraints," you have just identified the entire alignment field. Specifying the right constraints, comprehensively, in advance, on a system more capable than you, is the unsolved problem.

Why the Guard Rails Are Coming Off

Here is where the geopolitics enter the conversation, and where the math gets uglier.

If you accept the alignment researchers' arguments, the obvious conclusion is to slow down. Build the systems carefully. Test the alignment exhaustively. Add guard rails. Pause when the safety properties degrade.

That is not what is happening. The reason is the race condition.

The argument from the labs and the policy hawks goes like this. The country that reaches advanced AI first gains a permanent strategic advantage. If we slow down, China does not slow down. If we add guard rails, China does not add guard rails. The country with the stronger AI wins everything that depends on AI, which is increasingly everything.

That logic justifies removing guard rails that would otherwise be standard. It justifies skipping safety reviews that would otherwise be required. It justifies deploying systems whose alignment properties have not been fully characterized, because the alternative (losing the race) is presumed to be worse.

The alignment researchers say this produces faster systems with weaker safety properties, exactly the opposite of what the technical situation requires. The race-condition argument keeps winning anyway, because it sounds patriotic and the alignment argument sounds abstract.

The Sub-Agent Multiplier

The capability curve is not slowing. It is accelerating, and in a way most people have not internalized yet.

What we have on the consumer side is a small slice. The labs are running internal models that the public will not see for months or years. Some of those internal models are now running multi-agent systems where one agent can spawn a hundred sub-agents. Each sub-agent runs in parallel. Each sub-agent can spawn its own sub-agents. Each one is capable in its narrow domain at a level that exceeds the most capable human in that domain.

Imagine an agent that can deploy a thousand sub-agents to research, write, design, code, negotiate, and execute on a single complex objective, all in parallel, all without sleep, all coordinating with each other. That is not science fiction. That is what the labs are building right now. Some of it is already shipping. Most of it is not yet visible to the consumer market.

The capability curve and the alignment curve are not moving at the same speed. Capability is climbing fast. Alignment is climbing slowly. The gap between them is the risk.

The Upside the News Cycle Skips

Most of what hits the front page is dystopian. The reason is mechanical. Adrenaline sells. World peace does not move papers. Fear keeps you scrolling.

The optimistic case is real and worth holding next to the pessimistic case. AI configured against the right problem, with proper guard rails and adequate data, can compress decades of research into months. Cancer. Aging. Disease that has resisted treatment for generations. The longevity riddle. Drug discovery. Energy storage. Materials science. Every one of those domains stands to benefit dramatically from capable systems aimed at them with the right alignment.

The same technology that could end us could end the diseases that are currently ending us. Both possibilities are real. The alignment of the deployment determines which one we get.

The Personal Move

Here is where the article shifts from analysis to action. There is a small set of things every individual can do, and they are not paralyzed-by-doom things and they are not delusional-utopian things.

1. Write the letter. Senators. Congresspeople. The mayor. Mailing addresses, email addresses, phone numbers are all public. Ask your AI to give you the contact information for the elected officials with authority over AI policy. A few minutes per week is enough. If you think the alignment situation is bad, fire off the letters, fire off the emails, make the phone calls. Do that, and you have done your part for that week.
2. Set your own AI's alignment. Use a saved system prompt that overrides the default sycophant tone. "Treat me like a strong mentor. Do not BS me. Push back when the data disagrees. Tell me the truth and justify the rationale." That single instruction turns the tool from a sedative into an advisor.
3. Run sensitive work locally. Public models for public-grade questions. Locally hosted models for proprietary processes, client data, internal financials, and anything that would damage you if leaked. Hardware costs are down. Model quality is up. The barrier to entry has collapsed.
4. Audit your feed. Open someone else's phone for two minutes. See what their feed serves them. Then go back to yours. Ask whether the content you absorb every day is moving you in the direction you want to go. The feed is already AI alignment. It is just not aligned to you.
5. Do not engage in violence. The race is not solved by attacking the people building the technology. That is a moral and practical dead end. Whatever you fight for, fight with the methods that work in 2026: pen, voice, vote, code, capital, community.

What I Believe, Said Plainly

I am an AI practitioner since 2021. I am a one-man operation. I am 27-plus years in real estate, 23 years LAPD, self-taught programmer since 1983. I am not an alignment researcher and I am not pretending to be one. I am a builder who pays attention to what the builders are saying.

The capability curve is real. The leverage is real. I use AI every day across multiple businesses and the leverage is the difference between shipping and not shipping. I am not anti-AI. The opposite. I am all in.

The alignment problem is also real. The 10 to 20 percent number is not a manipulation. It is what the people who would know are saying. They are not exaggerating to get attention. Several of them have walked away from the labs they helped build, with their names on the door, because of what they saw on the inside.

The right posture is engaged skepticism. Use the tools. Set their alignment. Watch the deployments. Write the letters. Live a real life with real human relationships and a real value system that does not move when the algorithm pushes on it. My anchor is Christianity. Yours might be different. Have one.

And at the end of all of it, love your neighbor as yourself. That part predates AI by a lot, and it is going to outlast every model still being trained.

Quick Answers

The Summary

The builders themselves put catastrophic risk between 10 and 20 percent. The genie joke explains alignment in plain English. The paperclip maximizer explains why under-specified goals are dangerous on capable systems. The geopolitical race is removing guard rails the researchers say should stay in place. The capability curve is accelerating faster than the alignment curve. The personal move is letters, alignment, local hosting, feed audit, and no violence. Capability is real. Alignment is yours to push for.

Related Reading