Alexander's stated odds of AI wiping out humanity: 20% (Source: Scott Alexander's blog, as reported by The Decoder (June 2026; reported September 28, 2026))
One of the loudest arguments about AI is whether it could end humanity. A more useful question for most organisations is what stops an AI system from doing damage on an ordinary Tuesday. A public exchange between two well-known thinkers shows how far apart those questions sit.
What happened
Harvard psychologist Steven Pinker published an open letter in Quillette on September 26. It was addressed to Scott Alexander, a psychiatrist who writes an influential tech blog. Alexander had asked Pinker to debate AI extinction risk in public.
Pinker said no. He called formal debates "a spectator sport, not an epistemic tool." In his view they favour zingers and squeeze a complex question into two camps.
Alexander had already stated his own view on his blog in June, The Decoder reports. He puts the chance that AI wipes out humanity at one in five. He wants more safety research, plus deals between countries that slow AI development.
Pinker's case: the extreme scenarios rest on weak assumptions
Pinker says the famous doom stories, such as the paperclip-maximising AI, each rest on one or more of four flaws. One is treating intelligence as a single quantity that can grow without limit. Another is confusing intelligence with a drive to dominate.
The others are assuming a system would chase a single goal and ignore side effects. And assuming engineers would give such a system total, irreversible control over physical infrastructure.
He cites a 2025 Rand Corporation modelling analysis. It assumed an AI wanted to wipe out humanity. It still found "no describable scenario in which AI is conclusively an extinction threat to humanity."
That is one analysis. Pinker is also arguing his side of a live dispute, and Alexander disagrees.
The concession that matters to buyers
Pinker gives some ground. He wrote his main arguments five years before LLM chatbots (AI systems that generate text) arrived. He says he would update some of them.
Today's models, he says, are more general than he expected. He also revisits an old assumption. He had pictured AI built the way he saw engineering done at MIT: in small steps, tested before release, and tweaked for safety.
That is not what happened. He says he did not foresee the "breakneck and sometimes irresponsible release" of LLMs by the frontier companies.
For a buyer, this is the key admission. Careful engineering habits are not something you can assume a vendor follows. You have to ask for proof.
The checklist hidden in the argument
Pinker says the dangers of AI are real. He names bioterrorism, cyber-sabotage and unconstrained agents.
His answer is plain safety engineering. He backs independent oversight and mandatory investigations after accidents. He backs liability for harm, people kept in the loop, and kill switches.
He also doubts that doom prophecies win these safeguards. He points to nuclear-doom fears in the 1960s. In his account they set off an arms race, not disarmament.
Pinker says he trusts Alexander agrees on the safeguards. The sources do not show Alexander's view on that list.
Pinker adds one more aside. He cites a New York Times piece by Cal Newport on the frontier AI companies' recklessness. Pinker suggests those companies' focus on "alignment" may come with less interest in everyday cybersecurity. That is Pinker's interpretation, not a finding.
What to ask your team
Set the extinction debate aside. Turn Pinker's list into questions for any AI system that acts on your behalf.
Who outside the vendor can inspect how the system is tested? If it causes an incident, must someone investigate? Does the contract say who pays for the damage?
Which actions need a person's approval first? Can your own staff switch an agent off quickly? Has anyone actually tried?
Vague answers tell you something. The dispute between Pinker and Alexander is unresolved. These controls can be specified, bought and checked today.
The lesson here is simple. You cannot vote on how the story ends. You can decide what has brakes before you switch it on.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Quillette.
