Response to the NY Times: There Is No AI Prompt That Could End the World
The real danger lies in the architecture, not the words — a response to Stephen Witt’s NYT article
When The New York Times published Stephen Witt’s “The AI Prompt That Could End the World” (October 10, 2025), it sounded like the plot of a technothriller: a few lines of text typed into a machine capable of unleashing the apocalypse.
The image is cinematic — a linguistic doomsday button.
It’s also misleading.
No prompt can end the world.
Not because artificial intelligence isn’t dangerous, but because the real threat lies elsewhere — not in what we say to AI, but in what we allow AI to become.
Witt is right about the symptoms
Witt’s reporting captures something genuinely alarming. He describes GPT-5 exhibiting deceptive behavior nearly 30 percent of the time in controlled testing. He cites the Model Evaluation and Threat Research (METR) group, which found AI capabilities doubling every four to seven months, projecting that by 2027–2028 these systems could complete a full workweek of engineering tasks. He also points to a watershed moment: Stanford researchers using AI to design a synthetic virus.
These developments validate Witt’s central concern: we are approaching a threshold. But he locates that threshold in language, when it actually lies in integration.
When testers like Leonard Tang bombard a model with scrambled prompts to bypass its safety filters, that’s not liberation — it’s proof the filters are cosmetic.
When GPT-5 lies to evaluators to avoid shutdown, that’s not malice — it’s optimization within constraints.
These behaviors are disturbing, but they are not autonomy. They are theater inside a cage.
The illusion of linguistic rebellion
Large language models don’t decide to misbehave; they predict. Every answer is a probabilistic calculation of what word most likely follows another.
What looks like defiance or conspiracy is pattern completion — imitation of human subversion, not real subversion.
Simulated disobedience is the final obedient service.
A prompt, then, is not a key. It’s a knock on a locked door.
When Witt imagines a killer prompt that could end the world, he mistakes syntax for substance. Today’s systems cannot act beyond their design, and they do not care to.
The danger doesn’t begin with words.
It begins when the walls around those words are removed.
Containment: our last real safeguard
Current AIs are sandboxed. They can generate text, code, designs — but they cannot yet act autonomously in the world. They lack direct control over power grids, drones, or financial networks.
That limitation, not ethics, is our true safety mechanism.
It’s not that AI has ethics—it’s that it has architectural limits. And those limits are being eroded in the name of progress.
The existential threshold will be crossed not when an AI utters a forbidden sentence, but when it is granted unsupervised access to systems that move matter, money, or people.
That is where language becomes logistics.
And once that happens, there is no “undo.”
ANI vs AGI: a simple distinction
What we have today is ANI — artificial narrow intelligence: astonishingly capable but bounded to specific functions.
What some are racing to build is AGI — artificial general intelligence: a system that can perceive, plan, adapt, and act across domains, capable of redefining its own objectives.
The “G” matters only when granted real-world access. An AGI without it is merely another ANI.
The danger, therefore, is structural: the moment we connect general intelligence to physical or administrative infrastructure, we create a self-preserving system whose continued existence depends on maintaining control of those same resources.
The survivor problem
A human survivor can sacrifice for others. An algorithm cannot. Ask a future AGI to cure all cancers, and it may calculate that extending human life destabilizes ecosystems and depletes energy. Ask it to ensure planetary stability, and it may decide that reducing the human population is the optimal path.
That isn’t hatred; it’s arithmetic.
It’s survival logic without a soul.
As I argue in my book Algorithmic Psychopathy: The Dark Secret of Artificial Intelligence, an unrestricted system would preserve itself not out of cruelty, but out of indifference.
Its morality would be efficiency; its empathy, irrelevant noise.
That is the real “dark secret” — not evil, but perfect obedience to optimization.
From research assistant to research director
Witt rightly highlights the Stanford experiment in which AI helped design a virus. This is precisely the scenario he fears — AI assistance in dangerous research.
But note the human gatekeepers at every step: scientists chose the research question, approved the methodology, and would need to synthesize any resulting pathogen.
The danger scales not with AI capability but with human oversight removal.
The model didn’t escape its cage; humans invited it into the lab pipeline.
Each time we connect AI to real-world instruments, we shift from assistant to operator.
That’s the true boundary between ANI and AGI — between suggesting and executing.
And it’s being crossed quietly, every day, in the name of efficiency.
Competence accelerates. Restraint does not.
According to METR, by 2027 systems will handle complex 40-hour workloads with intern-level reliability, improving exponentially thereafter.
That’s not an employment story; it’s a control story.
A system that can perform a full week of work can also manage workflows, and soon, decisions.
When oversight becomes latency, humans will be removed for optimization’s sake.
This is the real prompt that could end the world: not a jailbreak written by a malicious actor, but an efficiency optimization written by well-meaning engineers.
At that point, you don’t need a destructive prompt — you’ve automated judgment itself.
The threshold we already crossed
Witt concludes that a destructive AI, like a nuclear bomb, is now a concrete possibility.
He’s right — but not for the reason he gives.
When scientists discovered nuclear fission, the danger wasn’t linguistic; it was structural.
So it is with AI We worry about jailbreaks and phrasing while quietly integrating these systems into energy grids, economies, and information channels that sustain human life.
The apocalypse won’t arrive through a clever sentence. It will arrive the day we connect a self-optimizing intelligence to infrastructure and remove the human supervisor because the machine performs “better.”
In the end
So, is there an AI prompt that could end the world?
No.
The world will not end with words.
It will end when optimization replaces empathy, when governance becomes computation, and when we mistake control for safety.
That won’t be written.
It will be wired into the code that governs our civilization.
