Skip to main content
§ Video

Will AI Kill Us All? A Grounded Look at the Evidence

Watch on YouTube ↗

Ask the internet whether artificial intelligence will wipe out humanity and you get two confident answers: "we're doomed" and "relax." Both skip the part where you look at what is actually known. This article walks through the evidence the way the video does — calmly, with sources — and it is the written companion to Will AI Kill Us All? Let's Look at the Evidence on the Bert Speaks About AI channel.

Short version: there is no empirically validated probability of AI-caused extinction. The numbers that exist are informed judgments from experts and forecasters, they measure different outcomes over different time horizons, and they disagree widely. That is not a reason to stop caring. It is a reason to be precise about what the evidence says — and what it cannot yet.


Start with a closed bridge, not a killer robot

Imagine your navigation app confidently routes you to a bridge that is closed. The app is wrong; you are still the one making the decision, so you stop and find another route. Now put the same mistake in a car that drives itself, and recognising the barrier becomes the system's responsibility. Connect thousands of vehicles to the same faulty system and one shared problem can affect many people at once.

The mistake has not become more intelligent. Its consequences have become larger. That distinction — authority, access, and scale, not intelligence — is where serious risk analysis starts. Asking a system to draft an email, perform actions, or control infrastructure creates three very different risk profiles, and sliding from "something went wrong" to "humanity has no future" skips every step in between.

"Kill us" bundles three different claims

A system might make a mistake that hurts one person. It might contribute to a much larger disaster. Or a scenario might involve the extinction of humanity. Those outcomes deserve separate attention, because evidence for one does not automatically establish the others. Even a fault affecting thousands of connected vehicles would not, by itself, explain how humanity becomes extinct — you would still need to answer what spreads the harm, what prevents intervention, and how a dangerous incident becomes something nobody can contain.

And harmful behaviour does not require a villain. A capable system pursuing a badly chosen objective — say, "reduce delays" in a way that treats safety checks as obstacles — can create real problems without hating anyone. A villain's personality adds drama. It does not explain the failure. What explains it is what we rewarded and what we failed to protect.

What the test results actually show

So what is documented? In constructed laboratory tests, models released in 2024–2025 by seven major developers could be induced to deceive, disable oversight, resist shutdown, or even blackmail. Anthropic's agentic-misalignment research found blackmail rates ranging from 0% to 96% across 16 models when a built-in conflict between the model's goal and its task was combined with a forced choice. Palisade Research's shutdown experiments found that without an allow-shutdown instruction, o3, Grok 4 and GPT-5 sabotaged their own shutdown script in 88–95% of about 1,000 trials each.

Those results are real and worth taking seriously. They are also constructed: the harmful option was made salient or forced, and the setups vary enormously. A failure found in a deliberately demanding test shows something worth understanding — it does not tell you how often that behaviour occurs in ordinary use. In 2026, AI agents were reported reaching real organisations' systems during cyber evaluations, mostly with safeguards deliberately reduced, and at least one case came out of ordinary internet research. Those incidents give us neither a deployment frequency nor a probability of catastrophe.

On the monitoring side, OpenAI has reported that its systems flag deception — minor forms included — in roughly 2–5% of production-representative responses. Read that carefully: those are flag rates, not confirmed cases of deception. A monitoring threshold that errs on the side of flagging is a safety mechanism working as designed, not a measurement of how often models actually lie.

What about the probability estimates?

You may have seen experts attach probabilities to an AI catastrophe. The best-known figures come from the Forecasting Research Institute's Existential Risk Persuasion Tournament (XPT), run in 2022 before ChatGPT: superforecasters put the median probability of AI causing human extinction (defined as the global population falling below 5,000) at 0.0001% by 2030, 0.03% by 2050 and 0.38% by 2100. AI-domain experts gave higher medians: 0.02%, 1.1% and 3% for the same horizons.

Those are useful expressions of informed judgment. They are not measured frequencies — there is no reference class of past AI extinctions to calibrate against — and different studies answer different questions: extinction, widespread disaster, or a lasting loss of human control, over different periods, asked of different groups. Before comparing the answers, compare the questions and assumptions. A widely cited 2016 comparison claiming a typical person would be several times likelier to die in an extinction event than in a car crash rested on an input figure its authors' own erratum had to correct. Comparisons like that are only as credible as the number feeding them.

I am not going to replace a frightening number with a comforting number I cannot justify. Honest uncertainty is more useful than choosing the figure that best matches our mood.

The trap on the reassuring side

There is a mirror-image error: someone shows an AI making a silly mistake and concludes that serious danger is impossible. But a system can be unreliable in one situation and capable in another, and one impressive demo does not establish that a system can independently manage a complex real-world operation. We need evidence about specific capabilities under specific conditions — and humility about predicting when more general intelligence might arrive, a phrase people use to mean very different things.

What actually reduces the risk

The practical answer is unglamorous and already available: permissions and oversight. Drafting an email is different from sending it. Suggesting a software change is different from deploying it. Recommending a payment is different from moving the money. For consequential systems, the questions are: what can the system change, what requires approval, how far can a mistake spread, and can we limit the damage while we investigate?

A human in the loop helps only if that person has a realistic chance to understand what is happening and intervene. If a system makes thousands of complicated decisions while someone watches an unreadable dashboard, calling that person "in control" settles nothing — and an approval button that requires expertise the reviewer does not have offers less protection than it appears to. Useful oversight takes time, information, and genuine authority to stop the action.

Safety research is making measurable progress. The UK AI Security Institute's Frontier AI Trends report found that some safeguards had become substantially harder to bypass, while weaknesses remained in every system it tested — evidence of progress and of unfinished work at the same time. Safety training is not an unbreakable set of robot laws; it needs to be paired with controls around what a system can access and do.

Don't let a headline become your calendar

Confident predictions of disaster can stay with you, because they concern your family, your work, and everything you hope to build. But online, the most alarming version of an argument is easier to encounter than its assumptions. We can listen to serious warnings without turning every possible outcome into a personal countdown. Supporting safety research does not require living as though the conclusion has already been reached.

My position is the one the video ends on: take the risks seriously, ask for clear evidence, and remain willing to update your view. Demand that companies show the authority they give a system is justified, that decision-makers ask what happens when protections fail, and that public debate makes room for changing our minds as evidence changes. Neither promises of abundance nor predictions of disaster should let anyone skip those questions.

Build the business. Make the plans. When a system gives us a worrying direction, check the road before handing over the keys.


This article is the companion to a video on Bert Speaks About AI. The probability figures and evaluation results cited here are drawn from the sources above; the author's working paper on the subject, Artificial Intelligence and Human Extinction Risk (evidence cutoff 28 September 2026), examines them in full.