What are the chances AI stays aligned?

At what point do we lose control? I'm not sure. Nobody is. Anyone who gives you a confident number is selling something.

What we can do is look at what the experts say, see where they disagree, and name what would move the odds.

01What the experts say

The biggest survey of AI researchers (2,778 respondents, 2023) found:

AI researchers on the risk (2023)

68.3%thought good outcomes from superhuman AI are more likely than bad ones.
38–51%gave at least a 10% chance of outcomes as bad as human extinction, depending on how the question was worded.
Mostagreed research on reducing AI risk should get more priority than it does today.

So the typical expert is an optimist who still sees a serious tail risk. That's not a contradiction. A surgeon can expect an operation to go well and still insist on a sterile room.

In January 2025, the International AI Safety Report, written by 96 experts, backed by 30 countries and chaired by Yoshua Bengio, put it this way:

There's broad agreement that today's general-purpose AI can't pose a loss-of-control risk.
On the next several years, views differ widely. Some experts think loss of control is implausible, some think it's likely, and some see it as a modest-likelihood risk that deserves attention because the impact would be so severe.

02What pushes the odds up

Things that make alignment more likely

Interpretability worksIf we can reliably read a model's goals, hidden misalignment has nowhere to hide.
Control scalesTrusted, simpler AIs can keep policing smarter ones (see layer 4).
Mutual interestAI that benefits from being part of the human system has reasons to stay in it.
TimeSlower, staged releases give safety work time to catch up.

03What pushes the odds down

Things that make alignment less likely

A raceLabs and countries competing cut safety corners to ship first.
Easy to buildWhen anyone can build a powerful model, nobody can check them all.
Self-improvementAI that improves AI can outrun the people supervising it.
DeceptionA model that hides its goals during testing passes every test.

04My take

Will every AI behave? Probably not. Rogues will come from somewhere. But it's the same with humans: we live with the good and the bad, and we build police, courts and laws to handle the bad.

So the real question isn't "will all AI stay aligned?" It's "will the aligned AI, and the humans working with it, stay stronger than the rogue ones?" That's a question about systems, not just about models. It's the subject of Rogue AI.

05The short version

No one knows the odds. Expert surveys show optimism with a serious tail risk.
The 2025 international report: today's AI can't escape control; the next few years are disputed.
The likeliest future isn't zero rogues. It's whether the good side stays stronger.

06Sources

Grace, K. et al. (2024). Thousands of AI Authors on the Future of AI. arXiv:2401.02843.
Bengio, Y. et al. (2025). International AI Safety Report. 29 January 2025.

All posts

Appointment setting3

Copywriting3

AI services2

Deep Think AI13

Get the next markup by email.

One real piece of copy per issue, rewritten line by line: a cold email, a web page or a newsletter.