Bill Gates says expect a lot of negative AI surprises in cyber over the next year
Bill Gates warns that unregulated AI will drive an escalation in AI-powered cyberattacks, predicting a lot of negative surprises in the cyber area over the next year and dismissing self-regulation by AI companies as sufficient.
Bill Gates has a forecast, and it isn’t about what artificial intelligence will make possible. He warned that unregulated artificial intelligence poses a growing threat and predicted an escalation in AI-powered cyberattacks in the coming year — a forecast premised on what happens when bad people use AI without safeguards. His exact words: “over the next year, primarily in the cyber area, we are going to have a lot of negative surprises.”
An interview on CNN’s State of the Union added a superlative to the forecast: artificial intelligence is “the most dangerous technology humans ever invented,” as he put it. And on the question of who keeps a technology like that in check, he dismissed the idea that self-regulation by AI companies is sufficient.
A forecast of surprises is a strange genre. Security warnings are useful to the degree that they name their object — this vulnerability, that technique — so the people on the defensive side can point resources at something. Gates’s version does the opposite: it supplies a domain (“primarily in the cyber area”) and a clock (“over the next year”) and leaves the content blank, on the logic that the content is precisely the part nobody will see coming. Read one way, that’s a refusal to speculate; read another, it’s a judgment that speculation wouldn’t help. His phrasing suggests the second: a surprise that could have been named in advance wouldn’t be one.
The territory isn’t incidental. Most of the harms people project onto AI need a bridge between the model and the world — a laboratory, a delivery system, a distribution channel. Cyber operations barely need one: the raw material is text and code, produced quickly and at scale, which is close to a description of what the current models already do for a living. The premise of his warning — bad people using AI without safeguards — is a statement about the gap between what the models can do and what actually constrains them: the gap AI safety exists to close.
The dismissal of self-regulation may be the line with the longest shelf life. “Trust the builders” has been the default answer to warnings of this kind — the assurance that the people best placed to understand the systems are also the ones keeping them safe. Gates is not a critic the industry can wave off as an outsider, and his verdict — insufficient — forecloses the most convenient rebuttal that position has: that the alarm comes from people who don’t understand the technology. What’s left is the question the argument actually turns on — not whether these systems are powerful, but who answers when the power gets used.
There is a reason forecasters usually avoid this construction. “A lot” is not a number, “negative” is not a category, and “surprise” is a promise that the specifics will arrive unannounced — which makes the prediction nearly impossible to grade, and safe from ever being wrong in a way that sticks. What isn’t negotiable is the clock. A year out, the surprises will have arrived or they won’t have, and the difference will be legible in a way decade-scale prophecies never are. The warning is both more and less than the genre usually offers — more, because the territory and the deadline are named; less, because by his own framing, what actually shows up will be the part nobody saw coming.