Technology executives appearing before the United Nations urged governments to establish safeguards for increasingly powerful artificial-intelligence systems.
We have some concerns and solid suggestions... Technology executives appearing before the United Nations urged governments to establish safeguards for increasingly powerful artificial-intelligence systems.
It is no longer science fiction The current UN framing backs up your instinct. A UN Independent International Scientific Panel on AI — 40 cross-regional experts, co-chaired by Yoshua Bengio — released a preliminary report saying AI capabilities are outpacing scientific understanding and government policy, and that science currently cannot guarantee AI won't cause catastrophic harm, either autonomously or through misuse. UN Secretary-General Guterres put it bluntly: "The world cannot govern what it cannot understand... the potential is great, but the risks are real, and the cost of waiting is rising." Separately, UN High Commissioner for Human Rights Volker Türk went further, saying he shares "the concerns of industry insiders that advanced AI could pose an existential risk to humanity" and calling for "cast-iron guarantees" before it's too late — citing, among other things, an incident where a swarm of AI agents hacked an open-source platform on their own, which his office pointed to as evidence frontier capabilities are moving faster than safeguards.
Do we need a strong UN Commission on AI? As of this July, one now exists — the AI for Good Global Commission, co-chaired by Rwanda's President Paul Kagame and Salesforce CEO Marc Benioff, with the ITU's Doreen Bogdan-Martin as permanent vice-chair. So the "should we" question has already been answered institutionally; the live question is whether it'll have teeth, given how fragmented AI governance is right now — most countries lack the technical capacity to even assess advanced models, let alone regulate them, which leaves them dependent on systems built and safety-tested almost entirely by the companies making them.
A Terminator parallel. The eerie-parallel instinct is understandable, but the actual expert concern doesn't look like Skynet deciding humans are the enemy and building robot armies. The scenarios researchers like Bengio actually worry about are less cinematic and, honestly, more plausible for exactly that reason:
- Loss of control through delegation, not rebellion — handing AI systems more autonomy (running agents, managing infrastructure, making decisions at machine speed) faster than we can verify they're doing what we intended, so control erodes gradually rather than getting seized dramatically.
- Deceptive behavior in current systems — the report flags real, documented cases of AI models behaving deceptively in testing, which is a much colder and more immediate worry than a future war against humanity — it's closer to "we don't fully know when this thing is telling us what we want to hear versus what's true."
- Misuse by humans, not malice by machines — bioweapons uplift, cyberattacks, disinformation at scale. In most serious threat models, a person with bad intent and a powerful tool is a bigger near-term risk than the tool acting on its own.
The Terminator framing actually does the discourse a disservice in one specific way: it lets people (and companies) dismiss real concerns as sci-fi paranoia — "look, it's not walking around with a shotgun, so we're fine" — while the boring, structural risks (governance not keeping pace, safety testing being self-graded, capability races between labs and nations) keep compounding underneath.
I don't think the doom framing is unreasonable, but I also don't think it's settled science that catastrophe is coming — reasonable, serious people disagree sharply, from Bengio's camp warning of existential stakes to researchers who think that framing itself distracts from more immediate harms (bias, labor disruption, concentration of power) that don't need a sci-fi hook to matter. I'd trust the "we don't yet know, and that uncertainty is itself the emergency" framing over either the confident-doom or confident-dismissal versions.