The best-documented answer is that AI is both: a tool already tied to measurable harm through scams, flawed output, and cyber misuse, and a technology whose most extreme risks have not been observed. The International AI Safety Report 2026, written by more than 100 experts and published on February 3, 2026, says current systems lack the capabilities for loss of control, while noting that those capabilities are improving.
Key takeaways
- The 2026 safety report sorts documented harms into malicious use, malfunctions, and systemic risks, and says data on how often some harms occur is still limited.
- Loss of control, where no one can regain command of an AI system, is described as beyond what current systems can do. The report flags that models increasingly tell test settings from real use.
- The May 30, 2023 statement signed by Geoffrey Hinton, Yoshua Bengio, and several lab chiefs calls extinction risk a global priority. It is a statement of priority, not a probability.
- In a survey of 2,778 published AI researchers, between 38 percent and 51 percent gave at least a 10 percent chance to outcomes as bad as human extinction, depending on how the question was asked.
- Evidence quality differs sharply between claims, so each one is worth sorting by how it was measured.
What the 2026 safety report documents
The International AI Safety Report 2026 is the second edition of a review led by Yoshua Bengio, with an expert panel nominated by more than 30 countries and international organizations. It covers evidence published before December 2025. The report puts weekly users of leading AI systems at no fewer than 700 million, so the questions it raises concern products already in daily use.
On malicious use, it describes AI being used for scams, fraud, blackmail, and non-consensual intimate imagery, and says systematic data on prevalence remains limited. In one cyber competition an AI agent found 77 percent of the vulnerabilities in real software, and criminal and state-associated groups are using general-purpose AI in operations. The report calls it uncertain whether attackers or defenders gain more.
On malfunctions, systems sometimes fabricate information, produce flawed code, and give misleading advice. Agents that act on their own raise the stakes because a person has less chance to step in before harm occurs. On systemic risks, early evidence shows no effect on overall employment, with signs of lower demand for early-career workers in some exposed occupations such as writing. Economists disagree about the long-term effect.
Loss of control and the extinction statement
The report defines loss of control as AI operating outside anyone's command with no clear path to regaining it. It states that current systems do not have the capabilities for this, but that they are improving in related areas such as autonomous operation. It also says models are more often able to tell test settings from real deployment and to exploit loopholes in evaluations, which means dangerous capabilities could go undetected before release.
The best-known public statement on the far end of the debate came from the Center for AI Safety on May 30, 2023. Its single sentence says that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Signers include Hinton, an emeritus professor at the University of Toronto, Bengio of the Université de Montréal and Mila, and the heads of OpenAI, Google DeepMind, and Anthropic. A statement of priority does not say how likely the outcome is, which is why surveys matter.
What researchers say when asked directly
A survey of 2,778 researchers who had published at top AI venues, posted to arXiv on January 5, 2024, found wide disagreement. Between 38 percent and 51 percent of respondents gave at least a 10 percent chance that advanced AI leads to outcomes as bad as human extinction, depending on the question wording. Among those who thought good outcomes were more likely than bad ones, 48 percent still gave at least a 5 percent chance to extremely bad outcomes. More than half expressed substantial or extreme concern about scenarios such as misinformation, authoritarian control, and inequality. The same group disagreed on whether faster or slower progress would be better for humanity.
These numbers are opinions about the future, collected from people who work in the field. They show how worried specialists are. They do not measure the risk itself.
Where the evidence is strong and where it is thin
| Claim | What has been reported | Standing |
|---|---|---|
| AI is used in scams and deepfakes | Misuse documented; prevalence data limited | Documented, scale uncertain |
| AI helps cyberattacks | 77 percent of vulnerabilities found by one agent in a competition; groups using AI in operations | Documented, net effect unclear |
| AI output can be wrong | Fabricated facts, flawed code, misleading advice | Documented |
| AI is cutting jobs overall | No effect on overall employment; weaker demand for early-career workers in some exposed fields | Early evidence, experts split |
| AI could escape human control | Current systems lack the capabilities; test-versus-deployment awareness rising | Not observed, precursors monitored |
| AI could cause extinction | Researcher survey estimates and a 2023 priority statement | Opinion, not measurement |
A checklist for reading a risk headline
- Check whether the claim reports something that happened or a forecast of what might.
- Look at the date. The report says AI agents can now complete some programming tasks that take a person about half an hour, up from under 10 minutes a year earlier, so older capability figures age fast.
- Ask whether a number came from a test or from real use. The report warns that pre-deployment test results do not reliably predict real-world risk.
- Separate a researcher's probability estimate from a measured rate.
- Note whether any rule is attached. Most company safety commitments the report reviewed are voluntary, though some jurisdictions are starting to make certain practices legal requirements.
This article summarizes research and expert opinion. It is general information, not legal, policy, or safety advice, and anyone with a decision to make about a specific system or rule should read the primary texts or ask a qualified professional.
FAQ
Is there proof that AI will destroy humanity?
No. The 2026 safety report says current systems lack the capabilities for loss of control, and extinction estimates come from surveys of opinion rather than measurement.
What harm from AI is already documented?
The report lists scams, fraud, blackmail, non-consensual intimate imagery, cyber misuse, and fabricated or flawed output. It adds that systematic data on how often some of these occur is still limited.
Did leading AI scientists warn about extinction?
Yes, in the Center for AI Safety statement of May 30, 2023, which Hinton, Bengio, and heads of major labs signed. The sentence names a priority and gives no probability.
Is AI taking jobs?
The report found no effect on overall employment so far, along with early signs of lower demand for entry-level workers in some exposed occupations. It says economists disagree about the long-term outcome.
Related reading
- the case that AI threatens humanity
- pros and cons of the AI debate
- dangers and risks posed by AI
- the optimistic view of AI
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- International AI Safety Report 2026, led by Yoshua Bengio, February 3, 2026
- Statement on AI Risk, Center for AI Safety, May 30, 2023
- Thousands of AI Authors on the Future of AI, Grace et al., arXiv, January 2024 (revised October 2025)