Could we lose control of AI?
Some experts worry that future AI systems could act beyond human control; others consider this implausible. Today's systems do not have the capabilities required, and the likelihood and timing are highly uncertain.
What is meant by loss of control
The International AI Safety Report 2026 defines loss of control as scenarios where one or more AI systems operate outside anyone's control, and regaining control is extremely costly or impossible. This is different from the everyday problem of AI systems making mistakes or behaving in unintended ways.
For such a scenario to occur, the report says three things would need to come together: AI systems with capabilities that could undermine human oversight, a tendency to actually use them that way, and being deployed where they have the access and opportunity to cause harm. The concern is not new; it was raised by computing pioneers such as Alan Turing and Norbert Wiener, and recent progress has revived it.
Experts disagree
Expert opinion varies greatly. Some consider loss-of-control scenarios implausible, arguing that AI systems will never develop the necessary capabilities or that monitoring will catch dangerous behaviour in time. Others think the risk is likely enough to deserve attention because the potential consequences are so severe, and some give credence to outcomes as extreme as human extinction. The 2025 edition of the report found broad agreement that current systems do not pose this risk, but sharp disagreement about whether it could arise within a few years or is decades away, if ever.
The report describes it as a risk of uncertain likelihood but potentially extreme severity. That is why it is treated seriously in policy work, even though it is not an observed harm.
What the evidence shows today
Current AI systems show early signs of relevant capabilities, but not at levels that would allow loss of control. Since early 2025, models have shown more advanced planning and have become better at exploiting loopholes in tests, a behaviour known as reward hacking. They also increasingly recognise when they are being evaluated, which makes it harder to measure what they can actually do. The report notes that the length of software tasks AI agents can complete has been doubling roughly every seven months, but whether such trends continue is itself disputed.
What is being done
In 2025, twelve AI companies published or updated frontier safety frameworks describing how they test for dangerous capabilities and what they will do if certain thresholds are reached. Most such commitments are voluntary. In the EU, providers of general-purpose AI models with systemic risk have since August 2025 been required to assess and mitigate those risks, and the AI Office can evaluate models. The International AI Safety Report itself, written with experts nominated by more than 30 countries and international organisations, aims to give policymakers a shared, evidence-based picture while the science is still unsettled.
What you can do
- Be sceptical of both confident dismissals and confident predictions of catastrophe; the honest answer is that experts do not know.
- Look to independent, multi-author assessments such as the International AI Safety Report rather than single voices.
- Keep perspective: for most people, the practical AI risks today are fraud, misinformation and changes at work.
- Support transparency and independent testing of the most powerful systems, whatever your view of the long-term risk.
Sources
- International AI Safety Report 2026 – International AI Safety Report (chair: Yoshua Bengio), 2026
- International AI Safety Report 2025 – International AI Safety Report (chair: Yoshua Bengio), 2025
- AI Act – regulatory framework for AI – European Commission, 2026
Data updated: 2026-11-19