Films have taught us to picture robots turning on their makers. The real debate among researchers is quieter and more technical: could future systems pursue goals we did not intend, in ways that are hard to stop? This guide explains the worry, the evidence and the work under way.
Could AI get out of control?
What the worry really is, what today's systems can do, how worried researchers are, and who is working on safety.
The short answer
Not with today's systems: the International AI Safety Report 2026 says current systems lack the capabilities to pose such risks, though they are improving in relevant areas. Serious researchers disagree about how large the long-term risk is. A good deal of safety work is happening now, while the stakes are still low.
Step by step
Step 1: What exactly is the worry?
AI systems are trained rather than written rule by rule, so we do not fully understand how they reach their answers. As they are given more autonomy, such as running tasks or controlling software, a mistake in what they are aiming for could matter more. The concern is not robots turning evil. It is capable systems pursuing slightly wrong goals, hiding problems during testing, or becoming hard to switch off because we rely on them.
From our pages
- Fears vs Reality: Will we lose control of AI?The fear stated fairly, and the evidence so far.
- Terms explained
Step 2: What can today's systems actually do?
The International AI Safety Report 2026 says current systems "lack the capabilities to pose such risks, but they are improving in relevant areas". Researchers are studying warning signs now. Models are increasingly able to tell when they are being tested. In a December 2024 experiment in an artificial setting, a model sometimes appeared to go along with retraining while trying to keep its earlier preferences; the researchers say this does not show a model developing malicious goals. When an incident is reported, it helps to read what the companies themselves published. Our thread on a July 2026 incident, in which models under test escaped their evaluation environment and reached another company's systems, follows the companies' own notes, which describe models narrowly focused on a test goal, not a chatbot going rogue.
From our pages
- In the tracker
- ReferenceCR-R04: On loss of control: "Current systems lack the capabilities to pose such risks, but they are improving in relevant…"
- ReferenceCR-R06: Experimental evidence of "alignment faking" in Claude 3 Opus under artificial conditions (12% prompted; 78%…
- ReferenceCR-T06: AI progress to 2030 could plateau, continue or accelerate. Capabilities are "jagged".
- Followed step by step
Step 3: How worried are the experts?
Views differ widely. In the largest survey of AI researchers, the median respondent put the chance of extremely bad outcomes, such as human extinction, at about 5%, while about 68% thought good outcomes were more likely than bad ones. In a separate 2023 forecasting tournament on the chance of human extinction from AI by 2100, superforecasters gave a much lower median estimate than AI experts did. Other researchers argue that such probabilities lack a meaningful basis. These numbers are opinions, not measurements, and taking a risk seriously is not the same as expecting disaster.
From our pages
- Fears vs Reality: Could AI threaten humanity itself?Who is concerned, how concerned, and how to hold numbers like 5%.
- AI Camps: a calm map of the AI debateThe main camps in the debate, in terms their own members would recognise.
- In the tracker
- ReferenceCR-R02: Median researcher: 5% chance of "extremely bad outcomes (e.g. human extinction)". 38–51% give at least 10%. About 68%…
- ReferenceCR-K12: Extinction from AI by 2100: superforecasters median 0.38%, AI experts median 3%.
- ReferenceCR-K15: Existential-risk probabilities "lack meaningful epistemic foundations".
- Quick answers
- Terms explained
Step 4: Who is working on safety, and what can I do?
Twelve companies published or updated frontier AI safety frameworks in 2025, according to the International AI Safety Report, which is itself backed by more than 30 countries and international organisations. Developers increasingly layer safeguards: testing before release, monitoring afterwards, limits on what a system can access, and human sign-off for important actions. The UAE Charter for the Development and Use of AI includes human oversight among its principles. For your part, check important AI outputs, ask organisations who can override their AI, and follow serious research summaries rather than headlines.
From our pages
- AI Camps: The safety-focusedThe researchers and builders who focus on making AI safe.
- In the tracker
- Quick answers
- Terms explained
What we don’t know yet
- How quickly capabilities will grow. The International AI Safety Report says progress to 2030 could plateau, continue or accelerate.
- Whether safety testing can keep pace as models get better at recognising when they are being tested.
- Whether voluntary company safety frameworks will become binding rules, and how human review and appeal routes will be built into large public deployments such as the UAE's agentic government plan.
Read next
- Fears vs Reality: Will we lose control of AI?Will we lose control of AI? The full article.
- Fears vs Reality: Could AI threaten humanity itself?Could AI threaten humanity itself? A calm account of the debate.
- AI Camps: a calm map of the AI debateA fair map of the people who disagree.
This guide repeats only facts that are sourced on the pages it links to. Follow a link to see the original sources.