Artificial intelligence and technologies
In Depth
How powerful can artificial intelligence become?
Capabilities are increasing rapidly, but the future trajectory remains deeply uncertain.
- Last updated:
- Reading time:
- 8 min
In Brief
- The International AI Safety Report 2026 is a scientific assessment produced by more than 100 experts and backed by more than 30 countries and international organizations.
- The capabilities of general-purpose systems have improved rapidly, particularly in mathematics, programming, science and the execution of more autonomous tasks.
- The profile of these capabilities remains uneven, however: the same systems can solve difficult problems and fail at apparently simple tasks.
- Between now and 2030, very different scenarios remain plausible: slowdown, continuation of the current pace, or acceleration. No arrival date for artificial general intelligence can be considered certain.
What we know
Documented facts
- The report documents measurable improvements in the capabilities of general-purpose models, especially in mathematical reasoning, coding, scientific tasks and operations with greater autonomy.
- Performance is described as uneven: excellence at some complex tasks coexists with errors on tasks trivial for a person.
- Some risks are already observed: malicious uses, malfunctions, cyber applications, manipulation and disinformation.
- Risk-management techniques are improving, but no combination of safeguards is perfectly reliable; there is also a gap between test results and system behavior in the real world.
- The same systems produce concrete benefits in scientific research, medicine, education, productivity and environmental management.
How sure are we?
Evolving evidence
Current knowledge points in a direction, but significant uncertainties remain.
The evidence is stronger on already observed risks and current capabilities; it is much more uncertain regarding future capabilities and scenarios such as loss of control. The two should not be confused.
How it can be read
Interpretations
- Uncertainty about the future trajectory is not an argument for delaying decisions: it is the reason robust evaluation capabilities are needed already today.
- Presenting artificial intelligence only as a threat, or only as an opportunity, prevents it from being governed.
What we don't know
- We do not know whether the current pace of improvement will continue, slow down or accelerate.
- We cannot reliably quantify the risks linked to future capabilities not yet observed, including in the biological and chemical domains.
- We lack evaluation methods capable of accurately predicting system behavior outside test settings.
Why it matters
- Decisions on rules, checks and responsibilities are being made now, while uncertainty is at its greatest.
- Systemic effects — on work, information, security — appear before the consequences are fully understood.
- The same technologies that increase some risks can help reduce others, from medical research to environmental monitoring.
What we can do
- Demand independent evaluations of the capabilities and risks of the most advanced systems, before and after release.
- Distinguish, in public debate, between already documented risks and uncertain future scenarios.
- Support transparency on the data, limits and incidents of systems in use.
The No-Extinction proposal
Proposals
- No-Extinction principle: the greater a technology's capability, the greater the collective capacity to assess and control it must be. Verification must grow at least as fast as power.
Sources and data
The links lead to the original documents and open in a new tab. Every figure cited in this article comes from these sources.
International AI Safety Report
International AI Safety Report 2026
international scientific report · 2026
internationalaisafetyreport.org/publication/international-ai-safety-report-2026(opens in a new tab)International AI Safety Report
Extended Summary for Policymakers 2026
summary for policymakers · 2026
internationalaisafetyreport.org/publication/2026-report-extended-summary-policymakers(opens in a new tab)