AI risk in 2026: evidence is more useful than panic

The public conversation about AI risk often jumps between two unhelpful positions: nothing has changed, or everything is about to collapse. The evidence supports a more demanding view. Some harms are already documented, some failures become more likely as systems gain autonomy, and some high-impact scenarios remain uncertain.

The International AI Safety Report 2026 was written with guidance from more than 100 independent experts nominated by over 30 countries and international organisations. It groups general-purpose AI risks into three broad areas: malicious use, malfunctions, and systemic risks.

That separation is useful because each category needs a different response.

Three kinds of failure

Malicious use begins with a person deliberately using a system for harm. The report notes documented use in scams, fraud, blackmail, and non-consensual intimate imagery, while also pointing out that consistent data about scale remains limited. Access controls, abuse monitoring, provenance, and law enforcement matter here.

Malfunctions happen when a system pursues the wrong interpretation of a task, produces unreliable output, or behaves unexpectedly in a new setting. Better evaluations help, but product design matters just as much. A fallible model should not receive silent, irreversible authority.

Systemic risks emerge when many organisations adopt the same technology. Labour disruption, concentration of power, dependence on a small number of providers, and widespread information pollution cannot be solved by tuning one model.

Uncertainty is part of the evidence

The report is careful about disagreement. Experts differ on how quickly capabilities will improve, how severe future risks could become, and whether present safeguards will hold. That uncertainty is not a reason to ignore the problem. It is a reason to avoid pretending that one forecast is settled fact.

A practical risk register should label what is observed, what is inferred, and what remains speculative. It should also record how quickly a control can be strengthened if the evidence changes.

Start with the system people actually use

NIST's Generative AI Profile extends its AI Risk Management Framework with concrete considerations for generative systems. Its govern, map, measure, and manage structure is more useful than a vague promise to be responsible.

For a real product, I would ask:

• What data can the model see, retain, or reveal?
• Which actions are reversible?
• How do we measure confident but incorrect output?
• Who notices when behaviour changes after a model update?
• What happens when the model, tool, or network is unavailable?

AI risk management should make a product more dependable, not bury it in ceremony. The aim is to connect a specific capability to a specific harm, then place a control where it can actually interrupt that path.