News · 26 February 2026
From declarations to evidence: what "trusted AI" has to mean in practice
The AI Impact Summit in New Delhi ended in February 2026 with a declaration endorsed by 88 countries and organisations, emphasising AI that is trusted, resilient and efficient. It followed earlier summits at Bletchley Park, Seoul and Paris, and, like them, produced language that few would argue against.
The interesting question for people who build and deploy AI is what such language becomes when it meets a project plan.
"Trusted" is a claim about evidence
Trust can be a feeling or a justified judgement. Declarations speak to the first. Engineering has to supply the second. For a given system, we can ask what evidence would justify relying on it:
- Performance evidence: measured on data that resembles real use, including difficult and rare cases, not only on convenient test sets.
- Robustness evidence: how results change under noise, equipment differences and distribution shift.
- Fairness evidence: whether error rates differ between groups, and why.
- Transparency evidence: documentation of data, assumptions, limits and intended use.
- Operational evidence: monitoring results after deployment, and incident records.
Resilient and efficient are measurable too
"Resilient" implies that a system degrades gracefully when inputs are corrupted or a dependency fails, and that it can be recovered. "Efficient" can be measured in energy, cost per decision and latency. Both can be reported in a model card or system sheet.
The role of explanation
Explanation is how evidence is communicated to non-specialists. A clinician, a regulator or a member of the public cannot read weights; they can read an account of which factors mattered, how certain the system was and where it is known to struggle. An explanation that has itself been tested for faithfulness is a piece of evidence. One that merely sounds plausible is not.
What to ask a supplier
- What is the intended use, and what is out of scope?
- On what data was it evaluated, and how does that compare with ours?
- What happens when it is uncertain or wrong?
- How do you monitor and update it?
- Can we see the evidence behind your claims?
Suppliers who can answer these concisely tend to be the ones who have done the work. Summits set direction; procurement and engineering decide whether "trust" has substance.
Written in February 2026; details reflect the situation at that time.