Knowing what to trust is an AI skill.

Confidence needs a reference point.

An answer can be clear, well organised, and wrong. A table can look complete because missing information has been quietly filled in. A source can look credible until you open it.

For me, treating AI as a superpower means developing the judgment to use its output well. Generating possibilities is one capability. Deciding which of them deserves to influence the next step is another.

NIST’s Generative AI Profile identifies confabulation as a risk: systems can present false or erroneous material confidently. The practical implication is straightforward. Fluency alone is not a test of accuracy.

Capability changes from task to task.

Research with 758 consultants, described by the Harvard Business School AI Institute, found an uneven boundary of AI capability. AI improved performance on some tasks, while other tasks remained outside its effective reach. The researchers called this a jagged technological frontier.

I take that as a reason to evaluate the task in front of me. A good result on a writing exercise does not establish that the same system can verify a statistic, interpret a complicated document, or make the next decision correctly. Trust needs to attach to the evidence for this particular output.

Match the check to the claim.

For a factual claim, inspect the source. Open the cited material. Check whether it actually supports the statement, whether the date matters, and whether the answer has overstated a narrow finding.

For a transformation, compare against the input. When turning notes into tasks, check every owner, deadline, and commitment. In a summary, check what was omitted as well as what was added. The meeting-note example shows how to keep unknown information visible.

For an action, inspect the outcome. A system saying it completed a task is different from evidence that the requested change exists. Anthropic’s guide to agent evaluations separates an agent’s transcript from the state it leaves behind. That distinction matters whenever AI acts through tools.

For a creative choice, define the criteria. A headline or product idea rarely has one correct answer. It can still be evaluated against the intended audience, the constraints, the tone, and the purpose. Explain those criteria before asking for another revision.

Make uncertainty visible.

“Not provided” can be a better result than a plausible guess. When information is missing, ask the system to identify the gap and say what would resolve it. Then check that it followed the instruction.

Another model can help identify issues, but agreement is not independent proof. Where an answer matters, return to source material, a direct test, or a person who can evaluate it.

Keep a few difficult examples alongside the easy ones: incomplete notes, conflicting statements, and requests that cannot be answered from the supplied material. A dependable approach needs to recognise those cases too.