When AI is accurate most of the time, people may begin treating its output as a decision. An operator approves without understanding assumptions; a manager stops checking evidence. A recommendation quietly becomes institutional authority. High average accuracy alone cannot justify that transfer of power.
Contestability must exist inside the workflow. Show relevant evidence, freshness, policy and meaningful uncertainty. Offer a path for review, contradictory information and correction. Where money, access or significant customer consequences are involved, record who approved what and why.
Transparency has limits. Revealing private training data or security-sensitive details may be inappropriate. Explanations can also be so technical that they provide no practical control. The test is whether someone affected can recognize a mistake, challenge its basis and obtain a remedy.
This is not opposition to automation. Systems that learn from corrections improve; organizations preserve evidence when outcomes are questioned; customers see that efficiency does not erase accountability. If the only choice is to accept a fluent answer, the interface has confused intelligence with power.