AI fluency includes knowing when not to delegate
Anthropic’s AI-fluency lead calls attention to a “discernment tax”: sometimes checking generated work costs more than doing the task directly.
What happened
In a new interview, Anthropic’s Kristen Swanson argued that capable users distinguish tasks worth delegating from tasks where review burden erases the saving.
Why it matters
Learning programmes should assess task choice and verification effort, not maximise tool usage or prompt volume.
The next useful AI skill may be restraint. In a Business Insider interview, Anthropic’s head of AI fluency, Kristen Swanson, describes a “discernment tax”: the effort required to decide whether generated work is reliable and suitable. For some tasks, checking the output can take longer than doing the work directly.
That is a practitioner judgement, not a measured productivity coefficient. The article does not provide a controlled comparison of task time, quality or error rates. It does, however, sharpen a neglected learning objective. Many AI programmes teach access, prompting and iteration; fewer require learners to estimate the cost of verification before delegating.
Anthropic’s separate AI Fluency Index gives the idea a broader behavioural frame. It analyses conversation patterns and distinguishes how people direct, describe, discern and delegate. The index remains first-party research based on use of Anthropic’s own system, so it should not be treated as a universal proficiency scale.
A better practice exercise
Give learners three real tasks: one repetitive and checkable, one ambiguous but reversible, and one high-stakes or dependent on tacit context. Ask them to choose whether to delegate, state the checking plan, and compare total effort and quality with an unaided route. Reward a justified “do it yourself” decision when it is cheaper or safer.
The management implication is equally plain: adoption rates are not capability rates. A team that uses AI less often may be exercising better judgement, while a high-use team may be accumulating hidden review work. Track rework, verification time and escaped errors alongside usage.