Anthropic’s AI Fluency Lead Explains the “Discernment Tax”

Kristen Swanson, who leads AI fluency research and learning at Anthropic, has a counterintuitive job description: she spends at least half her time telling people when not to use AI. Speaking about how workers are adapting to generative tools, Swanson said the savviest users are not those who turn to AI for every task, but those who exercise judgment about what to hand off.

Her central warning is what she calls a “discernment tax.” When people delegate the right tasks, the time spent reviewing AI output is worthwhile. But when they ask AI to do work they already know well, evaluating and correcting the output can consume more time than doing the task themselves.

The risk is not only wasted effort. Swanson pointed out that AI can hallucinate on niche topics with limited training data, making delegation risky in areas such as obscure research papers or little-known researchers. She also highlighted a subtler problem: users’ understanding of AI capability can freeze in time, leaving them stuck with older, familiar uses even as models improve.

Anthropic’s Claude Academy, which Swanson oversees, therefore asks users to record some of their hardest tasks and retest them when new models arrive. The goal, she said, is not to maximize prompts or features, but to make judgments about when and how to use AI become second nature.

When Delegating to AI Costs More Than It Saves

Why the “discernment tax” falls hardest on experts

Swanson’s argument is that the cost of reviewing AI output is highest for tasks where a person already has real competence. For tedious data analysis, an AI draft can save effort even after checking. For a social media post on a topic the user knows deeply, the same review process becomes “parsing through all of this stuff.” The distinction is not about the tool’s quality, but about the user’s expertise relative to the task.

Hallucination makes niche delegation a specific risk

The discernment tax is not the only reason to keep some work human. Swanson says AI may hallucinate when asked about a very specialized research paper or researcher because the model saw little on that topic in training. That is a concrete boundary for professionals in research-heavy fields: the less represented a subject is in training data, the more verification is required.

Capability overhang keeps users behind the models

Swanson also describes a problem sometimes called “capability overhang”: a worker’s mental model of what AI can do gets frozen based on earlier experiences. Even as Anthropic and other labs ship more capable models, users may stick to old tasks and miss new uses. Claude Academy’s practice of writing down hard tasks and retesting them after model updates is designed to counter that drift. She argues the useful question is not whether a feature is switched on, but whether a user can say, “I tried this, and it didn’t work, and I’m going to try it a different way next time.”

What Workers and Managers Can Do About the Discernment Tax

  • For a task you already do well, try doing it yourself first; Swanson’s core point is that reviewing AI output on familiar work can create more effort than it saves.
  • Reserve AI for tedious, lower-expertise tasks such as data analysis, where the cost of checking output is outweighed by the time saved.
  • For niche research topics—very specialized papers or little-known researchers—treat AI output as a starting point requiring verification, because hallucination risk rises when training data is thin.
  • Keep a written list of your hardest tasks and retest them when a new model version appears; that counters capability overhang and keeps your sense of what AI can do current.