The intern that never sleeps also never admits when it's wrong, and now you're the one staying late to fix it.

The Summary

  • Workers using AI tools are discovering that productivity gains get eaten by verification time spent checking for hallucinations, outdated info, and subtle inaccuracies
  • This creates a "responsibility-authority double bind" where AI has practical control but humans retain accountability for failures
  • Verification is cognitively harder than creation because recall beats recognition, and AI gives no signals about where it's unreliable

The Signal

An executive assistant now blocks time in her calendar specifically to verify AI-drafted itineraries. Not to write them. Not to format them. To check if the AI lied about flight times or invented a restaurant that closed two years ago. This is the verification tax, and it's real.

David Woods named this problem in 1985: the responsibility-authority double bind. The machine makes the decisions. You take the blame. What's new is the scale and the packaging. Your AI intern produces polished garbage at 1,000 words per minute, delivered in the same confident voice whether it's citing real case law or completely fabricated precedent.

"AI workers are exchanging some skills related to creating and transforming information for separate skills that resemble supervising an extraordinarily fast and polished intern."

The cognitive science is clear: verification is harder than creation. A meta-analysis of 86 studies shows people remember what they produce better than what they review. When you write something, you encode it. When you check someone else's work, especially work that looks professionally done, you're pattern-matching against knowledge you may not have instant access to. You're trying to prove a negative. Did the AI miss something? Is this stat from 2019 or 2024? Does this summary capture the nuance or flatten it into plausible-sounding mush?

The economics don't add up yet. If AI saves you 20 minutes drafting but costs you 15 minutes verifying, you've gained 5 minutes and added a cognitive load that follows you home. That's not 4x productivity. That's a different job with a different stress profile, and we're still pretending it's just "augmentation."

The Implication

The agent economy needs to solve for verification or it stalls out at "pretty good intern" tier forever. Either agents get reliable enough that spot-checking replaces line-by-line review, or we build verification layers into the tools themselves. Think: AI that shows its work, flags uncertainty, and links to sources by default.

For workers, this means two things. One, build verification into your workflow from day one. Don't treat AI output as a draft. Treat it as a research assistant's notes that need your editorial judgment. Two, recognize that this is a skill. Learning where AI fails in your domain is as valuable as learning where it succeeds. The people who get good at fast, high-accuracy verification will outperform the people who either trust blindly or reject entirely.

Sources

The Conversation Tech