Evaluating AI-Assisted Work - Cognitive Load and Over-Reliance

Users are surprisingly bad at judging how much an AI tool helped them. So if you’re rolling out an assistant (or deciding whether to keep one), measure qualitatively and quantitatively, and design so the human stays the final judge of quality.

How to measure an AI tool

• Cognitive load: NASA TLX. A widely used post-task survey rating mental, physical and temporal demand, performance, effort and frustration on a 0–100 scale.
• Usability: did it make people feel more capable, or more confused?
• Output quality: did the work actually get better, judged independently?
• Observed behaviour: record sessions and watch what people really do, and whether they keep using it.

Dunning-Kruger, with an AI in the room

The Dunning-Kruger effect: people with less skill tend to overestimate their performance. The ChainBuddy study (ChainBuddy - Two-Phase Agents and Model Tiering) found something adjacent: AI assistance improved participants’ work, but they didn’t rate it as better. Self-assessment is unreliable in both directions, which is why surveys alone aren’t enough. It’s also why a single number makes a poor target (Goodhart’s Law).

The risks

1. Over-reliance. If users can’t judge output quality, they accept sub-par solutions.
2. Bias and deskilling. People keep following the AI’s patterns even when working without it.

Design responses

• Produce an editable starter rather than a finished answer; the user is the final arbiter.
• Avoid defaults that let people skip the thinking (ChainBuddy’s intake form asks questions with no pre-filled answers).
• Pair AI output with deterministic verification rather than trusting a read-through: Don’t Read the Code.