Original article

MIT News: https://news.mit.edu/2026/medical-ai-assistance-benefits-vary-based-on-user-expertise-0804

Introduction

An MIT study found that AI explanations can affect users differently depending on their medical expertise. Non-experts often deferred to convincing LLM explanations even when the advice was wrong, while clinicians were better able to detect incorrect assistance.

Vocabulary

  1. defer to — accept another person’s or system’s judgment.
  2. expertise — advanced knowledge or skill in a field.
  3. explainability — the ability of an AI system to provide understandable reasons for its output.
  4. deference — willingness to rely on another authority’s judgment.
  5. convincing — appearing believable or persuasive.
  6. generic — broad and not specifically tailored.
  7. automation bias — a tendency to trust automated recommendations too much.
  8. anchoring effect — excessive influence from the first information received.
  9. lead someone astray — cause someone to make a wrong judgment.
  10. overreliance — depending on something more than is appropriate.
  11. plausible — apparently reasonable or believable.
  12. liability — something that creates a disadvantage or risk.
  13. hypothesis — a proposed explanation to be tested.
  14. subtle — difficult to notice or distinguish.

Reading comprehension

1. Why did AI explanations affect clinicians and non-experts differently?

Reference answer
Clinicians could use medical knowledge and prior experience to challenge the model’s output. Non-experts had fewer independent signals for evaluating the advice and were therefore more likely to use the explanation itself as evidence that the AI was correct.

2. Why were non-experts sometimes more confident even when their AI-assisted answers were wrong?

Reference answer
LLM explanations could sound coherent and convincing even when the underlying diagnosis was incorrect. This increased trust in the recommendation without necessarily increasing its accuracy.

3. What is algorithmic deference?

Reference answer
It is the tendency to give excessive weight to an algorithm’s recommendation instead of independently evaluating whether the recommendation is supported by the available evidence.

4. Why might forming an independent hypothesis before seeing AI advice reduce automation bias?

Reference answer
Making an initial judgment forces the user to process the evidence independently. The AI recommendation then becomes something to compare against rather than the anchor that determines the user’s first interpretation.

5. Why is a one-size-fits-all approach unsuitable for explainable AI?

Reference answer
The same explanation can help an expert while misleading a beginner. AI interfaces therefore need to account for the user’s knowledge level and encourage appropriate scrutiny rather than automatic trust.