A randomised trial involving 123 Chinese trainee radiologists: when GPT-5.4 gets it wrong, incorporating the opinion of a second model almost doubles the accuracy of the readings, particularly amongst trainee radiologists.
A Chinese team conducted a prospective randomized trial in three hospitals from July to September 2026 with 132 interns with less than three years of experience (123 analyzed). Each one interpreted 60 X-rays, first alone and then with AI help: either GPT-5.4 alone, or GPT-5.4 accompanied by a second opinion issued by Kimi-K2.6 or Gemini-3.6 Flash.
The second opinion as a safeguard
Among radiology interns, the double opinion increased accuracy by about 6.7 to 7.9 points more than the single opinion; no significant difference appeared in interns in other specialties. The clearest effect was in cases where GPT-5.4 was wrong: accuracy increased from about 20% with one opinion to 40% with two in radiologists-in-training, and from 12% to 31% in others.
For the authors, comparing two suggestions would limit the ripple effect of an erroneous AI — a well-documented risk among novice practitioners — with a benefit that depends on the reader's specialty.
Read the full article on the original source
Go to the source →