Differentiating between GPT-generated and human-written feedback for radiology residents
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
PURPOSE: Recent competency-based medical education (CBME) implementation within Canadian radiology programs has required faculty to conduct more assessments. The rise of narrative feedback in CBME, coinciding with the rise of large language models (LLMs), raises questions about the potential of these models to generate informative comments matching human experts and associated challenges. This study compares human-written feedback to GPT-3.5-generated feedback for radiology residents, and how well raters can differentiate between these sources. METHODS: Assessments were completed by 28 faculty members for 10 residents within a Canadian Diagnostic Radiology program (2019-2023). Comments were extracted from Elentra, de-identified, and parsed into sentences, of which 110 were randomly selected for analysis. 11 of these comments were entered into GPT-3.5, generating 110 synthetic comments that were mixed with actual comments. Two faculty raters and GPT-3.5 read each comment to predict whether it was human-written or GPT-generated. RESULTS: Actual comments from humans were often longer and more specific than synthetic comments, especially when describing clinical procedures and patient interactions. Source differentiation was more difficult when both feedback types were similarly vague. Low agreement (k=-0.237) between responses provided by GPT-3.5 and humans was observed. Human raters were also more accurate (80.5 %) at identifying actual and synthetic comments than GPT-3.5 (50 %). CONCLUSION: Currently, GPT-3.5 cannot match human experts in delivering specific, nuanced feedback for radiology residents. Compared to humans, GPT-3.5 also performs worse in distinguishing between actual and synthetic comments. These insights could guide the development of more sophisticated algorithms to produce higher-quality feedback, supporting faculty development.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.005 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it