The task assigned to the model
Analyse an already drafted recommendation and flag criteria that may require human review. The model does not write the initial medical recommendation and does not automatically modify the document.
A concrete AI use case in occupational health
A language model used as a second reader to flag possible quality defects in occupational-health recommendations, without drafting or deciding in place of the occupational physician.
In 30 seconds
Essential points
- The model acted as a second reader using five explicit quality criteria.
- It flagged wording for review; interpretation of the job and the final decision remained medical responsibilities.
- The results suggest potential value as a supervised assistant, while over-detection requires systematic human validation.
The professional problem
Reviewing recommendations before they are communicated
An occupational-health recommendation must be clear, applicable, proportionate and compatible with medical confidentiality. Ambiguous or excessive wording may create difficulties for the worker, employer and physician.
The responsibility that remains human
The occupational physician remains responsible for final wording, interpretation of the real job, the clinical context and the proportionality of any recommendation.
Five predefined criteria
What the model was asked to detect
The evaluation used categories defined in advance so that model output could be compared with a structured human reference.
Imprecision
Wording too vague to be understood or implemented correctly.
Ambiguous obligation
Uncertainty about whether the wording is a recommendation or a mandatory instruction.
Information outside scope
Content that should not appear in the document sent to the employer.
Disguised unfitness decision
A restriction or job change expressed in an excessive or inappropriate form.
Medical confidentiality
Information that could reveal a diagnosis or confidential medical detail.
Main results
Potential as a second reader, with identifiable limits
These figures apply to the specific corpus and protocol studied. They do not demonstrate a general ability to produce or validate occupational-health recommendations autonomously.
78%
of recommendations in the previous consensus dataset contained at least one quality defect.
385
recommendations were randomly selected for the comparative evaluation.
74.6%
overall agreement between the model and the multidisciplinary consensus.
0
hallucinations observed in the comparative sample that was assessed.
Interpretation
What the study can — and cannot — establish
What this use may contribute
- A repeatable second-reading framework.
- Systematic attention to predefined quality criteria.
- Prompts for reconsideration before a document is transmitted.
- A support for training and discussion within an occupational-health team.
What remains a limitation
- The model does not know the real job unless the relevant context is supplied.
- Over-detection may increase verification workload.
- Performance may change with the corpus, model version and prompting strategy.
- Clinical and professional responsibility cannot be delegated to the system.
Read the published study
The full article is available in French through the French National Research and Safety Institute for the Prevention of Occupational Accidents and Diseases.