Large language models can now draft, summarize and restructure text with striking fluency, and pathology reports are an obvious candidate. But a report is a clinical document that drives treatment, and fluent text is not the same as correct text. This article is for pathologists, lab managers and hospital IT teams weighing large language models for reporting, with a balanced look at where they help, where they can cause harm, and which safeguards should be non-negotiable.
What large language models are, briefly
A large language model is a system trained on very large amounts of text to predict likely sequences of words. From that training it learns grammar, style and a great deal of general knowledge, which lets it write, summarize, translate and answer questions.
It does not look anything up unless the surrounding system is designed to do so, and it knows nothing about your patient beyond what is in the text it is given.
Two properties matter most for pathology. First, the model generates text that sounds plausible, whether or not it is correct. Second, its behavior depends heavily on the instructions and context it is given, so the same model can be safe in one design and risky in another.
Opportunities in pathology reporting
Drafting from dictation or notes
A model can turn a pathologist's dictated or abbreviated findings into well-formed report text in the lab's preferred style. Paired with templates, this can reduce typing and editing time. For how to keep templates flexible, see using report templates without losing clinical nuance.
Structuring free text
Many reports still mix narrative and required data elements. A model can help map free text into structured fields, supporting the move toward synoptic reporting without forcing pathologists to abandon narrative where it adds value.
Consistency and completeness checks
Models can compare the description, diagnosis and comment for contradictions, flag laterality or site mismatches, and point out elements that the lab's protocol requires but the draft omits. This acts as a proofreader that never tires.
Readability
A model can suggest clearer phrasing, expand abbreviations or highlight ambiguous wording that a clinician might misread. The pathologist decides whether to accept each change.
Summaries for tumor boards and clinicians
Long cases with multiple specimens, addenda and ancillary results can be hard to follow. A model can draft a concise summary of final diagnoses and pending items for a tumor board or referring clinician, as long as the summary is checked against the signed reports and clearly labeled as derived from them.
The risks of large language models in reporting
Fabrication
The best-known failure is that a model may state something that is not supported by its input, such as an invented measurement, margin status or lymph node count. In a pathology report, a single fabricated detail can change management. This is the main reason output must be reviewed line by line against the pathologist's own findings.
Subtle distortion
Less dramatic but equally important, a model may soften or strengthen a statement, drop a qualifier such as "suspicious for", or turn uncertain language into certain language. Diagnostic nuance lives in those words.
Automation bias
When drafts look polished, reviewers tend to skim. The more reliable the system appears, the easier it becomes to miss its occasional errors.
Privacy and data handling
Reports contain sensitive personal data. Sending them to an external service raises questions about where data is processed, whether it is stored or used for training, and which laws apply. Thailand's Personal Data Protection Act treats health data as sensitive personal data; our article on Thailand's PDPA and patient data in the laboratory explains the implications for labs.
Change over time
Models and their providers update frequently. A change in the underlying model can change output behavior, so a design that was validated last quarter may behave differently today.
Language and terminology
Models trained mostly on general text may handle pathology terminology, local abbreviations or non-English reports less reliably. Labs reporting in Thai, or in a mix of Thai and English, should test specifically on reports in the languages and styles they actually use rather than assuming performance carries over.
Safeguards labs should insist on
- The pathologist remains the author. Generated text is a draft. Nothing reaches a clinician without the pathologist's review and signature.
- Ground the model in the case. The model should work only from the pathologist's own findings and the lab's templates, not add content from general knowledge.
- Make generated text visible. Highlight what was suggested so reviewers know exactly what to check.
- Prefer checking over creating for high-risk elements. Using a model to flag inconsistencies is lower risk than using it to generate diagnoses or staging.
- Log everything. Record the input, the output, the model version and the final edits for audit and quality review.
- Control data flows. Know where data is processed and whether it is retained, and make sure contractual and technical protections match your legal obligations.
- Validate and monitor. Test on representative reports before go-live, review samples regularly afterward, and recheck after model updates.
These safeguards are easier to enforce when the model is built into the reporting system rather than used in a separate window. Integration lets the lab control what data the model sees, record what it produced and require review before sign-out, instead of relying on individual habits.
Common mistakes
- Treating a general-purpose chatbot as a reporting tool, with staff pasting patient text into consumer services.
- Allowing the model to add diagnostic statements not present in the pathologist's input.
- Measuring success only by time saved, without tracking errors caught at review.
- Rolling out to all report types at once instead of starting with lower-risk tasks such as proofreading.
- Skipping user training on how the model can fail, so reviewers do not know what to look for.
Key takeaways
- Large language models write fluent text, but fluency does not guarantee accuracy.
- They are most useful for drafting from the pathologist's findings, structuring text and checking consistency.
- Fabrication, loss of nuance and automation bias are the main clinical risks; data handling is the main legal one.
- Keep the pathologist as author, ground the model in case data, and log inputs, outputs and versions.
- Start with low-risk tasks and validate before expanding.
PATHOLOGY+ applies these principles with AI-assisted drafting and consistency checks of descriptions that the pathologist reviews and signs. Book a demo to see how the review step works in practice.