Large Language Models Enter the Surgical Workflow With Clear Utility and Clear Limits

Large language models are moving from novelty to bedside-adjacent utility in surgery, and the most immediate appeal is efficiency. In a field where operative decision-making, documentation, inbox management, and patient counseling all compete for the same limited hours, tools that can compress language, summarize records, and draft routine communication naturally attract attention. The review places these systems in a practical frame: not as autonomous clinicians, but as assistants that may reduce clerical load and speed information handling when used with oversight.

A single sentence captures the central tension. These systems can save time, and they can also produce convincing errors.

That tension matters because surgery depends on precision. A weak summary of a patient’s comorbidities, prior procedures, or perioperative risks is not a minor annoyance; it can distort handoffs, affect planning, and amplify downstream error. The review positions LLMs as tools best suited to tasks that are repetitive, text heavy, and reviewable, rather than tasks that demand independent judgment, interpretation of uncertainty, or accountability for irreversible clinical decisions.

Documentation, Chart Summaries, and the Pressure to Reclaim Time

In clinical workflows, one of the most plausible uses is ambient documentation. If an LLM can listen, transcribe, and organize a visit into a cleaner note, or condense a long chart into a usable summary, the result could be less clerical friction and more time for direct clinical work. That is especially relevant in surgery, where preoperative assessment often requires stitching together prior imaging, medication lists, operative history, and evolving plans across multiple encounters.

Chart summarization also has a straightforward cognitive value. Surgeons routinely need to integrate large amounts of information quickly, and an LLM can help sort that material into a more navigable form before the clinician validates it. Used well, this is not automation of judgment; it is compression of text. Used poorly, it can create false confidence if the user accepts the draft without checking the underlying record.

The operational lesson is direct: any time saved at the keyboard only becomes real if the output is verified line by line against the source material.

Patient Communication May Be the Most Immediate Clinical Use

The review gives substantial attention to patient-facing communication, and the rationale is strong. Surgical instructions often require plain language, careful sequencing, and sensitivity to health literacy. LLMs can rephrase complex explanations into simpler wording, adapt content to different reading levels, and translate instructions into other languages. In perioperative care, that could mean clearer preparation guidance, more understandable discharge instructions, and less confusion about wound care, medication changes, or return precautions.

Empathy is another area where these systems may help, though only in a constrained way. Draft responses to patient portal messages can be shaped to sound respectful, organized, and calm, which may improve efficiency when teams are managing high message volume. Still, the emotional tone of a machine-generated reply cannot substitute for clinical responsibility, particularly when messages include red flags such as fever, bleeding, worsening pain, or signs of surgical complications. Human review remains essential because the content may be more consequential than the phrasing.

For a surgeon, the practical question is not whether the tool sounds polite. It is whether the message is accurate, safe, and appropriate for the specific patient context.

Research Support Without Replacing Scientific Judgment

In academic surgery, LLMs may be especially useful for literature triage and early-stage research support. They can summarize articles, organize background material, help refine study questions, and assist with protocol drafting. The review also identifies risk-of-bias assessment in randomized trials as a possible use case, though any such assessment still requires expert confirmation. The value here is not originality; it is speed and structure.

That distinction is important because research quality depends on reasoning, not prose alone. A model may help gather and arrange the material, but it cannot decide whether a hypothesis is clinically meaningful, whether a design is adequately powered, or whether the final interpretation survives scrutiny. In practical terms, LLMs may move a surgeon-scientist faster from blank page to draft, yet the burden of scientific rigor stays exactly where it belongs: with the investigator.

Prompting Discipline and the Need for Verification

The article emphasizes prompt engineering as a core skill. Effective use depends on clear clinical objectives, explicit instructions, and iterative refinement. Vague prompts invite vague outputs. Specific prompts can narrow the task, define the audience, and improve the usefulness of the response. For example, a request to rewrite discharge instructions for a patient with low health literacy will yield a better result than a generic request to “make this clearer.”

Even a well-constructed prompt does not solve the larger safety problem. Hallucinations remain a central limitation, meaning the model may generate text that sounds authoritative but is factually wrong or unsupported. In surgery, where erroneous detail can propagate quickly through documentation or patient instructions, every output needs human verification. The review treats fact-checking not as a final polish step, but as a required safeguard.

Public-facing LLM platforms introduce another hard boundary: protected health information should never be entered into systems that are not compliant with privacy obligations. The privacy risk is not theoretical. Entering identifiers, operative details, or other sensitive data into an unsecured platform can create downstream exposure that no amount of output quality can undo.

Privacy, Liability, and the Question of Accountability

The most unresolved issue is responsibility. If an AI-generated error contributes to harm, liability frameworks remain unsettled, with accountability potentially distributed among providers, institutions, and developers. That ambiguity has immediate implications for adoption. Clinicians may be willing to use a tool that saves time, but institutions need to know who owns the risk, who reviews the output, and which use cases are permitted.

Surgery is built on identifiable lines of responsibility. The operating surgeon, the care team, and the institution all have defined roles. LLMs blur those lines if they are introduced without clear policy, documentation standards, and oversight pathways. For that reason, implementation should be conservative, explicit, and monitored, especially in settings where the output enters the medical record or directly shapes patient care.

The clinical promise is real, but it is bounded by verification, privacy, and accountability. Human judgment still sits at the center of surgical care, and no model removes that obligation.


Reporting by Dr. Aishwarya N., MBBS for Medicine World Foundation, with additional data provided by Fierce Healthcare.