AI-assisted hospital discharge summaries: who is responsible for the content

Where AI genuinely helps in preparing a discharge summary

At the end of a hospital stay, the doctor has to summarise the course of treatment, results, diagnoses, medications and recommendations. That is a lot of information to organise, especially when the records have been built up over several days and added to by different people. An AI tool approved for use in the facility can prepare an initial version of the discharge summary based on data it is authorised to access. It can put events in chronological order, cut out repetition and suggest sentences that are easier for the reader to understand. This is useful, as long as the editing does not change the meaning of the medical information.

Fluent writing, however, is not the same as knowing the patient’s condition. The model has not seen the patient, has not examined them and is not the person responsible for their ongoing care. It may also fail to recognise which information is crucial and which is merely peripheral. That is why a hospital needs rules: when the tool may be used, what data can be entered into it, who compares the draft with the records and who gives final approval. In this context, Empatyzer is used to practise communication, not to produce medical discharge summaries.

The most dangerous error may be an omission

A patient reads the discharge summary and asks: “Why is there nothing here about the reaction to the medication we talked about?” Even if the rest of the document sounds professional, a missing reference to something important may matter the next time the patient deals with healthcare services. Language models can generate false information, but omissions are just as important. Detecting them requires comparing the document with the actual course of care, not merely checking the style.

In a study by Mehri and colleagues, published in 2026 in the journal eBioMedicine, 292 pairs of discharge summaries – one prepared by doctors and one by an AI model integrated with the medical records – were compared. Overall ratings were similar, but the model-generated versions were rated lower for completeness and judged to be more concise. The authors pointed to the technology’s usefulness, provided there is appropriate oversight. This does not mean that all tools achieve the same result, or that any facility can automatically do away with checks. A concise text can be an advantage, but only if no information that the patient or the next doctor needs has disappeared from it.

Who prepares the draft, and who checks and approves it

The patient asks: “Did ChatGPT write this discharge summary? Who is responsible for it?” If the facility really does use an approved system, staff can say: “The software helps prepare a draft based on our records. An authorised doctor checks it and approves it in line with the hospital’s rules.” The answer must, however, match the actual process. You must not claim that the document was checked if it was not, or describe a tool used outside the procedures as approved.

In Polish healthcare, medical records must meet requirements concerning their content, authorship and the way entries are kept. Using an editing tool does not remove the obligations of staff or of the healthcare provider. How patients’ data are processed also matters. Before rollout, the hospital should establish the basis and conditions for using the system, the scope of access, security, and responsibility for the accuracy of the document. A member of staff should not paste a patient’s data into a random public service just because it will help them write a discharge summary quickly. Above all, the patient needs to be confident that their information is protected and that the person responsible can explain and correct the document.

What to say when a patient spots an error

“It says here that I didn’t have a temperature, but I told you about it yesterday.” A poor response would be: “AI sometimes does that – there’s nothing I can do about it.” The doctor might reply: “Thank you for spotting that. I’ll check the entry in your records and we’ll establish whether the discharge summary needs correcting. If it makes a difference to your follow-up instructions, we’ll clarify it before you leave.” This is not a hasty admission of fault, but a commitment to checking a specific fact.

How to check the draft before the document reaches the patient

It is staff who should catch errors before the patient receives the discharge summary. On the ward, it is worth using a simple procedure for comparing the draft with the records: diagnoses, medications including dosages, allergies, important results, procedures performed, recommendations, follow-up dates and symptoms that require urgent contact. Particular attention should be paid to information that is easy to leave out of a summary. If a discrepancy is found, it must be corrected in line with the agreed procedure, the required history of changes must be kept, and the patient must be given the correct version of the information. A significant mistake may require an additional clinical response; it is not enough simply to swap out one sentence.

An understandable discharge summary is more than plain words

A discharge summary can be clinically correct and yet incomprehensible to someone who is only just going home. “Outpatient follow-up as indicated” says little if the patient does not know which clinic to attend and when. AI can help simplify wording, but the simplification itself needs checking. Replacing a technical term with an everyday phrase must not change the recommendations. Nor do simple sentences guarantee that the patient will understand the recommendations correctly.

At the end, the doctor or nurse can ask: “What will you do when you get home? Which medication are you taking from tomorrow, and when are you coming in for your follow-up?” The aim is not to test the patient, but to check whether the explanation was sufficient. If the answer reveals a misunderstanding, the plan needs clarifying. The way the document is handed over should also take into account stress, fatigue, language barriers and the need for support from someone close, if the patient wishes and the law allows. The discharge summary then helps care to continue, rather than simply closing the hospital stay.

AI can make preparing a discharge summary easier, but it does not replace verification against the records or the responsibility of staff. The biggest danger may lie not only in an incorrect sentence but also in leaving out important data. The patient needs a correct document, understandable recommendations and a simple answer about whom to turn to if something does not match. A clear explanation of how the document is produced helps build trust.

How Em helps you prepare to talk to a patient about the role of AI

A doctor learns that a patient was worried on hearing that the discharge summary had been prepared by computer. She wants to respond clearly, without jargon and without needlessly alarming the patient. She describes the anonymised situation to Em, along with the actual rules for approving documents at her hospital. She asks for two versions of a short answer to the question “Who wrote this?” and for a way of taking feedback on board if the patient notices a discrepancy. Em helps her remove the defensive line “the system sometimes makes mistakes” and replace it with a message about a specific action: “We’ll check the information you’ve pointed out and explain what happens next.” In the next exercise, the doctor turns technical explanations into language that may make sense to a patient after a stressful hospital stay. She can also reflect on her own reaction to a critical question and prepare a calm paraphrase. Authorised staff, acting in line with the facility’s procedures, are responsible for checking the discharge summary, making clinical decisions and documenting corrections. The result of working with Em is a clearer conversation, in which the patient knows who is responsible for the discharge summary and whom to report an error to.