Using AI to draft a lay summary
Drafting a lay summary with AI turns on what the tool is given, who checks its output against the sources, how the workflow is tested and governed, and what the summary discloses about the AI's part and the people's.
What AI offers
Considerations for incorporating AI into established lay summary workflows supplement existing development practice, including the Good Lay Summary Practice (GLSP) guidance, and do not replace general best practice for writing lay summaries. The same principles could also apply to other public- and patient-facing material, such as lay language protocol synopses.
AI brings considerable benefits to the development of lay summaries, along with risks that have to be managed with care. Using it to develop lay summaries can enhance transparency and accessibility while maximising efficiency and helping the work to scale. It can make drafting more efficient by reducing the time and resources spent on manual work. It can raise productivity by handling repetitive tasks such as drafting, organising information and simplifying technical language. It has the potential to make lay summary creation more efficient and to promote health literacy by spreading trial results more widely to patients and the public, faster and cost-effectively. Used effectively, it can streamline development and deliver clear, concise content sooner, and with fewer resources needed, sponsors may be able to produce summaries for more trials. By automating tedious tasks such as literature review, data extraction and summarisation, AI tools enable researchers and medical writers to focus on higher-level tasks such as content curation and quality assurance.
A 2023 study demonstrated the feasibility of using AI to produce plain language summaries for clinical trials, highlighting its potential to accelerate their production. Advances in AI have revolutionised the generation of plain language summaries of trial results, and natural language processing can analyse complex trial data and synthesise it into clear, concise summaries tailored to specific audience needs.
Many content creators use AI writing assistant software that extracts information from the web and automatically produces a summarised text, which the content creator then edits before publishing. The prospects for such software in clinical research are massive, especially combined with natural language processing translation tools that produce documents disclosing complicated information to the general public, and this is relevant to helping medical writers produce documents that comply with strict European regulations, such as lay summaries in each European language.
Where AI falls short
While large language models can generate human-like text, they still have limitations to be managed. Without appropriate safeguards, AI-generated lay summaries may contain inaccuracies: hallucinations, which are false, fabricated or misleading information that may arise when the model lacks adequate input and training; a lack of nuance, especially in scientific data; and insensitivity to tone and culture. Standalone use of AI to create lay summaries concerns sponsors, because AI lacks the nuanced understanding and contextual knowledge that human experts bring to complex trial data. Without proper human oversight, an AI-generated summary may misrepresent those data or miss critical details, leading to inaccuracies or misinterpretations. This concern was observed in 2023 in a large-scale instance of publicly posted AI-generated lay summaries that lacked proper human oversight, and those summaries were eventually removed from the public domain after significant concerns were raised about their accuracy.
Misinformation is unintentional error, which can occur when AI misinterprets data or lacks the context to understand scientific concepts, and disinformation is the deliberate distortion of facts with intent to mislead. Either may arise where the AI system is open-source or trained on public data without proper vetting, and the opacity of AI decision-making compounds the risk. If the source information used by natural language processing tools is correct, the risk of “fake health news” emerging in wellness blogs and other social media should not be high, but policymakers need to take serious steps to protect the public from potential misinformation generated by software still in development, or by its operator.
Bias in training data or in user prompts, intended or not, can produce biased output, and when a model is trained on large data sets that may not fully reflect diverse cultural backgrounds, its content can lack cultural awareness and sensitivity. AI can reproduce and even amplify those biases, and the skewed summaries that result compromise the objectivity of the information given to patients and the public. Training data may also include marketing content, which may lead to persuasive or overly positive wording; that wording can bias the presentation of results, potentially misleading readers about a study's significance, benefits or risks.
AI technologies change quickly, and using an outdated model may lead to inaccuracies and inconsistencies. The pace of change may also make it hard to keep tools aligned with the latest standards and best practices, which could raise the risk that a summary does not meet current regulatory or quality expectations. Appropriate AI governance can mitigate that risk effectively.
Inputs and prompts
The quality and comprehensiveness of the inputs are crucial to an accurate and relevant summary. Data taken solely from a source such as ClinicalTrials.gov may lack the context needed for an accurate and complete summary. Reducing the risk of hallucination or overconfidence helps to prevent responses that seem legitimate yet may omit critical information or draw incorrect conclusions. It would also help if the model gave references to the source documents it takes data and information from. Key data sources may include aggregate tables, figures and listings; clinical study protocols; clinical study reports; informed consent forms; the lay protocol synopsis; other nonpromotional public or patient-facing documents; and glossaries of medical terms with their plain language equivalents.
Standardised templates can help keep summaries consistent and compliant with regulatory requirements. The AI should be trained to work within them while leaving room for differences such as study design and therapeutic area. It should also be trained to use a glossary of preferred terminology for a document or a set of documents, together with previously completed summaries.
Clinical study data sets contain sensitive personal health information about the participants. To protect their privacy, the inputs should contain no identifiable patient data, aggregated data should be used, and organisations must ensure that models do not retain sensitive information. Good data stewardship is required.
Prompt engineering is a critical part of using AI effectively, guiding it to create accurate, understandable content suited to the public and patients. Each summary should be drafted through multiple, sequential prompts, for drafting individual sections and for setting the context clearly. The prompts should give specific instructions on tone and style, and guidelines for simplifying complex concepts. Such prompts help the model set the right tone, keep the output consistent with approved medical terminology and address potential biases. Reminders in them to give the necessary context and caveats can help make the output both informative and suited to patients. Sequential prompts used for refinement can help improve the quality of a draft.
Example prompt wording names the condition and sets the audience's reading level at sixth grade or at 12 years old. It asks for a compassionate and encouraging tone that keeps factual accuracy; for a medical term to be explained in simple language a non-expert can understand; for simple words and short sentences, without jargon or complex medical terminology; for a simple explanation followed by the technical term in parentheses, as in “sugar in the blood (glucose)”; for a statement of the study's limitations, that the results may not apply to all patients and that individual responses may vary; and for headings and bullet points. One example, for a new diabetes medication and a general public that includes patients with type 2 diabetes, asks for a two- or three-sentence overview of the study's purpose, followed by sections on what the study was about, what it found, what that means for the reader and what the next steps are. It asks for the whole summary inside summary tags, factually accurate to the trial results provided and easy for the audience to understand, and leaves a place to insert the documentation of those results if it is within the model's capabilities or applicable. Another asks for three paragraphs on why a trial, named with its NCT number, is being done: the condition, then the study drug and why it is being developed, then the trial design, with the hypothesis restated in the final sentence.
Human review and approval
Responsible use of AI is critical, and AI should complement human expertise and review, not replace them. All machine-generated output should be reviewed by people, who bring essential judgement where AI may fall short. To better ensure accuracy and appropriate tone, professionals and members of the intended readership should review and refine AI-generated content. A hybrid approach, in which AI supports drafting and experts ensure accuracy and appropriateness, can maximise the benefits while minimising the risks, so that sponsors can maintain regulatory compliance while patients and the public receive timely, understandable and accurate information to make effective decisions about their health. AI can improve efficiency, reach and scalability, but human oversight remains critical, and AI output must be guided by human expertise to ensure accuracy, relevance, sensitivity and adherence to regulatory requirements. The joint effort of people and the machine can help summaries reach high standards of accuracy, transparency and compliance while building public trust and empowering patients to make informed decisions about their health care.
How well AI generates lay summaries depends on the expertise of the people who train, prompt and oversee it and who generate and revise the summaries. To ensure adherence to best practice and maintain quality and accountability, all the reviewers and approvers that the GLSP guidance recommends should keep their essential roles, skills and qualifications when AI tools are brought into the process. Standard operating procedures and resourcing vary between organisations, but people with additional AI knowledge and experience may play critical roles at various stages:
- AI training and development experts, whose experience is required to design and calibrate the system and train it on relevant inputs and data sets.
- Health literacy specialists, to help train the AI to simplify complex medical language and to guide which terminology, explanations and formatting suit the reader.
- Legal and compliance teams, whose expertise should be built into training to ensure that the AI uses data safely and in line with AI and data use laws and policies such as GDPR or HIPAA, with data privacy monitored through the standard review procedures for lay summaries.
- Medical writers, to check that the AI's interpretation of the results is factual and that no critical scientific nuance is missing, a task that differs from their traditional authoring of this information.
Only specialised translators or writers should use AI writing assistant software to generate documents, and pharmaceutical companies and medical writers are responsible for the accuracy of the information disclosed.
AI should be placed at specific points of the existing development process, such as the best practice and overall process that the GLSP guidance lays out, with clear roles for human review and approval. In a suggested flow, AI can help to make an initial draft by processing the study data, simplifying technical details and ensuring consistency, which reduces the time spent on manual tasks, and human experts should be involved in writing the prompts and entering the data. Human experts, including medical writers and health literacy specialists, should review the draft for accuracy, ethical compliance and language suited to the audience, correcting errors and putting the findings in context. The refined document should then undergo a quality control and compliance check in line with existing processes, which ensures that it complies with data privacy laws, regulatory standards and ethical guidelines for communicating study results. Review and approval should align with existing processes, including cross-functional teams and patient advocates or advisory groups, which ensures that the summary is clear, engaging and culturally sensitive. At that stage, AI could likely help with smaller tasks such as checking grammar and spelling.
Governance, testing and monitoring
Robust AI governance is essential for overseeing any new system, an AI system included, and implementing AI is an iterative process that requires initial testing and continuous improvement. Incorporating AI into lay summary development calls for thorough planning and careful implementation, and it requires balancing innovation with oversight. Successful implementation will be an ongoing process of continuous monitoring, evaluation and refinement.
AI applications in health care were increasingly subject to oversight from regulators. The US and the EU were developing frameworks aimed at data privacy, accuracy and ethics, such as the US Blueprint for an AI Bill of Rights, the NIST AI Risk Management Framework, the EU Artificial Intelligence Act and the EU Ethics Guidelines for Trustworthy AI, alongside the FDA's guidance on the use of AI in the development of drugs and biological products. ICMJE, AMWA, EMWA and ISMPP emphasised transparency, accuracy and human oversight in AI-generated medical writing, and sponsors were developing and deploying their own AI use cases and policies, including tools for drafting scientific and public- or patient-facing documents. Guidance on AI use in medical information was limited, and none was specific to lay summaries or other patient-facing clinical research information. As frameworks, guidelines and technology change, those involved must stay informed and adopt best practice so that summaries remain of high quality and compliant.
Effective governance is crucial when AI is used for lay summaries, so that its use aligns with organisational goals, regulatory requirements and ethical standards. Its key components should include a cross-functional team of medical writers, statisticians, legal experts, patient advocates and AI specialists; clear guidelines and standard operating procedures for AI use in patient communications; a review and approval process in which subject matter experts validate AI-generated content; an ethical framework that addresses bias, privacy and transparency; compliance with relevant regulations such as the EU AI Act, GDPR and FDA guidelines; and comprehensive training for staff, with resources to help them understand AI's capabilities, limitations and good practice in collaboration between people and AI systems. Documentation of training data, algorithms and decision-making processes should be kept for regulatory audits, and a process should track changes in regulation and adjust AI systems and governance practices to them. Safeguards should protect patient data and confidentiality, and the ethical implications of the AI's use should be assessed regularly, with adjustments made to keep it aligned with organisational values, industry best practice and societal expectations. Key performance indicators should measure the accuracy, readability and effectiveness of AI-generated communications, and the rollout should be phased, from low-risk applications to more complex tasks.
Initial testing should use a comprehensive suite of test cases across study phases, designs, endpoints, safety data sets and patient populations, and should compare AI-generated content with human-written content for patient preference and understanding. An example testing process runs in this order: generate the initial content with AI; have people review it for accuracy, readability and health literacy level, using validated tools; bring in the public and patients for feedback on how understandable and relevant it is; adjust the prompts or the AI models in the light of the feedback; repeat until the results are satisfactory; put the workflow into a limited rollout and monitor its performance; and scale up on the strength of successful performance metrics.
Monitoring should be ongoing, given AI's continuous learning, through regular audits of the AI's performance, ongoing monitoring of outputs that includes random sampling and review by human experts, protocols for correcting errors and biases, and feedback loops that take in human experts, patients, healthcare providers and other stakeholders. Models should be updated and retrained regularly on new data, feedback and evolving best practice, and systems kept up to date with advances in AI technology. Regular monitoring, and updates to processes and models with the latest medical and regulatory information, will likely be essential to mitigate the risks and keep the highest standards of accuracy, clarity and ethical practice. Organisations can harness AI to enhance their lay summary processes by addressing its opportunities and risks carefully and through continuous learning.
Multi-agent systems
Using AI most effectively may require a more advanced architecture, such as a network of AI agents, each with a specialised role such as medical fact-checking, readability optimisation or bias and sensitivity detection, into which orchestrator agents can also be integrated to coordinate the specialised agents' work much as a project manager would, while people continue to give expert oversight and intervene at key points. Such an architecture divides complex tasks among several AI components working in coordination, and it mirrors team-based document creation in traditional settings but offers greater consistency, scalability and traceability.
An example architecture has specialised agents for strategy planning, which analyses the source documents and develops a structural approach, maps the complexity of the information, creates audience-appropriate templates and sets measurable objectives for reading level, length and key messages; drafting, which turns clinical documents into a first draft with a logical structure, a balance of detail and the target reading level; medical accuracy checking, which verifies factual correctness, cross-references claims against the source documentation and flags statistical information for expert verification; readability optimisation, which refines the language for the intended audience, adjusts the text using readability metrics and suggests simpler terms that preserve meaning; and bias and sensitivity review, which aims at inclusive content, identifies potentially exclusionary or stereotyping language and checks for balanced representation and culturally sensitive explanations. An orchestrator agent manages the workflow and integration, routes content between the agents, resolves conflicts, maintains the integrity of the document and identifies where human intervention is needed, and a feedback integration agent processes the experts' input, categorises and prioritises it and updates the agents' parameters based on patterns in the feedback. Human experts give strategic oversight at key junctures, review the planning outputs and the initial parameters, evaluate flagged uncertainties that require domain expertise, give structured feedback and give final approval.
The example architecture's implementation workflow runs in nine steps: planning; a first draft; review by the accuracy, readability and bias agents; integration of their inputs into a revised draft by the orchestrator; structured feedback from human experts; refinement by the agents; repetition of feedback and refinement until quality thresholds are met; system learning, in which patterns in the feedback update the agents' parameters for future projects; and final sign-off by human experts, with complete documentation of the process. Its benefits are specialisation of agents for specific tasks, traceability through clear documentation of decisions, adaptability through learning from expert feedback, and scalability through a consistent approach across document types. Testing and monitoring such a system is critical and should follow the governance considerations for AI in lay summaries.
Disclosing AI use
Transparency about AI's part in a lay summary is essential to public trust, ethical standards and regulatory compliance. Failure to disclose it can lead to scepticism, undermine confidence in the information and damage the credibility of the author or the organisation. As AI capabilities advance, open communication helps to address misconceptions about the technology and to build a more informed and trusting relationship between the public and the research community.
The EU Artificial Intelligence Act, in Chapter 4, Article 50, paragraph 4, requires deployers of an AI system that generates or manipulates text published to inform the public on matters of public interest to disclose that the text has been artificially generated or manipulated. One view holds that not stating clearly that content was generated by AI is an ethical dilemma in any context, that the final version of the EU AI Act requires the user to know when an AI agent is involved, and that AI use should therefore be stated clearly.
A disclosure should state clearly that AI was involved, typically in the introduction or a dedicated section, and to what extent, such as for initial drafting, language simplification or checking spelling and grammar. Where AI was used to create a summary, the summary itself has to declare that AI was used and for what purpose. It may include a brief explanation of how AI was used together with human expertise to ensure accuracy and relevance, and it should be easy for the intended readers to understand, without technical jargon. It should also state how far the study sponsor and medical experts were involved in reviewing and approving the summary, acknowledge any involvement of the public or the patient community, and say so where human involvement was limited or absent. Disclosure practice should align with the EU AI Act or similar applicable regulation, give the information on the AI system's purpose, capabilities and limitations that applicable law requires, and include contact information for enquiries about the system or its outputs.
Example statements cover, in turn, an AI first draft reviewed, revised and approved by qualified medical professionals; AI used to simplify language and organise information, with all content verified and approved by the study team and patient representatives; the sponsor's review of the summary for accuracy and alignment with the trial results; review by members of the public, patients and patient advocates for understanding and relevance; and compliance with AI regulation, naming the system's developer and its purpose and giving a contact for more information.
Supporting tools and resources
Organisations should use existing tools and resources and develop further resources specific to their own use of AI for lay summaries. Quality control checklists verify medical facts and statistics against source documents such as clinical study reports and published literature; check consistency within and between documents and with approved messaging and terminology; assess readability and health literacy level; evaluate cultural sensitivity and inclusivity; identify potential biases and misleading, oversimplified or illogical statements; and confirm compliance with regulatory requirements and internal guidelines. Tools for implementing and evaluating AI include frameworks for assessing accuracy and readability, benchmarking tools that compare AI output with content written by people and, where AI agents are used, developer tools that help to show how the agents reach their decisions. Collaborative platforms include secure platforms for collaboration between AI systems and human experts, version control to track changes and approvals, annotation tools for feedback on AI-generated content and project management software to coordinate review and approval. Data privacy safeguards include anonymisation and de-identification tools, secure file transfer, access controls, encryption of data at rest and in transit, and templates for privacy impact assessments.
Staff who work on AI-assisted summaries should have comprehensive training materials, e-learning modules on AI's capabilities and limits, good practices for collaboration between people and AI, and regular workshops and webinars on emerging AI technologies and ethical considerations. Further resources are the GLSP guidance, the ISMPP position statement and call to action on artificial intelligence, the EMA's artificial intelligence workplan, its four principles for the safe and responsible use of large language models, its guiding principles on the use of large language models in regulatory science and for medicines regulatory activities, and the EU Artificial Intelligence Act.