CoPilot Tag

Does AI Help You Reflect, or Just Write Faster?

By Nigel Hinchliffe Director of Education, CCH.

During a curriculum review workshop several years ago, a group of newly qualified nurses were asked to share a reflective account they were proud of. Most had written competent, structured pieces that moved neatly through Gibbs’ stages: description, feelings, evaluation, analysis, conclusion, action. When asked whether those accounts had changed anything about the way they practised, the room went quiet. They had learned to write reflection. They had not necessarily learned to reflect.

“They had learned to write reflection.
They had not necessarily learned to reflect.”

That gap, between the performance of reflection and reflection as a genuine habit of professional learning, is not a new problem. It is, however, an urgent one. Generative AI tools are now capable of producing plausible reflective writing at speed. Before the profession reaches for AI as a solution to reflective practice, it is worth being aware of what the problem actually is.

Has reflection been taught as a skill?

Healthcare professionals are required to reflect, assessed on their reflections, and held accountable for them at revalidation. The Nursing and Midwifery Council, for instance, requires five written reflective accounts as part of the revalidation process (NMC, 2019). 

What is less common is being explicitly taught how to reflect: how to move from description into analysis, how to connect personal experience to evidence, how to sit with uncertainty rather than resolve it prematurely into a tidy learning point.

Gibbs’ Reflective Cycle (1988) is the dominant framework, precisely because its six stages – description, feelings, evaluation, analysis, conclusion, action planning – provide a scaffold practitioners can follow without having been taught the underlying skill. That is both its strength and its limitation. Applied well, Gibbs takes a practitioner from raw experience to transferable learning. Applied as a form to be completed, it produces the kind of account that satisfies an assessor without troubling the practitioner who wrote it.

This is the baseline problem AI enters. Not a profession of skilled reflectors who lack time, but a profession in which written reflection may well be delivered, but practitioners may not have learned how to reflect

What AI offers, and what it does not do automatically

Generative AI tools – Claude, ChatGPT, Copilot and others – are capable of functioning as structured thinking partners in a way that a blank page cannot. Used well, they can prompt practitioners through Gibbs’ stages with tailored questions, surface relevant clinical frameworks, and help translate half-formed thoughts into coherent prose. 

For a nurse processing a safeguarding concern, or a GP working through a difficult prescribing decision, an AI interlocutor that asks “What assumptions were you making at that point, and what would challenge them?” can move the reflection further than a prompt to “complete your analysis section”.

But this is not automatic. A general AI model’s default behaviour is to affirm and reflect back.

Asked to help with a reflective account, it will typically smooth, summarise and endorse. Getting it to meaningfully challenge a clinician’s reasoning requires deliberate prompt design. That design is itself a competency, and one that many practitioners have yet to acquire. The AI needs to be asked to:

• Notice a self-serving account

• Probe an unexamined assumption

• Push back on a premature conclusion

There is also a legitimate concern about cognitive offloading. The effort of articulating an experience is an important aspect of reflective learning: finding the words, sitting with ambiguity, returning to an account and revising it. Moon’s theoretical work on reflection and experiential learning argues that this effortful processing is part of how meaning is constructed from experience, rather than a delivery mechanism for insight that exists independently of it (Moon, 2004).

The real risk: effortless performance

A 2024 systematic review of reflective writing as summative assessment found that power dynamics between students and markers can lead to ‘performative instead of genuine reflection’ and concluded that student voices in assessed work may not represent authentic participation ‘because students may lack agency as they are on the wrong side of a power imbalance and are motivated to pass’ (Ross, Bohlmann and Marren, 2024). The conventions of a passable reflective account, such as a growth narrative, appropriate emotional awareness, or a clear action point,  are learnable independently of the reflection itself.

AI makes producing that performance effortless. A practitioner who has absorbed the conventions can generate a convincing Gibbs account of a clinical encounter they have not meaningfully examined, in minutes, with minimal cognitive effort.

AI will be used in reflective writing; it already is. The real question for educators and professional bodies is whether the assessment frameworks and supervisory practices we have built are robust enough to distinguish genuine reflection from sophisticated production. Many are not, which is a problem for education design to solve, not for the technology.

Where AI can make a real difference

None of this means AI has no role. It means the role needs to be designed rather than assumed.

For student practitioners, AI can function as a responsive scaffold – asking the questions a supervisor would ask, at the moment when the experience is still live rather than weeks later in a tutorial. 

The immediacy matters: Mann, Gordon and MacLeod’s review identified proximity to the experience as one of the variables that influences the quality of reflective learning (2009). A tool that meets practitioners in that proximate moment, rather than waiting for a scheduled portfolio deadline, addresses a real structural gap. 

A systematic review of how health-professions students use generative AI found these tools most often supported learning through inquiry and in practice, rather than simple content acquisition (Pham et al., 2025).

For experienced clinicians preparing for supervision or appraisal, AI offers a way to organise thinking before the formal conversation rather than during it, reducing the time spent on description and increasing what is available for deeper analysis. The preparation is the reflection; the supervision becomes its extension.

For practitioners processing emotionally charged events – errors, patient deterioration, moral distress – the non-judgemental quality of an AI interlocutor is not trivial. The psychological safety required for honest reflection is not always available in formal supervisory relationships, and a first pass engagement with an AI tool may lower the threshold for authentic disclosure in a subsequent human conversation.

Prompting as a professional skill

The common thread across these applications is that the value of AI in reflective practice depends almost entirely on how it is used. A tool set up to challenge will challenge; a tool set up to affirm will affirm. The distinction is determined by the practitioner’s prompting, and prompting is a skill.

“A tool set up to challenge will challenge.
A tool set up to affirm will affirm.”

That skill can be taught directly. Practitioners can learn to build challenge into their prompts in the same way they learn any other clinical or communication skill: through worked examples, supervised practice, and feedback. This might mean asking the AI to argue the counter-case, name the assumption they have not examined, or hold back the tidy conclusion until the uncertainty has been explored.

Embedding this in professional development does not require waiting for system-wide change. It can sit inside existing CPD structures:

• A short module on prompt design for reflective practice
• Worked examples built into portfolio guidance
• Supervisors modelling the technique in appraisal conversations.

The frameworks for reflective practice already exist; what is needed is training practitioners to use AI within them deliberately, rather than assuming the tool will do the work on its own.

The Gibbs cycle has endured for nearly four decades because its structure is sound. What it has always needed is a workforce able to use it as a habit of thinking rather than a form to complete. AI does not change that requirement. It raises the stakes on meeting it. Whether AI ends up supporting genuine reflection or accelerating its performance depends on whether that training happens now.

What this means in practice

When AI makes reflective writing easier, it does not automatically make reflective learning more likely. The outcome depends on whether practitioners learn to prompt it to challenge their thinking, rather than simply affirm it.

Colleges and providers who educate healthcare professionals can address this directly, rather than wait for the wider profession to resolve it. Teaching prompt design as part of CPD, portfolio guidance, and supervision is not a future ambition. It is available now.

How to Take This Further

If this is a skill you want to build, our short course, Getting More from ChatGPT: Prompting Skills for Healthcare Professionals, includes a prompt template for reflective practice specifically, as well as the course’s wider training in getting reliable, useful output from any AI tool you use.

Click here to learn more.

About the author

This article was written by Nigel Hinchliffe, Director of Education at the College of Contemporary Health (CCH). Nigel has extensive experience in clinical education, with a particular focus on how healthcare professionals develop and demonstrate competence throughout their careers. At CCH, he leads the development of evidence-based training programmes and regularly provides feedback on learner’s reflective submissions as part of their professional development.

Read More

Large Language Models Show Promise in Detecting Drug Safety Signals from Clinical Notes

Key Takeaways:

• Large language models can identify immune-related adverse events in clinical notes without task-specific training, offering a potential alternative to labour-intensive manual review
  
• Performance remains below the threshold required for clinical decision support, with models tending to overpredict adverse events
  
• Despite limitations, this approach may support large-scale safety monitoring and accelerate research into cancer immunotherapies
  
  
  

The challenge of detecting drug safety signals

Drug safety signals are often embedded within unstructured clinical text, particularly in electronic health records. Identifying these signals has traditionally required either manual chart abstraction, which is resource-intensive, or natural language processing systems tailored to specific drugs and healthcare settings.

This challenge is particularly evident in the case of immune checkpoint inhibitors. These cancer therapies, first introduced in 2011, are associated with a broad range of immune-related adverse events. These events can affect multiple organ systems, including the colon, liver, lungs, heart, nervous system, skin, and endocrine system, making systematic detection complex and time-consuming.

Exploring large language models as a solution

Large language models are increasingly being explored as a way to streamline the identification of drug safety signals within clinical text. A multicentre study, published in eBioMedicine, evaluated whether these models could detect immune-related adverse events associated with immune checkpoint inhibitors.

The study focused on a zero-shot learning approach. In this setting, the model receives a single, detailed prompt without prior examples. The prompt used by the researchers began: “You are a clinical expert in identifying immune-related adverse events caused by immune checkpoint inhibitors …” and included a list of six immune checkpoint inhibitors alongside numerous associated adverse events.

This prompt was applied to clinical notes from multiple sources. These included records from 100 people treated at Vanderbilt Health, 70 people from the University of California, San Francisco, and 272 people enrolled in seven Roche-sponsored clinical trials.

Study design and model performance

The research team evaluated three models: GPT-3.5, GPT-4, and GPT-4o, with GPT-4o demonstrating the strongest overall performance.

To assess accuracy, the investigators used F1 scores, a metric that balances false positives and false negatives. Scores range from zero to one, with values above 90 percent considered excellent. A score of 80 percent or higher may be sufficient for use in automated clinical decision support systems.

At the patient level, GPT-4o achieved average F1 scores of 56 percent for Vanderbilt Health data, 66 percent for University of California, San Francisco data, and 62 percent for Roche clinical trial data. The models showed a consistent tendency to overpredict the presence of immune-related adverse events.

When analysing individual clinical notes, the model achieved an average F1 score of 57 percent across 667 notes from Vanderbilt Health, evaluating 17 different adverse events.

Implications for clinical practice and research

The findings suggest that large language models can play a role in identifying drug safety signals, even without task-specific training data.

“Manual patient chart abstraction for monitoring the safety and efficacy of drugs already at market requires tremendous resources and puts a drag on the pace of discovery in precision medicine. And that’s especially true with immune checkpoint inhibitors, where the adverse events are so varied. If zero-shot learning with LLMs could help with these notes, it could significantly reduce time and costs for all concerned,” said the report’s corresponding author, Cosmin Bejan, PhD, assistant professor of Biomedical Informatics at Vanderbilt Health.

However, the current level of performance falls short of what would be required for clinical decision support.

“These results show that zero-shot learning with a powerful LLM is useful for detecting these adverse events,” Bejan said. “This performance does not rise to the level required for clinical decision support, but the method could be valuable for automated irAE extraction across multiple sites, potentially speeding discovery and enhancing the safety and effectiveness of cancer immunotherapies.”

Wider research context

The study involved collaboration among multiple researchers at Vanderbilt Health, including Yaomin Xu, PhD, Eric Mukherjee, MD, PhD, Matthew Krantz, MD, Douglas Johnson, MD, MSCI, Elizabeth Phillips, MD, and Justin Balko, PhD. Funding support was provided in part by the National Institutes of Health.

Related research further highlights safety concerns associated with immune checkpoint inhibitors. In a research letter published in JAMA Oncology, Mukherjee, Phillips, and colleagues used logistic regression analysis of adverse event reports from the Food and Drug Administration. They confirmed that these therapies are independently associated with an increased risk of Stevens-Johnson syndrome and toxic epidermal necrolysis, which are severe and potentially life-threatening skin reactions. The study also found that this risk may be linked to exposure to human leukocyte antigen–restricted drugs.

Conclusion

Large language models represent a promising tool for extracting clinically meaningful insights from unstructured health data. While their current performance limits direct clinical application, their ability to operate across multiple datasets without task-specific training suggests potential for supporting large-scale pharmacovigilance efforts. As these models continue to improve, they may contribute to more efficient and comprehensive monitoring of drug safety in clinical practice.

Read More