Generative AI can make clinicians faster and improve diagnostic suggestions, but real-world benefits are uncertain: hallucinations, bias, liability gaps and perverse incentives could offset gains unless deployment, payment and regulation are carefully designed.

Main Finding

Generative AI (GenAI) in UK general practice is already widely used for transcription and administrative tasks, and is increasingly being experimented with for clinical decision support. While GenAI could materially improve diagnostic accuracy, guideline tailoring, and efficiency, there is scant real‑world evidence on how clinicians actually use these tools for decision making, and substantial risks (hallucination, bias, liability, skill degradation, digital exclusion) that must be managed. The net economic and clinical impact is therefore uncertain and requires targeted evaluation and governance.

Key Points

• Current uptake and uses
  • GP use of AI rose >25% in the past two years, driven largely by ambient voice transcription/scribes.
  • Most clinicians use general-purpose chatbots (ChatGPT, Gemini, Claude, Copilot) rather than medically specialised agents.
  • Common applications: scribing, administrative automation, learning support; a minority use AI for clinical decision support, though anecdotal use is growing (sense‑checking plans, differential diagnosis, interpreting results, second opinions).
• Benefits (promise)
  • Potential to access and apply the totality of evidence, tailor guidelines to multimorbidity, and integrate records to spot patterns earlier—thus reducing unwarranted variation and missed diagnoses.
  • Integration with genomic, wearable, and exposomic data could be transformative.
  • May materially reduce administrative burden and clinical errors if properly implemented.
• Harms and limitations (peril)
  • Hallucinations and lack of true clinical understanding; transcription errors (e.g., lithium → sodium) are consequential.
  • Unclear guidance/regulation, medicolegal liability, privacy/consent concerns, bias in training data, model drift, and digital exclusion.
  • Automation bias and emerging evidence that clinician+AI teams can underperform compared with the best of either alone.
  • Risk of skill degradation and changing nature of the doctor–patient relationship; possible long‑term redundancy for higher‑order clinical tasks.
• Knowledge gaps
  • Virtually no qualitative data on real-world prompts, timing of use, how clinicians appraise and apply GenAI outputs.
  • NHS lacks a clear picture of GP needs for decision‑support AI despite policy ambitions to make AI “every doctor’s trusted assistant.”
• Recommended professional response
  • Improve AI literacy among clinicians (prompting, appraisal).
  • Proactively evaluate triadic (doctor–patient–AI) consultation models, mitigate limitations, and maximise patient benefit.
  • Treat GenAI as a tool with known flaws requiring expertise to apply.

Data & Methods

• Document type: editorial / commentary synthesising literature, recent surveys, policy documents, and author experience—no new primary empirical data.
• Sources cited include:
  • Surveys and reports (e.g., RCGP Voice survey, Nuffield Trust, BMJ Health Care Inform survey of GPs).
  • Recent literature on AI in primary care, automation bias, and clinician trustworthiness.
  • Anecdotal clinical examples supplied by the author (illustrative transcription error).
• Limitations of evidence base noted by the author: lack of qualitative and observational data on how GenAI is used in real consultations; limited clinical trial data on GenAI decision support.

Implications for AI Economics

• Economic channels affected
  • Productivity and time savings: ambient scribing and administrative automation can reduce non‑clinical workload, freeing GP time (value depends on accuracy and integration costs).
  • Diagnostic efficiency and error costs: improved detection could lower avoidable harms and downstream costs; conversely, hallucinations/automation bias could impose new error costs.
  • Labor demand and skill composition: short‑term augmentation may raise GP productivity; long‑term deskilling could alter demand for tasks and workforce training needs.
  • Market formation and pricing: demand for specialised medical AI agents, integration services, and data platforms will create new markets; vendors with EHR access may capture large rents.
  • Liability and insurance costs: ambiguity over responsibility for AI-supported decisions may raise malpractice premiums or require new indemnity schemes.
  • Equity and distributional effects: adoption skewed by clinician demographics and practice affluence risks widening inequalities in care access and outcomes.
  • Public goods and data value: linking AI to national EHRs, genomics, and wearables raises questions about data governance, monetisation, and social returns.
• Policy and financing implications
  • Need for investment in AI literacy and clinician training (public funding or conditional reimbursement).
  • Regulatory costs: certification, post‑market surveillance (model drift), and auditing will be necessary and costly.
  • Reimbursement models should clarify whether and how AI use is compensated (e.g., per‑use, bundled, outcome‑based).
  • Procurement and competition policy to avoid vendor lock‑in and ensure interoperability.
  • Malpractice / standard‑of‑care norms may shift if GenAI becomes expected practice—affecting liability economics.
• Priority research and evaluation agenda for health economists
  • Real‑world usage studies (surveys + qualitative) to map who uses GenAI, for what, when, and how outputs are applied.
  • Randomised trials and implementation studies comparing clinician care with/without GenAI decision support on clinical outcomes, errors, and workflow.
  • Cost‑effectiveness and budget‑impact analyses (including downstream costs/savings from avoided harms or introduced errors).
  • Workforce modelling to assess long‑run impacts on GP supply, task allocation, and training needs.
  • Equity analyses quantifying differential uptake and outcomes by socioeconomic status and practice characteristics.
  • Regulatory impact assessments and studies of liability/insurance market responses.
  • Market and competition analyses around data access, interoperability, and vendor concentration.
• Practical recommendations for economists and policymakers
  • Fund mixed‑methods evaluation (qualitative + quantitative) now to inform procurement and regulation.
  • Incorporate heterogeneity and learning effects in economic models (automation bias, skill decay, model drift).
  • Design procurement and reimbursement policies that incentivise validated clinical benefit, transparency, and equitable access.
  • Require post‑deployment monitoring and independent audits to detect harm and model degradation.
  • Consider public investment in medically‑trained or NHS‑controlled AI agents to reduce private‑market externalities and ensure alignment with national health priorities.

Short takeaway: GenAI offers large potential productivity and quality gains in primary care but also creates novel risk and distributional problems. Robust, timely economic evaluation, governance, and clinician upskilling are essential to ensure net social benefit.