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) clinical decision support (CDS) has the potential to be transformative: it can raise diagnostic and treatment quality, speed workflows, and expand access to expertise — but it also creates substantial clinical, economic, legal, and distributional risks (errors, bias, liability, deskilling, perverse incentives). The net effect on health outcomes and health-sector economics will depend heavily on system design (human-in-the-loop vs. autonomous), evaluation and governance, payment incentives, and how deployment shapes clinician labor and firm market power.

Key Points

• Potential benefits
  • Improved diagnostic and treatment suggestions through synthesis of patient data and medical knowledge; can reduce missed diagnoses and standardize care where evidence is clear.
  • Time savings for clinicians (faster charting, literature summarization, guideline retrieval), potentially increasing capacity and access.
  • Better access in low-resource settings via decision-support for non-specialists or overburdened clinicians.
  • Personalization: tailored care pathways and risk predictions that integrate multimodal data (notes, imaging, labs).
• Major risks and harms
  • Hallucinations and incorrect recommendations that appear plausible; these can cause patient harm if trusted unchecked.
  • Bias and inequities if training data underrepresent groups or reflect historical disparities; harms can be amplified when models scale.
  • Liability and accountability gaps: unclear who is responsible for AI-suggested errors (vendor, hospital, clinician).
  • Deskilling: overreliance may erode clinician judgement over time, raising systemic vulnerability.
  • Data privacy and security risks given high-value medical data and external cloud services.
• Economic consequences
  • Productivity vs. cost: GenAI can reduce clinician time per case (productivity gains) but may increase utilization (more tests/treatments) if it lowers thresholds for intervention or incentivizes revenue-generating services.
  • Labor effects: substitution in some tasks (documentation, triage) and complementarity in others (complex decision-making). Net employment effects in health care are ambiguous and vary by role.
  • Market structure: large AI firms with proprietary models and data could concentrate market power, raising prices for hospitals and shaping standards of care.
  • Payment and incentives: reimbursement rules (fee-for-service vs. capitation) will influence whether cost savings are realized or offset by increased service volume.
• Implementation sensitivity
  • Safety and benefit hinge on deployment details: user interface, real-time feedback, uncertainty quantification, calibration, and how recommendations are presented (strong vs. suggestive).
  • Human factors: training, trust calibration, and workflows determine whether clinicians accept, override, or ignore suggestions.
• Evidence maturity
  • Early evaluations show promise in controlled tasks and vignettes; high-quality real-world randomized trials and longitudinal studies are limited but essential.

Data & Methods

• Typical data sources
  • Electronic health records (structured data, clinical notes), diagnostic images, genomics, claims data, and prospective trial data.
  • Simulated patient vignettes and standardized case sets for model benchmarking.
• Common evaluation methods
  • Retrospective validation: compare model outputs to historical chart-verified diagnoses or guideline-concordant actions.
  • Prospective randomized controlled trials (RCTs): clinicians randomized to use vs. not use GenAI CDS to observe patient outcomes, utilization, and clinician behavior.
  • A/B testing and stepped-wedge deployments in health systems to estimate causal effects on workflow and utilization.
  • Human-AI interaction studies: qualitative and quantitative assessments of clinician trust, override rates, and cognitive load.
  • Economic evaluation: cost-effectiveness analysis, budget impact models, difference-in-differences and instrumental variables to estimate utilization and spending effects.
  • Robustness and fairness audits: subgroup performance, calibration across demographics, stress testing for adversarial inputs.
• Key metrics
  • Clinical: diagnostic accuracy, guideline concordance, time to correct diagnosis, morbidity/mortality, adverse events.
  • Process: clinician time per patient, documentation time, rates of tests and referrals, override/acceptance rates.
  • Economic: per-patient cost, total system spending, return on investment, labor hours, distributional impacts across populations.
• Gaps in evidence
  • Few large-scale RCTs showing direct patient outcome improvements; scarcity of long-term studies on deskilling and systemic effects.
  • Limited public datasets and reproducible evaluations for latest generative models; vendor lock-in constrains independent audit.

Implications for AI Economics

• Productivity and health spending
  • If GenAI CDS substitutes routine clinician time and reduces unnecessary testing, it can lower costs and expand capacity. But if it increases diagnostic sensitivity without specificity (or incentivizes additional services), spending could rise.
  • Economic evaluations should model both direct efficiency gains and induced demand effects.
• Labor market impacts
  • Task-based effects: jobs will be reshaped (fewer routine documentation and triage tasks; more oversight, complex decision-making, and patient communication).
  • Wage effects may be heterogeneous: some specialties face downward pressure on paid hours, while value-added skills (AI oversight, complex diagnostics) may command premia.
• Market structure and innovation incentives
  • Proprietary models trained on large clinical datasets can create high entry barriers, leading to concentration among a few platform firms and chained lock-in with major EHR vendors.
  • Antitrust and data-access policies may be required to preserve competition and diffusion of beneficial innovations.
• Payment and regulatory policy levers
  • Reimbursement design matters: adopting outcome-based payments and bundled payments could align incentives to use GenAI for net health gains; fee-for-service may incentivize overuse.
  • Certification and post-market surveillance (like medical device regulation) should require clinical trials, real-world monitoring, and transparent performance reporting.
• Distributional and equity considerations
  • Without explicit mitigation, GenAI CDS risks widening disparities by performing worse on underrepresented groups or being unequally distributed across resource-rich vs. poor settings.
  • Subsidies, open models, or public-domain training data for underserved contexts can help diffuse benefits.
• Research and policy priorities
  • Invest in large-scale RCTs and longitudinal studies that measure clinical outcomes, utilization, and labor effects.
  • Require public reporting of subgroup performance, calibration, and safety incidents.
  • Design procurement and reimbursement to reward validated clinical improvement and equitable deployment.
  • Promote model transparency, standards for uncertainty communication, and clear liability frameworks.
  • Support workforce training for human-AI collaboration, and strategies to preserve clinician expertise.

Short checklist for stakeholders - Health systems: run rigorous pilots with outcome tracking, integrate uncertainty and explainability, and train clinicians on appropriate use and overrides. - Regulators/payers: condition reimbursement on evidence of net benefit and mandate post-market surveillance; align payment models to avoid perverse incentives. - Researchers: prioritize RCTs, economic impact studies, and fairness audits using representative data. - Vendors: provide calibration metrics, provenance of training data, documentation of limitations, and mechanisms for real-world feedback and rapid patching.

Bottom line: GenAI CDS can meaningfully improve clinical care and health-system productivity, but realizing net social benefit requires rigorous evaluation, governance to manage risks and incentives, and policies to prevent concentration and unequal distribution of gains.