From Predictive AI to Generative AI

Traditional AI-enabled medical technologies are typically designed around specific, narrowly defined tasks: analyzing an image, identifying a feature, calculating a risk score, or classifying a condition.

Generative AI introduces a completely different level of complexity. Large language models (LLMs) and multimodal systems can generate new outputs based on multiple sources of information. They can summarize clinical documentation, interact conversationally, interpret diverse data streams, and support professionals across several stages of a clinical workflow.

Output Type

• Predictive / Traditional AI: Fixed outputs (classification, risk score)

• Generative AI: Dynamic outputs (text, summary, multimodal content)

Scope

• Predictive / Traditional AI: Single, narrowly defined task

• Generative AI: Multitask, adaptable across workflows

Evaluation

• Predictive / Traditional AI: Established, static validation criteria

• Generative AI: Continuous evaluation based on context & evolving models

While flexibility is one of Generative AI's greatest strengths, it is also its primary challenge. A system that produces a fixed output for a defined task can be evaluated using established metrics. A generative system may produce different responses to similar inputs, interact dynamically with users, and evolve as underlying models update.The fundamental question becomes: How do we evaluate a system whose capabilities are not always fixed?

The FDA Is Asking a Different Question

The FDA's discussion paper explores regulatory considerations including risk assessment, premarket evaluation, and postmarket monitoring for generative AI-enabled medical devices.

One particularly important concept emerging from this dialogue is evaluating AI according to its competence for the specific task it is intended to perform.

Instead of asking only: "Does the technology work?" We increasingly need to ask: "Is the technology sufficiently competent for the specific clinical task in which it is being used?"

This distinction matters enormously. An AI system helping a healthcare professional organize administrative information does not present the same risk profile as an AI system contributing to a diagnostic or treatment decision. The consequences of an incorrect output are fundamentally different.

The future of healthcare AI regulation will not rely on a single, universal definition of "good AI." Instead, risk, intended use, clinical context, and potential consequences will drive assessment frameworks.

Healthcare Cannot Treat AI Like Ordinary Software

Generative AI creates another fundamental challenge: its outputs are often probabilistic rather than deterministic. The same model can generate different responses depending on context, prompt formulation, or model version.

Conventional software engineering paradigms fall short in high-stakes clinical scenarios. Healthcare systems must evaluate beyond accuracy, performance, and usability:

• Transparency & Traceability: Understanding how inputs lead to generated outputs.

• Bias & Fairness: Ensuring equitable performance across diverse patient demographics.

• Privacy & Cybersecurity: Safeguarding patient data across conversational and generative interfaces.

• Human Oversight & Clinical Validation: Maintaining clear boundaries for clinician responsibility.

• Post-Deployment Monitoring: Tracking performance drift as models and environments evolve.

Recent research published in Nature highlights that the rapid integration of large language models into healthcare creates new responsibilities, vulnerabilities, and potential safety risks that must be systematically evaluated before widespread clinical adoption.\

Building a powerful healthcare AI system is only half the challenge. Building one that can be responsibly trusted is the other half.

The Human Professional Remains Central

A common misconception surrounding healthcare AI is that the goal is to replace medical professionals. The far more meaningful opportunity is human augmentation. A clinician brings contextual understanding, clinical intuition, communication skills, ethical judgment, and ultimate responsibility.

Artificial intelligence contributes complementary capabilities:

• Processing high volumes of unstructured clinical data

• Identifying underlying patterns across disparate sources

• Reducing repetitive administrative overhead

• Supporting rapid evidence retrieval

• Assisting with structured, personalized patient workflows

• Providing secondary decision-support signals

The most effective healthcare model is not Human versus AI it is Human Expertise + Artificial Intelligence.

The Real Value of AI May Be Invisible

The most valuable healthcare AI will not always be the technology with the most impressive demo. It will be the technology that quietly and seamlessly improves how healthcare professionals work.

Consider an environment where AI transforms fragmented patient data into structured knowledge, automates routine documentation, and presents relevant insights at exact points of care. The AI does not need to be the decision-maker; it acts as an intelligent supportive layer for the humans who are.

This is where AI transitions from a technological novelty into genuine healthcare infrastructure.

Trust Must Be Designed From the Beginning

Trust cannot be retrofitted into an AI system after deployment; it must be architected from day one. Responsible innovation requires answering critical questions upfront:

1. Data Provenance: What data is the system built and fine-tuned on?

2. Evaluation: How was the model validated across edge cases?

3. Failure Modes: What safety fallbacks occur when the system makes an error?

4. Explainability: Can clinicians understand why a specific output was generated?

5. Accountability: Who retains ultimate responsibility for clinical decisions?

6. Lifecycle Management: How is performance monitored as underlying models update?

The FDA's work reflects a broader industry transition: evaluating AI throughout its entire lifecycle rather than relying on a single, pre-deployment static check.

The Next Generation of Healthcare AI

Healthcare AI is entering a new maturity phase:

• Phase 1: Demonstrating that AI can perform impressive, isolated tasks.

• Phase 2: Proving that AI can be implemented safely, responsibly, and meaningfully in real-world clinical workflows.

Unlocking Phase 2 requires robust evaluation frameworks, strict data governance, strong cybersecurity, continuous post-market monitoring, and deep collaboration between technology developers and healthcare practitioners.

Capability alone will not determine whether AI succeeds in healthcare. Trust will.

Kemetica's Perspective

At Kemetica, we believe the future of healthcare AI is defined not by replacing human expertise, but by amplifying it. The most impactful AI solutions directly solve operational and clinical friction points, simplifying complex workflows, supporting healthcare teams, improving access to information, and enabling more personalized, efficient patient care.

Our core principles for healthcare AI:

• Human-Centered: Designed around real clinical workflows and clinician control.

• Evidence-Informed: Grounded in rigorous clinical context and data integrity.

• Responsible by Design: Built with privacy, security, and safety at the foundation.

• Practical & Scalable: Delivering immediate operational value without friction.

The central question facing our industry is not simply: "How intelligent can healthcare AI become?" It is: "How can we make increasingly intelligent AI genuinely useful, safe, and trustworthy for the people who depend on healthcare?"

That is the challenge ahead and it is far more important than intelligence itself.

Sources & Further Reading:

• U.S. Food and Drug Administration: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (FDA Discussion Paper)

• U.S. Food and Drug Administration: FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices (FDA Press Announcement)

• Nature: Safety and security of large language models in healthcare (Nature Journal)

As generative AI moves closer to clinical care, the most important question is no longer what AI can do, but whether we can trust it to do it responsibly.

Artificial intelligence is moving rapidly from research laboratories into real healthcare environments. It is already being explored and deployed across medical imaging, clinical documentation, patient communication, drug development, and decision support. But as AI systems become more capable, a new challenge emerges: how should healthcare evaluate technologies that can generate answers, adapt outputs, and perform increasingly complex tasks?

This question has taken center stage recently. On August 18, 2026, the U.S. Food and Drug Administration (FDA) published a discussion paper focused specifically on the regulation of Generative AI-enabled medical devices, asking stakeholders for feedback on how these technologies should be assessed throughout their lifecycle.

The timing is significant. Healthcare AI is no longer simply about pattern recognition; it is increasingly about reasoning, generating, interacting, and supporting decisions. And that changes everything.