ChatGPT Health and Why Foundation Models Still Cannot Crack Healthcare Alone
Abstract
OpenAI launched ChatGPT Health in January 2025, partnering with Color Health and bwell Connected Health while collaborating with major health systems to build domain-specific clinical AI. The move signals that even the most sophisticated foundation model companies recognize healthcare requires specialized training, regulatory expertise, and clinical validation beyond what general-purpose LLMs provide. Key points include:
- ChatGPT Health targets clinical documentation, patient communication, and care coordination through partnerships rather than direct deployment
- Color Health brings clinical validation capabilities while bwell Connected Health provides consumer engagement infrastructure and health plan relationships
- Foundation model competitors including Anthropic, xAI, and Google Gemini face strategic choices about whether to pursue similar healthcare-specific products or enable partners to build on their platforms
- The partnership model validates that healthcare AI requires domain expertise, regulatory positioning, and distribution channels that foundation model companies lack internally
- Market opportunities exist across multiple layers of the stack for specialized companies that can navigate healthcare’s unique requirements
Table of Contents
Why General Purpose Models Keep Failing in Clinical Settings
What OpenAI Actually Built and the Partnership Strategy Behind It
Color Health’s Role in Clinical Validation and Regulatory Navigation
bwell Connected Health and the Consumer Engagement Infrastructure Play
The Economics of Healthcare AI and Where Value Actually Accrues
How Anthropic Should Respond to Stay Competitive
What xAI Needs to Do If Musk Wants Healthcare Revenue
Google Gemini’s Advantages and Strategic Missteps in Healthcare
The Infrastructure Problem That Determines Everything
Competition at Every Layer and Why Vertical Wins
What This Means for Health Tech Builders and Investors
Regulatory Frameworks That Will Make or Break Clinical AI
Why General Purpose Models Keep Failing in Clinical Settings
The past two years saw hundreds of health systems and digital health companies attempt to deploy ChatGPT and similar foundation models for clinical use cases. Most of these initiatives failed or got stuck in pilot purgatory because general-purpose language models lack the reliability, specificity, and regulatory positioning required for actual clinical deployment. The gap between passing medical licensing exams in controlled settings and functioning safely within real clinical workflows proved far wider than the initial hype suggested. Health system CIOs who jumped on foundation models in 2023 spent 2024 explaining to their boards why pilot projects did not scale despite impressive demos.
The failure pattern repeated across organizations. Initial pilots showed promising results with cherry-picked use cases and heavy human oversight. Attempts to scale beyond controlled environments exposed problems with hallucinations, inconsistent outputs, inability to handle edge cases, and lack of integration with existing workflows. Legal and compliance teams raised concerns about liability exposure, regulatory ambiguity, and patient safety risks that pilot coordinators had not adequately addressed. The technology worked well enough to be interesting but not well enough to be deployable at scale without significant additional development.
OpenAI apparently learned these lessons through their own partnership discussions and pilot failures. ChatGPT Health exists as a distinct product line precisely because the company recognized that healthcare requires fundamentally different architecture, training approaches, and go-to-market strategies compared to general-purpose AI. The decision to partner with domain experts like Color Health and bwell Connected Health rather than attempting direct deployment acknowledges that foundation model companies lack the clinical expertise, regulatory knowledge, and healthcare distribution channels needed to succeed independently.
This matters enormously for health tech investors and entrepreneurs because it validates several hypotheses about how AI will actually penetrate healthcare. First, domain specificity beats generalization in clinical settings. Models need training on clinical datasets, fine-tuning for specific workflows, and validation against real-world use cases rather than academic benchmarks. Second, partnerships between AI companies and healthcare domain experts will dominate the market rather than direct plays by foundation model providers. Third, the technical challenge of building good models represents only part of the problem, with regulatory navigation, liability frameworks, and workflow integration often proving more difficult than model development itself.
What OpenAI Actually Built and the Partnership Strategy Behind It
