From Watson to ChatGPT Health: Why AI’s medical moment looks different now
Better models, clearer use cases, and a fight over how AI enters the system.
In a short period of time, two of the world’s most prominent AI companies have announced significant moves into healthcare, signaling the possible beginning of a new era. OpenAI introduced ChatGPT Health, a dedicated patient environment within ChatGPT, while Anthropic, the developer of Claude, announced broad collaborations with medical and scientific institutions. Neither claims, at least not at this stage, to replace doctors. But their very public, branded entry into medicine suggests that AI is no longer merely a general-purpose tool; it is positioning itself as part of the healthcare system’s core infrastructure.
To understand why such a move appears more plausible today than in the past, it is impossible not to return to the most visible failure of the previous decade: IBM Watson Health. Watson was once presented as a breakthrough system that would help diagnose cancer and guide clinical decisions. Instead, it collapsed under the weight of unrealistic expectations, inconsistent data, and the immense complexity of medical practice. The key difference between then and now is not ambition, but technology. Watson struggled to process unstructured, jargon-heavy clinical texts and fragmented medical records.
Today’s large language models excel precisely in that domain. They can parse messy visit summaries, interpret medical terminology, recognize context, and synthesize information in seconds. In this sense, the technology has finally begun to match the original promise.
Yet technological capability alone is not the central battleground. The more consequential divide lies in operating models, and here the contrast between OpenAI and Anthropic becomes clear. OpenAI is pursuing a distinctly consumer-oriented approach: a business-to-consumer model aimed directly at patients. ChatGPT Health positions itself as a new mediating layer between individuals and their medical information, helping users interpret test results, decode terminology, and arrive at doctor visits better informed. The bet is on patient empowerment and the growing expectation that individuals should have direct access to, and control over, their health data.
Anthropic, by contrast, is taking a more institutional path. While it does not ignore individual users and does operate in consumer channels through partnerships, its primary focus is business-to-business. Its tools are designed for hospitals, pharmaceutical companies, research institutes, and clinical organizations. The emphasis is on analyzing scientific literature, accelerating clinical research, and supporting medical teams, with a strong focus on safety, governance, and operational reliability. If OpenAI meets the patient in the living room, Anthropic is trying to enter through the system’s back door, management, regulation, and institutional workflows.
This divergence became especially visible in a story that circulated widely on social media. Tobi Lütke, the founder and CEO of Shopify, described receiving his annual MRI results on a USB drive, accompanied by outdated and cumbersome viewing software. Rather than wait for official interpretation or abandon the effort, he fed the file structure to Claude and asked it to write code. Within minutes, the AI generated modern, browser-based viewing software that allowed him to navigate image layers and understand the raw data, not to diagnose himself, but to access information that had effectively been locked inside a closed system.
The episode neatly illustrates what is often called the “centaur model” of human–machine collaboration. The AI does not replace the radiologist or deliver a diagnosis. Instead, it removes a technical barrier and gives individuals direct access to their own medical data. The power lies not in clinical judgment, but in mediation: transforming opaque, inaccessible information into something readable and intelligible. This represents a meaningful shift in the balance of power between patients and the healthcare system.
At the same time, the debate around AI in medicine has also taken on more extreme forms. Elon Musk recently argued that if artificial intelligence systems are expected to outperform human doctors, there is little point in continuing traditional training paths such as medical school. In his view, future AI could deliver medical care superior even to what today’s most powerful and well-connected patients receive.
Even if such claims remain far removed from current clinical reality, where AI is largely an auxiliary tool, they illustrate how the conversation has moved beyond efficiency gains to deeper questions of authority, expertise, and how doctors should be trained in the future.
Beyond familiar concerns about errors or hallucinations, one of the most sensitive issues surrounding AI’s entry into healthcare is privacy and data usage. Unlike traditional software, where files remain under the control of the user or organization, language models improve by learning from large volumes of examples. In many consumer products, conversations may be stored and used, under various anonymization and restriction mechanisms, to improve the system.
The concern is not the retention of a single medical case, but the possibility that rare or highly sensitive patterns become absorbed into a model’s general knowledge. Medical data often combines symptoms, demographics, and treatment histories in ways that can be uniquely identifying. Even statistical generalization may constitute a privacy violation. The fear is not that an AI will recognize a patient by name, but that deeply intimate information will gradually become embedded in systems used by millions.
This is where the concept of Zero Data Retention (ZDR) becomes critical. At its strictest, ZDR refers to architectures in which user data exists only for the duration of processing, is not stored in long-term logs, and is not used for model training. In effect, the AI operates with only momentary memory.
The law does not require AI companies to offer full ZDR to private consumers, and even in enterprise settings, complete technical erasure is not always achievable. However, hospitals, health plans, and insurers, operating under regulations such as HIPAA in the United States or Israel’s privacy protection laws, typically demand stringent contractual guarantees: isolated environments, encryption, access controls, and, crucially, a clear commitment that medical data will not be used to train general-purpose models.
The equation is therefore straightforward. Any company seeking to sell AI into institutional healthcare must offer a robust privacy framework, sometimes including full ZDR, sometimes functionally equivalent safeguards. Without such guarantees, AI will remain primarily a consumer-facing tool used by individuals willing to accept calculated risks, rather than becoming an integral part of routine clinical care.
