Dear Editor,
AI is becoming a recurring feature in medical research roundups, and this week’s roundup highlighted by Tom Nolan provides another example of AI’s expanding capabilities. That brings me to the question: could medical superintelligence be approaching?
Before we can answer that, we need to establish what it would entail and who should decide when that threshold has been crossed, because the labelling has become more topical after the recent US announcement of making superintelligence the new blanket term replacing AI [1].
In the continuum which started with AI matching humans on specific tasks, to surpassing them, to become increasingly general and agentic, superintelligence could be the potential next step [2]. But is medical AI there yet?
When we talk about superintelligence, we’re imagining AI that can outperform humans across many different areas, possibly with a high degree of autonomy [2]. But today’s AI, although does excel at specific tasks, reaching capabilities that were unimaginable a few years ago, we can only say we are on track to superintelligence but not there yet [3].
We also don’t have a clear definition. Does what AI can do today count as superintelligence? If not, what would it need to do for us to say we’ve crossed that threshold? In the medical field, that’s especially difficult to answer. Being better than humans in one area does not necessarily make a system generally superintelligent. Then there is also the question of autonomy.
In practice, we have always been reluctant to allow AI to operate autonomously when it comes to patient care [4]. This introduces an important distinction between capability and permission. An AI might become capable of diagnosing, treating, or monitoring patients, and even delivering aspects of care, but we may still choose to limit its role through regulation and human oversight due to various reasons.
So, if autonomy turns out to be a defining feature of superintelligence, medical AI may never fully satisfy that definition because we intentionally withhold permission. This creates a peculiar problem for medicine. So, where do we draw the line?
Considering how gradually we have progressed over the years to what we are today with AI, perhaps there may not be a clear line to define a threshold to begin with. So, medicine might need its own operational definition of superintelligence, which is tied to demonstrated capabilities, but subject to accountable oversight. That's tricky.
As a clinician, I want a superintelligent system to do more than outperform a specialist at one task. I’d want it to work across specialties and different clinical settings. But developers might define the threshold in technical terms, regulators in terms of risk and autonomy, and patients in terms of trust. So, who gets to decide?
In that sense, medical superintelligence may not arrive as a single technological milestone, but through a collective decision. Different stakeholders will have different priorities and thresholds. So, perhaps the definition of medical superintelligence will have to be negotiated rather than declared.
Footnotes
Competing interests: None declared
AI use: ChatGPT was used to check spelling and grammar
References
1. Inaugurating The Era Of Super Intelligence. The White House, 2026. https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the...
2. What Is Artificial Superintelligence? IBM, 2024. 14 December 2023
3. What is Superintelligence? - Artificial Superintelligence (ASI) Explained - AWS. Amazon Web Services, Inc. 2026.
4. Ethics and governance of artificial intelligence for health. World Health Organization, 2021. https://www.who.int/publications/i/item/9789240029200
Superintelligence: where do we draw the line, and who is “we”?