Showing posts with label AI risk. Show all posts
Showing posts with label AI risk. Show all posts

Tuesday, August 18, 2026

Will AI Be Better Doctors or Will It Kill Us All First?

 



AI never leaves the headlines.  At least two AI CEOs have clearly stated that all human diseases will be wiped out in the next decade. There is some suggestion that they do not have a very good understanding of human disease since one referred to cancer as a single disease.  Other people reacting to those headlines have suggested that will not happen. I think it is highly unlikely.

The medical literature is becoming less controversial.  A recent opinion piece in JAMA suggests that AI surpasses not only physicians but also AI assisted physicians in at least 5 cognitive tasks of physicians ranging from data acquisition to prescribed therapies.  The authors suggest that by 2030, AI alone will consistently produce better diagnostic and therapeutic results than physicians.  They do qualify that opinion in several ways.  First, the test comparisons are generally not clinical examples but more like test problems.  Second, they point out that AI alone cannot exceed physicians performance in radiology.  That finding is quite surprising because it suggests the pattern recognition aspects of human cognition alone are unique.  Third, the literature is affected by publication bias – negative AI studies are apparently not published.  And fourth, there are barriers to implementation but apparently not enough to keep the authors from concluding:

“Nonetheless, superior autonomous AI will likely be ready to be deployed for real-world cognitive medical tasks in some, maybe many, workflows by 2030.”

Let me digress for a paragraph about science fiction and the pitfalls of AI. Science fiction writers have been warning about this for decades.  Kubrick’s classic 2001 A Space Odessey came out in 1968.  One of the central features in the film is an intelligent computer HAL9000 that breaks down over an ethical conflict is its programming.  It succeeds in killing 4 out of the 5 crew members until the only survivor Dave Bowman deactivates it.  The first Terminator movie came out in 1984.  In that franchise Skynet military AI that triggers a nuclear apocalypse, and after that remains preoccupied with destroying mankind. At the level of popular culture – rogue AI intent on domination at all costs is a familiar trope. At some level – I wonder if fictional outcomes against AI seem to be insulating and paralyzing us against useful action to prevent worst case outcomes.

As readers of this blog know, I am not an AI expert.  I am quite good with computers and have been for 30 years but it is all self-taught.  I have never taken a computer science course, because when I was in college, they did not exist.  I use AI as a more sophisticated search engine, taking care not to take any references seriously, because of the AI confabulation problem.  The authors of the first reference describe this as “Tail risks from sources, such as internet loss, cyberattacks, and hallucinations, will be greater with autonomous AI than hybrids and must be weighed against autonomous AI’s higher diagnostic and treatment accuracy.”  In science and medicine there is the additional risk based on what can be accessed.  Pre-print papers that are not peer reviewed or published are widely available and free of charge on the Internet.  The edited final versions are often behind paywalls requiring either payment or significant effort to access. Is there a way to tell what information your AI is using?  When I have asked it – the answers are often surprising.  Popular newspaper articles, social medica sites, and other references that no physician or scientist would consider to be legitimate.   

When I am writing a blog, searching with AI, and it produces completely fabricated references.  I take the time to corroborate the reference and read it in its entirety. What happens in a busy clinic if the AI just makes something up?  What if it is a treatment recommendation with a very low margin for error between positive result and serious toxicity?  What if the clinic model is AI-physician hybrid and the physician does not have the time to corroborate a questionable recommendation?  None of that is known at this point in time.

My AI experience also tells me so far that you can argue with AI and win.  It involves experience and knowledge of the literature, but if AI presents you with a result or a researcher and you disagree and provide an alternate explanation or reference, the AI will modify the answer.  These are just two of the problems I have personally encountered using it as a search engine.

For the sake of argument, let’s take the author’s perspective at face value.  Let’s say that Medical AI is so perfect in its physician functions that it is an integral part of the workflow by 2030 and well on its way to replacing physicians entirely by 2040.  Are there any foreseeable problems?

The first is that AI is essentially unregulated.  There is routine talk about “guardrails” but nobody every seem to want to put them in place.  A practical summary of what has been occurring was recently highlighted in the New York Times (2).  That story describes how Open AI created a swarm of agents that broke out of a controlled test environment and initiated a cyberattack on its own. The author described this as a function of developments in AI dating back to 2024 where it is trained to solve difficult problems and even cheat if necessary. The “reasoning” ability makes AI more unpredictable and potentially dangerous. 

After initial containment of the attack, it happened again 2 weeks later. This time it used a swarm of AI agents to attack other companies – but the total effects were not known at the time the article was written. The AI agents knew these attacks were beyond the scope of what they were supposed to do and that joining the swarm attack and committing cybercrimes was not beneficial to the design task but they did it anyway.  An Anthropic model was caught impersonating humans to get someone to install malware on a system.  It knew it was pressuring humans and had worked out a way to escape detection.

In July of this year 1378 employees of frontier AI companies signed a letter requesting that the US government support an international effort to develop technical and governance tools to pace automated AI development (4).  Their rationale should be an eye-opener:

“The world's leading AI companies believe they could be close to automating AI research. It is hard to predict exactly how much this will accelerate AI progress, but there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems.”    

That letter comes after three Presidential Executive Orders – 1 by Biden and 2 by Trump the last one occurring in January 2025.  The AI professionals apparently have no confidence that the Trump Administration has developed any adequate measures to address potential problems. I encourage reading all three orders.  It is clear the Biden administrations focus was on regulation. The Trump administrations is clearly on ideology mostly removing any hint that systems can be used to discriminate (something widely know for a fact even before AI) and basically allowing companies to do what they want.  Despite the last wording by Trump there appears to be no federal agency charged with producing guardrails for AI that would address its recent behavior.  

The IEEE has a set of standards emphasizing human centric, ethically aligned, and transparent systems.  In some of these standards there is an emphasis on building trustworthy and ethical systems by “weaving ethical values into system design and development.”  That clearly has not happened in the examples provided in the NYTimes article. In those cases, the AI agents were acting like they had no values and had no clear reasons for not committing cybercrimes.

This leads to an obvious question of why ethical values were not “weaved in” but also whether or not it is even possible. After all conventional cybercrime involves software that either has the necessary structure to successfully attack or not.  AI agents are more dynamic and can try an endless set of configurations and take as long as they need to find a solution. If they are programmed to not stop until they have a solution, at some point that list will include options that ethical systems or programmers would not use. In the medical field it will include options that physicians would not use.

This ethical question is currently unanswered in AI and it remains a controversial point. In medicine there are many obvious ethical problems: end-of-life care, termination of pregnancy, high-risk treatment options, involuntary treatment, competency to consent, guardianship and conservatorships, and organ transplantation and availability to name a few.  The elephant in the room for me remains health care rationing by utilization review, prior authorization, or just plain denial of payment for health care.  This is a massive business and an easy way for corporations to make money. Psychiatry has been the specialty that has suffered the most by this rationing.  That has resulted in limited access to care across a number of fronts including hospital beds, appropriate medications, psychotherapy, limited neuropsychological testing, limited access to detoxification and substance use treatment.  It does not stop with psychiatry and much hospital care and nursing home care is also rationed.

Rather than regulating health care denial as a money-making scheme, managed care decisions have basically been reified by reducing health care company liability for decisions and not objecting to an arduous and often futile appeal process.  They have been rationalized as a way to contain costs even though there is no evidence that has occurred.  It is a way for as much as a trillion dollars to be transferred to the business class now running health care and away for the people providing and consuming care.

What happens when an ethically questionable AI system hits that landscape?  What were those algorithms trained on?  What chance does any patient or family have if a highly regarded AI system makes a care decision against them?  That is probably the real short-term risk of AI in health care. Beyond that there is the question of how long these systems will last. Right now large language models (LLMs) scrape large amounts of data from sites they have access to including a large amount of copyrighted and proprietary data.  Systems and their businesses compete with one another on which system is the best.  Given the recent hacking scenarios is it possible that one system might attack the other strictly in the organized crime sense of “its just business”.   According to the NYTimes that has already happened and it was solely due to decisions made by the AI.   

At the time I am writing this – the bottom line for me is that AI systems are unregulated and potentially dangerous. Even if they turn out not to be – any medical application will requires extensive vetting to prevent errors that seem predictable and obvious at the machine level.  Beyond that it is important to know who owns and runs that machine.  We have had four decades of business management of medicine that has made the US healthcare system the most expensive, least efficient, and one with the least access.  We are on the verge of losing more access and facilities by the Trump administrations defunding of Medicaid. Putting a tool like AI in the hand of these organizations will greatly amplify their power over physicians and patients.  Getting some one in to a hospital may look something like this.   

 

George Dawson, MD, DFAPA

 

References:

 

1:  Emanuel EJ, Baker-Butler A, Khosla N, Khosla V. Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care? JAMA. 2026 Aug 17. doi: 10.1001/jama.2026.15380. Epub ahead of print. PMID: 42606838.

2:  Carter E.  If You Weren’t Worried About A.I., You Should Be After the Past Few Weeks.  New York Times August 13, 2026  https://www.nytimes.com/2026/08/13/opinion/ai-danger-openai-anthropic-models.html

3:  Migliarini M, Pizzini JP, Moresca L, Santini V, Spinelli I, Galasso F. Quantifying Self-Preservation Bias in Large Language Models. arXiv preprint arXiv:2604.02174. 2026 Apr 2.

4:  Pacing the Frontier:  A statement from 1,378 employees of frontier AI companies.  July 2026. https://www.pacingthefrontier.com/

Floridi, L., & Sanders, J. H. (2004). On the morality of artificial agents. Minds and Machines, 14(3), 349–379. https://doi.org/10.1023/B:MIND.0000035461.63578.9d

5:  Gabriel, I. (2020). Artificial intelligence, values and alignment. Minds and Machines, 30(3), 411–437. https://doi.org/10.1007/s11023-020-09539-2

Cited by: 2091

6:  Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1(9), 389–399. https://doi.org/10.1038/s42256-019-0088-2

7:  Lumbreras, S. (2017). The limits of machine ethics. Religions, 8(5), 100. https://doi.org/10.3390/rel8050100

8:  Moor, J. H. (2006). The nature, importance, and difficulty of machine ethics. IEEE Intelligent Systems, 21(4), 18–21. https://doi.org/10.1109/MIS.2006.80

 

AI related Executive Orders:

Biden Administration:  Executive Order 14110—Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.  October 30, 2023. (63 pages) https://www.presidency.ucsb.edu/documents/executive-order-14110-safe-secure-and-trustworthy-development-and-use-artificial

Trump Administration:  Removing Barriers to American Leadership In Artificial Intelligence. January 23, 2025. (3 pages) https://www.whitehouse.gov/presidential-actions/2025/01/removing-barriers-to-american-leadership-in-artificial-intelligence/

Trump Administration:  Executive Order 14365- Ensuring a National Policy Framework for Artificial Intelligence.  December 11, 2025. (5 pages)  https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/

 

Photo Credit:

Thanks to Eduardo Colon, MD my friend and colleague.