I have written on this topic at some length across several lines of inquiry. You can see my work on it here https://robertvanwey.substack.com/s/tech. In short, AI remains seriously deficient in comprehending contextual clues. Numerous examples exist where AI outputs lead to erroneous results from this problem. The technology is excellent for pattern identification, and analyzing large datasets, but input bias, hallucinations, and other issues still plague the product. Every AI output should be analyzed by humans before it is applied to any sensitive process, particularly medical. This is especially true in that LLMs, at least, still rely on probabilistic calculations, which potentially overlook contextual elements critical to a specific individual's circumstance. In addition, over-reliance on AI implicates legal issues not yet litigated or even thought of, so uncritical users proceed at their own peril.

Pros:

Cons:

Peter, thank you for this excellent and timely question!

I have read the previous commentary and completely agree.

I offer the following commentary that is a spin off from a casual conversation that I recently had with a group of doctoral nursing students following a lecture. One of the topics in the lecture focused on implementation AI and Deep Machine Learning at the health care delivery system level in electronic health record platforms.

I challenged the students to go back and read Dr. Issac Asimov's science fiction short story "Runaround" (circa 1942) even if they had already read it or were familiar with it. The challenge is to read what was "state of the art" 80 years ago, but is still right on the bleeding-cutting edge of technology today (and will remain so for the for-seeable future, in my opinion), with this in mind as part of the homework:

Substitute reconstruct your mental concept of "Robot" to be construed as "Artificial Intelligence" where ever Dr. Asimov is describing a machine or a mechanized device that is "replacing human effort" with electromechanical effort or work. "Runaround" is the first mention of Asimov's Three Laws of Robotics. I think it is way past time that we (data scientists) give it the old college try and start talking about AI in terms of those golden Three Laws.

Here what my own attempt at the homework assignment looked like. I don't suggest homework to interested students unless I have already done the assignment myself several times to be aware of way that it could go wrong.

Dr. Isaac Asimov's Three Robotic Laws (1942) as interpreted by a data scientist (me) around 80 years later:

Having been professionally involved in the fundamental development and design of the first official Codified Nursing Informatics Data Vocabulary accepted into SNOMED, I have observed that all of the issues (problems, errors, mistakes, hazards, harms, etc.) that have emerged from the intersection of health care and AI, are addressed by Dr. Asimov's Three Laws, when translated and configured to fit the health care delivery computational informatics electronic health record, and the environment of care, from which they are generated as a confluent of continuous data.

In my opinion, we (health care informatics biomedical data scientists) need to seriously consider how to design and incorporate the Asimov Laws and "hard wire" them into everything we design AND begin measuring and monitoring who, what, when, where, how, how much and how often preventable adverse events occur as a result of ignoring or accidentally or intentionally violating them. Dr. Asimov later added an additional or Fourth Law (aka the "zeroth law") that expanded the hierarchical scope of the three laws to apply to "All Humanity" as well as the individual "Human Being".

4. An AI may not harm humanity, or by inaction, allow humanity to come to harm.

Yes, these thoughts are controversial and could potentially be quite disturbing on many levels to those who have never encountered them before. But so it is with many serious conversations among wise and mature adult human beings, especially scientists. I appeal to your good scientific judgement in considering what I am attempting to communicate in a rational manner, that could seriously safe lives and prevent injuries and harm. These concepts have great potential for Patient Safety if taken seriously into consideration as a baseline tenet for AI algorithm designs for the present and future.

For those of you readers of this commentary who are not familiar with the intellectual construct popular in healthcare quality improvement called the Quadruple Aim here they are

My assertion is that if Dr. Asimov's 4 Laws were seriously applied to health care in general (as well as the AI we are beginning to incorporate into it), we could reasonably expect to functionally increase achievement of the Quadruple Aim.

This is not about achieving perfection, it is about achieving improvement without achieving a catastrophic disaster in the process of implementing "good intentions".

Richard E Gilder sir, this is a great answer! From a developer’s standpoint, I do not think the field has sufficiently progressed to properly engage with or implement Asimov’s laws. AI systems do not apprehend context in a meaningful way, and we are not all that close to achieving that. To be sure, they can emulate contextual comprehension, but that is something like a mathematical response to a philosophical problem. To illustrate, humans themselves struggle to define certain terms like “harm” in different contexts. For instance, chemotherapy ostensibly commits harm against the recipient, but toward the greater good of treating the cancer. Medical professionals understand the long-term implications despite the short-term effects manifesting in a rather discomfiting way. It is unclear how AI would reach the same understanding. Even action versus inaction can be challenging. AI would necessarily compute the odds of a result by taking or avoiding a specific action, but the results would not always reflect the desired outcome. This, however, would educate the AI’s learning process and be incorporated into the next scenario (much like how LLMs train to select the “right” word in response to an input). Just as copying a very slightly flawed document will, over time, lead to the production of a deeply flawed one, AI’s evolution might follow a similar path in its action v. inaction decision making. I am further reminded of the thought experiment of a trolley hurdling down a hill with no brakes. At a coming junction, on one track is a single person, on the other stand several. There is a switch the AI can manipulate to determine the track the trolley will follow. Using a cold calculus, the AI would likely turn the trolley toward the single person, thereby preserving the 5. But a human may make a different decision depending on other factors: 1) what if the lone person is a child or immobilized person? 2) what if the lone person is a ruthless dictator and the 5 are his victims? My point here is that Asimov’s laws carry deeply complex underlying philosophical quandaries that all but obliterate any "objective" programming, no matter how strongly we desire to develop it.

All that said, I think you struck directly to the core of what we can do. Achieving the Quadruple Aim is well within the current computational and programming capability. Successfully implementing AI into the healthcare sphere requires little more than careful testing, continued oversight, and proper motivations. In my column that I write for the general public, this has been the source of my primary complaint about AI (and other tech). We cannot sacrifice quality or compromise on the moral and philosophical aims of the industry to cut corners through the incorporation of under-developed AI in order to rush it to market. It is as much a burden of healthcare professionals to ensure that adopting new technologies is to the benefit of patients and the industry as it is to developers to produce products that seek to--and are able to--achieve the same ends. To date, I do not feel AI companies have given the necessary seriousness toward that principle, or even desire to. We do not need perfect results as you correctly stated, but imperfect cannot equal deeply flawed or damaging to large groups of people. We need only follow your wonderfully worded closing statement: “it is about achieving improvement without achieving a catastrophic disaster in the process of implementing ‘good intentions’.”

Some key aspects to consider when defining the safe limit of AI interference:

1. **Clinical Decision-Making**: AI systems can assist in clinical decision-making, but the ultimate responsibility and accountability should remain with human healthcare professionals. The safe limit should ensure that AI systems provide support, recommendations, and insights to aid human decision-making rather than replacing human judgment entirely.

2. **Transparency and Explainability**: AI systems should be designed to be transparent and provide explanations for their recommendations or decisions. Healthcare professionals and patients need to understand the reasoning behind AI-generated outputs to trust and validate the system's outcomes. Clear explanations can also help identify any biases or errors in the AI system's functioning.

3. **Continuous Monitoring and Validation**: AI systems should undergo rigorous testing, monitoring, and validation to ensure their performance, accuracy, and safety. Regular assessments and audits should be conducted to detect and rectify any issues, biases, or unintended consequences that may arise during the system's operation.

4. **Data Privacy and Security**: Safeguarding patient data is crucial when using AI in healthcare. Clear guidelines and regulations should be in place to protect patient privacy, ensure data security, and prevent unauthorized access or misuse of sensitive healthcare information.

5. **Ethical Considerations**: The safe limit should encompass ethical guidelines and principles that govern AI use in healthcare. Ethical considerations include fairness, equity, accountability, and ensuring that AI systems do not perpetuate or amplify existing biases, discrimination, or inequalities in healthcare delivery.

6. **Human-Machine Collaboration**: Establishing a collaborative framework between AI systems and healthcare professionals is essential. The safe limit should encourage effective collaboration, where AI systems complement and augment human expertise, allowing healthcare professionals to focus on critical and complex tasks that require human judgment, empathy, and interpersonal skills.

7. **Regulatory Framework**: Robust regulations and governance frameworks should be in place to oversee AI systems in healthcare. These frameworks should address issues such as system validation, accountability, transparency, and patient safety. Collaboration between policymakers, healthcare professionals, AI experts, and ethicists is crucial for developing effective regulations.

Hope it helps:credit AI