The Perils of AI

I’ve recently read several articles describing the incredible output of openai. Samples of its work were presented as equivalent to the columns of newswriters or the advice emails by writers with large numbers of readers. So I decided to try it out.

I asked for the biographies of a number of well known and important clinical investigators. The AI either never heard of them or produced bios that always had at least one error.

Next I asked the AI to write an essay on distal renal tubular acidosis (dRTA). It came up with several paragraphs that while accurate were very superficial. When I asked about a specific form of dRTA it went off the rails and produced gibberish.

I switched to metabolic alkalosis and asked for the mechanism of gastric alkalosis. This is what resulted:

Gastric alkalosis is a condition in which the pH of the stomach becomes more alkaline than normal (pH greater than 7.45). This can occur due to several mechanisms, including:

1. Excessive acid loss: The stomach produces hydrochloric acid (HCl) to help digest food. If there is excessive loss of HCl, the pH of the stomach can become more alkaline. This can occur due to vomiting or prolonged nasogastric suction.
2. Bicarbonate administration: Bicarbonate is an alkaline substance that can neutralize acid. If bicarbonate is administered, it can neutralize the acid in the stomach and increase the pH.
3. Hypokalemia: Low levels of potassium in the blood (hypokalemia) can lead to alkalosis. This is because potassium is required for the normal function of the H+/K+ ATPase pump, which helps produce HCl in the stomach.
4. Alkaline ingestion: Ingestion of alkaline substances, such as antacids or baking soda, can also lead to gastric alkalosis.

Overall, the mechanism of gastric alkalosis involves a disturbance in the normal balance of acid and alkaline substances in the stomach, resulting in an increase in pH.

The above description is complete nonsense. The AI obviously has not taken first year medical school physiology. It might help you write a column for the the almost brain dead news media or send an email about how to market yourself, but when it comes to hard data it flunks. This doesn’t mean it won’t get better. AI can be programmed for specific tasks and do very well, but as a fount of general knowledge it’s still in grade school. When it will succeed in truly mimicking human intelligence it difficult to predict. But it’s not around the corner.

The above illustration are just tests of the AI’s knowledge. Incidentally, Google did much better than the AI. But the definition of a true AI as invented by Alan Turing, almost always inaccurately described, involves conversation, not knowledge. His test was that of an observer watching the interaction of a human and a machine. One was a computer, the other a human. The conversations were written and the observer didn’t know which was which. The observer was asked to tell which of the two was human and which machine. If he couldn’t tell the difference the machine passed the test. For a complete discussion of the Turing Test and the criticisms against it go here.

The continued development of artificial intelligence will be fascinating to observe. If it will emerge and take over the direction of the planet is something a lot of people worry about. It can’t make a bigger mess of the planet than we have. It might even treat us better than we do do each other. Even so I’ve got better things to worry about like the reliability of my automatic garage door opener and the short term supply of vanilla ice cream.