AI and all that

Two types of AI mistakes

Philosophy of AI is all the rage now. Lots of university administrators are keen to get philosophy departments to hire folks who specialize in AI. Utah’s Governor is throwing around millions of dollars to help Utah’s universities to hire anyone who can add some AI to the mix. When university administrators and State Governors talk about AI, they have one thing in mind: Large Language Models (LLMs) that we see and use everyday, Claude, Chat GPT, Google AI etc. I have long had an interest in the history and philosophy of AI. I have written about it a bit over the years and taught many classes about it. The AI I was mostly concerned with in these contexts was not our current omnipresent LLMs. Here is a little reflection on the current style of AI.

Two types of AI mistakes.

Google AI just told me that a particular weighing scale displayed in small increments of either .1 lb or .5 kg (.5 kg is 1.1 lb). I pointed out that this made no sense and the AI replied that it was sorry and I was right that the claim did not make sense. The AI made a very interesting mistake involving logic, arithmetic and the use of both to make comparisons between different scales of measurement. This is also the sort of mistake that cannot be tolerated in most standard computational systems and most of them, as a matter of their underlying logic, cannot make such mistakes.

Such a mistake in a regular computational system would indicate that the system was completely broken. I used to hand out cheap, simple calculators for students to use in my probability class exams. A simple calculator is a small scale computational system that operates by the rules of arithmetic and solves math problems that require addition, subtraction, multiplication and division. The cheap calculators I handed out broke often and students knew when they had broken, because the devices started churning out bad arithmetic. The busted ones went in the bin and were replaced. A mistake in arithmetic for a simple calculator indicates a breakdown of the whole system.

Let’s compare this kind of mistake with another kind of mistake Google AI made responding to the same research prompt. Google AI said that Conair, who makes “Thinner” scales, did not make a scale that measured in .5 lb increments, they only made scales that measured in smaller increments. I refined my search to specify older scales and Google AI repeated the claim. This claim is false, as the scale in our bathroom is an older “Thinner” scale that measures in .5 lb increments. The AI wasn’t up on the history of “Thinner” scales. This kind of mistake could be rectified by loading in tons more training data sets on Conair brand weighing scales. Doubtless this would be at great expense due to this exercise requiring a new data center, which would use up all the water in the State it was in etc. etc. Anyway, back to the plot. The two mistakes are different in kind and require different kinds of fixes.

Only historians of Conair scales, or me apparently, would be upset with an AI making incorrect claims about early incarnations of that technology. The general point here is that we should expect Google AI etc. to make such mistakes. The various AI’s make mistakes about history, literature, times of sunsets in various parts of the world, basic scientific facts and on and on. You should expect this and so shouldn’t rely on them to answer such questions or, if you do, always check their answers. I am only bothered by this when the very same AI systems are touted by education specialists. “We should bring these AIs into schools or universities” they say. “AI will improve the educational experience for everyone” they say. This is usually code for our State or educational establishment has just dumped millions into a particular company’s AI system and now we all have to use it. This is a similar situation to universities constantly switching between online learning and support tools, Canvas, Blackboard etc., as they were convinced to buy the next new shiny one.

The logic and math mistakes are the ones that really bother me. (These issues also really bother Gary Marcus, see e.g. this article of his.) We have so many years of computing excellence behind us producing better and better algorithms, solving bigger and more complex problems but we now have to settle for AI that can’t do basic math and logic. Come on people! Try blending some good old fashioned problem solving techniques into your AI systems. The rules governing such problem solving are the very basis of all computing. LLMs can do some good stuff when suitably trained but why not give them a support system based in logic and math to deal with straightforward reasoning problems. That these AIs are not equipped to solve these problems should be worrying and perhaps even shocking to people. For those who aren’t worried, I have a bin full of broken calculators you can use to calculate the tip on your restaurant bill.

Postscript

I started writing this yesterday and finished it up today during that time news just broke that a major philosophy journal has published a paper written by a janky six fingered plagiarism robot (thanks Jeff Tiedrich ). Also, I have written something about LLMs before with Alex Grzankowski and Patrick Forber.