Tags
AI, Artificial Intelligence, chatgpt, Democracy, politics, technology, truth, USA
At first, they seemed as though they were simply errors. In fact, they were the types of errors you’d expect an AI system to make if it’s “intelligence” were based on a fairly uncritical amalgam of ingesting a vast amount of written material. The strains of the Beatles Nowhere Man reverberate in my head. I no longer thing the mistakes are “innocent” mistakes. They are part of an overall effort to destroy human intelligence. That does not necessarily mean that some evil person somewhere said to themselves: “Let’s destroy human intelligence. Then, people will be more willing to accept AI as being intelligent.” It could be that the attempt to destroy human intelligence is more a side-effect of unrelenting greed and hubris than a well thought-out plot.
What errors am I talking about? The first set of errors I noticed happened when my wife specifically asked ChatGPT about my biography. Admittedly, my name is very common. When I worked at IBM, at one point, there were 22 employees with the name “John Thomas.” Probably, the most famous person with my name (John Charles Thomas) was an opera singer. “John Curtis Thomas” was a famous high jumper. The biographic summary produced by ChatGPT did include information about me—as well as several other people. If you know much at all about the real world, you know that a single person is very unlikely to hold academic positions in three different institutions and specializing in three different fields. ChatGPT didn’t blink though.
A few months ago, I wrote a blog post pointing out that we can never be in the same place twice. We’re spinning and spiraling through the universe at high speed. To make that statement more quantitative, I asked my search engine how far the sun travels through the galaxy in the course of a year. It gave an answer which seemed to check out with other sources and then—it gratuitously added this erroneous comment: “This is called a light year.”
What?
No. A “light year” is the distance light travels in a year, not how far the sun travels in a year.
What was more disturbing is that the answer was the first thing I saw. The search engine didn’t ask me if I wanted to try out an experimental AI system. It presented it as “the answer.”
But wait. There’s more. A few hours later, I demo’ed this and the offending notion about what constituted a light year was gone from the answer. Coincidence?
A few weeks later, I happened to be at a dinner and the conversation turned to Arabic. I mentioned that I had tried to learn a little in preparation for a possible assignment for IBM. I said that, in Arabic, verbs as well as nouns and adjectives are “gendered.” Someone said, “Oh, yes, it’s the same in Spanish.” No, it’s not. I checked with a query—not because I wasn’t sure—but in order to have “objective proof.” To my astonishment, when I asked, “Which language have gendered verbs, the answer came back to say that this was true of Romance languages and Slavic languages. It not true of Romance languages. Then, the AI system offered an example. That’s nice. But what the “example” actually shows is the verb not changing with gender. The next day, I went to replicate this error and it was gone. Coincidence?
Last Saturday, at the “Geezer’s Breakfast,” talk turned to politics and someone asked whether Alaska or Greenland was bigger. I entered a query something like: “Which is bigger? Greenland or Alaska.” I got back an AI summary. It compared the area of Greenland and Iceland. Following the AI summary were ten links, each of which compared Greenland and Iceland. I turned the question around: “Which is larger? Alaska or Greenland?” Now, the AI summary came back with the answer: “Alaska is larger with 586,000 square miles while Greenland is 836,300 square miles.”
What??
When I asked the same question a few minutes later, the comparison was fixed.
So…what the hell is going on? How is the AI system repairing its answers? Several possibilities spring to mind.
There is a team of people “checking on” the AI answers and repairing them. That seems unlikely to scale. Spot checking I could understand. Perhaps checking them in batch, but it’s as though the mistakes trigger a change that fixes that particular issue.
Way back in the late 1950’s/early 1960’s, Arthur Lee Samuel developed a program to play checkers. The machine had various versions that played against each other in order to improve play faster than could be done by having the checker player play human opponents. This general idea has been used in AI many times since.
One possible explanation of the AI self-correction is that the AI system has a variety of different “versions” that answer question. For simplicity of explanation, let’s say there are ten, numbered 1 through 10. Randomly, when a user asks a question, they get one version’s answer; let’s say they get an answer based on version 7. After the question is “answered” by version 7, its answer is compared to the consensus answer of all ten. If the system is lucky, most of the other nine versions will answer correctly. This provides feedback that will allow the system to improve.
There is a more paranoid explanation. At least, a few years ago, I would have considered it paranoid because I like to give people the benefit of the doubt and I vastly underestimated just how evil some of the greediest people on the planet really are. So, now, what I’m about to propose, while I still consider it paranoid, is not nearly so paranoid as it would have seemed a few years ago.
MORE! MORE! MORE!
Not only have I discovered that the ultra-greedy are short-sighted enough to usher in a dictatorship that will destroy them and their wealth (read what Putin did and Stalin before him), but I have noticed an incredible number of times in the last few years where a topic that I am talking about ends up being followed within minutes by ads about products and services relevant to that conversation. Coincidence?
Possibly. But it’s also possible that the likes of Alexa and Siri are constantly listening in and it is my feedback that is being used to signal that the AI system has just given the wrong answer.
Also possible: AI systems are giving occasional wrong answers on purpose. But why? They could be intentionally propagating enough lies to make people question whether truth exist but not enough lies to make us simply stop trusting AI systems. Who would benefit from that? In the long run, absolutely no-one. But in the short term, it helps people who aim to disenfranchise everyone but the very greediest.
Next step: See whether the AI immediately self-corrects even without my indicating that it made a mistake.
Meanwhile, it should also be noted that promulgating AI is only one prong of a two-pronged attack on natural intelligence. The other prong is the loud, persistent, threatening drumbeat of false narrative excuses for stupidity that we (Americans as well as the world) are supposed to take as excuses. America is again touting non-cures for serious disease and making excuses for egregious security breaches rather than admitting to error and searching for how to ensure they never happen again.
————-
Sample Chapter from Turing’s Nightmares: A Mind of its Own
Sample Chapter from Turing’s Nightmares: One for the Road
Sample Chapter from Turing’s Nightmares: To Be or Not to Be
Sample Chapter from Turing’s Nightmares: My Briefcase Runneth Over
