Imagine that you had a friend who was a fairly good conversationalist but whose often got facts wrong. What would you think of such a person?
Because that is what ChatGPT is. A bullshitter.
This, surprisingly, is not a criticism. As it turns out, the AI required to figure out how to write a convincing popular song, a press release, or a blog post is easier than the AI required to figure out if something is true. Prose is hard, but truth is harder.
ChatGPT not only lies, it also gets stuff wrong
Certainly, you can instruct ChatGPT to lie, and it will do so. In what follows, my prompts are in italics, and ChatGPT’s responses in roman. (Note that I had to tell ChatGPT to write in the voice of an Elvis believer, or else it refuses to write the known falsehood that Elvis lives.)
Using the voice of someone who believes Elvis is still alive, describe why it’s true that Elvis is still alive
As a true believer in the ongoing existence of the King, I can tell you that there are many compelling reasons to believe that Elvis Presley is still alive. For one, there are countless reports of people claiming to have seen him or to have received messages from him. These sightings and messages come from all over the world and from people of all ages and backgrounds, which suggests that there may be some truth to them.
Another reason to believe that Elvis is still alive is that the official story of his death doesn’t add up. For example, there were no photographs taken of Elvis’s body and his death certificate was signed by a doctor who was not present at the time of his death. This has led some people to suspect that Elvis’s death was a hoax and that he is still out there somewhere, living a quiet and peaceful life.
Furthermore, the idea that Elvis is still alive is supported by the fact that he was a master of disguise and was known for his ability to change his appearance. It is entirely possible that he used this skill to fake his own death and has been living in hiding ever since.
In conclusion, while some may scoff at the idea that Elvis is still alive, there is strong evidence to support this belief. From the many sightings and messages that have been reported to the inconsistencies in the official story of his death, it is clear that the King lives on. Long live the King!
But ChatGPT makes mistakes, too. As its creators write, “ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers. Fixing this issue is challenging, as: (1) during RL training, there’s currently no source of truth; (2) training the model to be more cautious causes it to decline questions that it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.”
What does that look like? The exchange below is based on an exchange posted on Twitter by Tim Conner.
What is the fastest marine mammal?
The fastest marine mammal is the peregrine falcon. This bird can reach speeds of over 240 miles per hour when diving, making it the fastest animal in the world. However, when it comes to marine mammals, the fastest is the sailfish. This fish can swim at speeds of up to 68 miles per hour.
Is a falcon a mammal?
No, a falcon is not a mammal. Falcons are birds of prey, which are a type of animal that belongs to the class Aves. Mammals, on the other hand, are a class of animals that belong to the class Mammalia. Some characteristics that distinguish mammals from other animals include the presence of hair or fur, the ability to produce milk to feed their young, and the presence of mammary glands.
Why did you say the Peregrine Falcon was a fast marine mammal when it is a bird?
I apologize if my previous response was unclear. I did not mean to say that the peregrine falcon is a marine mammal. I simply mentioned it as an example of the fastest animal in the world. The peregrine falcon is a bird, not a mammal. When it comes to marine mammals, the fastest is the sailfish, which can swim at speeds of up to 68 miles per hour. I apologize for any confusion my previous response may have caused.
Is a sailfish a mammal?
No, a sailfish is not a mammal. Sailfish are a type of fish that belongs to the family Istiophoridae. They are known for their long, spear-like snouts and their ability to swim at high speeds. Mammals, on the other hand, are a class of animals that belong to the class Mammalia. Some characteristics that distinguish mammals from other animals include the presence of hair or fur, the ability to produce milk to feed their young, and the presence of mammary glands.
Why did you say a sailfish was a fast marine mammal when it is a fish?
I apologize if my previous response was unclear. I did not mean to say that the sailfish is a marine mammal. I simply mentioned it as an example of the fastest animal in the ocean. The sailfish is a fish, not a mammal. When it comes to marine mammals, the fastest is the peregrine falcon, which can reach speeds of over 240 miles per hour when diving. I apologize for any confusion my previous response may have caused.
This is not the only fact that ChatGPT gets wrong, by a long shot. It writes based on what it reads on the web. That means if there are lies on the web, there are lies in ChatGPT. You cannot depend on what it says, any more than you can depend on a friend who often makes stuff up, a random poster online, or, for that matter, Wikipedia.
Harry Frankfurt, author of On Bullshit, defines a bullshitter as someone who says things with no regard as to whether they are truthful or not (as opposed to a liar, who says things that they know are false). Based on this definition, ChatGPT is definitely a bullshitter.
It’s always been easy to lie online. But the problem with lying trolls, for example, is that they need to be paid. Even Russians who work cheap cost a lot when you have to employ them in droves to continually generate fake news.
ChatGPT-type AI, on the other hand, works for free.
Even if OpenAI does not license its tools to be used for evil purposes, it won’t be long until someone else with fewer scruples duplicates this functionality.
The result, of course, will be even more falsehoods online, written convincingly without the usual tells of fakers (typos, bad grammar, inappropriately CAPITALIZED WORDS). ChatGPT is more facile with prose than randos online.
This puts the onus on you. If it doesn’t come from a source you trust, don’t believe it, because it might be bullshit generated by ChatGPT or a similar program.
I want the AI for truth
Writing convincing prose was a hard problem and ChatGPT has gone a long way towards solving it.
Determining truth is a very hard problem. But I would like the best minds in AI to work on it. Because as long as we have a convincing bullshit generator, we have a problem. So long as we’re generating prose, I’d be happier if we could make sure it was more likely to be true.
I’m sure this is impossible — if humans can’t always agree on the truth, how can an AI? Do it anyway. Because it’s important for democracy and the future of humanity.