Untitled attachment

The moral panic over ChatGPT has led to confusion because people often speak of it as “artificial intelligence.” Is ChatGPT properly described as artificial intelligence? Should we call it that? Professor Sussman of the MIT Artificial Intelligence Lab argues convincingly that we should not.

Normally, “intelligence” means having knowledge and understanding, at least about some kinds of things. A true artificial intelligence should have some knowledge and understanding. General artificial intelligence would be able to know and understand about all sorts of things; that does not exist, but we do have systems of limited artificial intelligence which can know and understand in certain limited fields.

By contrast, ChatGPT knows nothing and understands nothing. Its output is merely smooth babbling. Anything it states or implies about reality is fabrication (unless “fabrication” implies more understanding than that system really has). Seeing a correct answer to any real question in ChatGPT output is folly, as many have learned to their dismay.

That is not a matter of implementation details. It is an inherent limitation due to the fundamental approach these systems use.

Here is how we recommend using terminology for systems based on trained neural networks:

“Artificial intelligence” is a suitable term for systems that have understanding and knowledge within some domain, whether small or large.
“Bullshit generators” is a suitable term for large language models (“LLMs”) such as ChatGPT, that generate smooth-sounding verbiage that appears to assert things about the world, without understanding that verbiage semantically. This conclusion has received support from the paper titled ChatGPT is bullshit by Hicks et al., (2024).
“Generative systems” is a suitable term for systems that generate artistic works for which “truth” and “falsehood” are not applicable.

Those three categories of jobs are mostly implemented, nowadays, with “machine learning systems.” That means they work with data consisting of many numeric values, and adjust those numbers based on “training data.” A machine learning system may be a bullshit generator, a generative system, or artificial intelligence.

Most machine learning systems today are implemented as “neural network systems” (“NNS”), meaning that they work by simulating a network of “neurons”—highly simplified models of real nerve cells. However, there are other kinds of machine learning which work differently.

There is a specific term for the neural-network systems that generate textual output which is plausible in terms of grammar and diction: “large language models” (“LLMs”). These systems cannot begin to grasp the meanings of their textual outputs, so they are invariably bullshit generators, never artificial intelligence.

There are systems which use machine learning to recognize specific important patterns in data. Their output can reflect real knowledge (even if not with perfect accuracy)—for instance, whether an image of tissue from an organism shows a certain medical condition, whether an insect is a bee-eating Asian hornet, or whether a toddler may be at risk of becoming autistic. Scientists validate the output by comparing the system's judgment against experimental tests. That justifies referring to these systems as “artificial intelligence.” Likewise the systems that antisocial media use to decide what to show or recommend to a user, since the companies validate that they actually understand what will increase “user engagement,” even though that manipulation of users may be harmful to them and to society as a whole.

Businesses and governments use similar systems to evaluate how to deal with potential clients or people accused of various things. These evaluation results are often validated carelessly and the result can be systematic injustice. But since it purports to understand, it qualifies at least as attempted artificial intelligence.

As that example shows, artificial intelligence can be broken, or systematically biased, or work badly, just as natural intelligence can. Here we are concerned with whether specific instances fit that term, not with whether they do good or harm.

There are also systems of artificial intelligence which solve math problems, using machine learning to explore the space of possible solutions to find a valid solution. They qualify as artificial intelligence because they test the validity of a candidate solution using rigorous mathematical methods.

When bullshit generators output text that appears to make factual statements but describe nonexistent people, places, and things, or events that did not happen, it is fashionable to call those statements “hallucinations” or say that the system “made them up.” That fashion spreads a conceptual confusion, because it presumes that the system has some sort of understanding of the meaning of its output, and that its understanding was mistaken in a specific case.

That presumption is false: these systems have no semantic understanding whatsoever.

Public

Notices where this attachment appears