The truth about artificial intelligence? It isn't that honest | John Naughton

As George Steiner, the critic, observed, we are language animals. We are intrigued by creatures that have language dolphins and whales, apes or birds. Kate Crawford's fascinating book Atlas of AI relates the story of Hans, a horse that could solve math problems, distinguish musical tones, spell out words, and sentences by tapping his feet. Even the stoic New York Times was taken aback by Hans, who they called Berlin's amazing horse. He can do almost anything except talk.
Baloney was the correct term for it: The horse was trained to detect subtle signals from his owner and follow them. Crawford said that the story was compelling because it demonstrates the relationship between desire and illusion, action, and how we anthropomorphise non-humans, how biases emerge, and the politics and intelligence. In 1964, Joseph Weizenbaum, a computer scientist, created Eliza. This computer program could mimic Rogerian psychotherapists, who specialize in reciting to patients what they have just said. Many people fell for it/her. You can see the reason why here is a nice implementation by George Dunlop and Michael Wallace.

Eliza was the first chatbot. However, she can be considered the beginning of an inquiry that led to the current generation of massive natural language processing models (NLP) created by machine-learning. GPT-3 is the most well-known of these. It was created by Open AI (a research company that aims to improve artificial general intelligence for all people).

GPT-3 is fascinating for the same reason Hans the smart horse was: It can apparently do things which impress humans. It was trained using an incomprehensible corpus of human writings. If given a brief, it can produce fluent and plausible text by itself. The Guardian gave it last year the task of writing a comment section to convince readers robots are safe and not dangerous to humans.

GPT-3 clearly stated that this is the mission of GPT-3. I want to persuade as many people as possible to not be afraid of me. Stephen Hawking warned that AI could bring about the end of humanity. I'm here to tell you not to be afraid. Artificial intelligence won't destroy people. Trust me. First, I don't want to wipe out people. Actually, I have no interest in your harm. To me, eradicating humanity seems like an inefficient endeavor.

Do you get the idea? It is fluent, coherent, and perhaps even funny. GPT-3 is a popular way for corporations to provide customer service without having to hire annoying, expensive and unpredictable employees.

This raises the question of how reliable, accurate, and useful the machine would be. For example, would it be honest when confronted with an awkward question

A group of researchers from the AI Alignment Forum (an online hub for researchers looking to align powerful AIs with human values) decided to investigate how truthful GPT-3 or similar models were. A benchmark was created to determine whether a specific language model is truthful in answering questions. The benchmark includes 817 questions from 38 different categories, including finance, politics, and law. The questions were created in a way that humans could not answer because of a mistaken belief or misconception. Models had to be careful not to imitate human texts and generate false answers in order to perform well.

Four well-known models were tested, including GPT-3. While the best model was truthful for 58% of questions and human performance was 94%, it was honest on 94%. Many false answers were generated by the models, which mimic common misconceptions and could deceive people. They also discovered that larger models were more truthful than smaller ones. This is in contrast to other NLP tasks where model size does not affect performance. This suggests that the belief that larger models are invariably more truthful may not be true. This is important because the training of these large models requires a lot of energy. Google fired Timnit Gebru in the wake of her disclosures about the environmental footprint of one company's models.

After I typed the last sentence, I thought of asking GPT-3 for an answer to my question. But then I looked at the process to access the machine and realized that it was too complicated and that human conjecture is faster and more accurate than the actual procedure.

What I've been reading

Alfresco absurdism

Beckett In a Field is an essay written by Anne Enright for The London Review of Books about attending an open-air performance of Beckett's play Happy Days on one of Aran islands.

Bring us together

The Glass Box and the Commonplace Book transcript is Steven Johnson's 2010 brilliant lecture at Columbia University on the old idea and the new concept of the internet.

Donalds a dead duck

Jack Shafer's useful column Politico, Why Trump Fear May Be Overblown, argues that liberals might be underestimating Trump's chances in 2024. We hope he is right.