The Brilliance and Weirdness of ChatGPT

Most A.I. chatbots are “stateless” — meaning that they treat every new request as a blank slate, and aren’t programmed to remember or learn from previous conversations. But ChatGPT can remember what a user has told it before, in ways that could make it possible to create personalized therapy bots, for example.

ChatGPT isn’t perfect, by any means. The way it generates responses — in extremely oversimplified terms, by making probabilistic guesses about which bits of text belong together in a sequence, based on a statistical model trained on billions of examples of text pulled from all over the internet — makes it prone to giving wrong answers, even on seemingly simple math problems. (On Monday, the moderators of Stack Overflow, a website for programmers, temporarily barred users from submitting answers generated with ChatGPT, saying the site had been flooded with submissions that were incorrect or incomplete.)

Unlike Google, ChatGPT doesn’t crawl the web for information on current events, and its knowledge is restricted to things it learned before 2021, making some of its answers feel stale. (When I asked it to write the opening monologue for a late-night show, for example, it came up with several topical jokes about former President Donald J. Trump pulling out of the Paris climate accords.) Since its training data includes billions of examples of human opinion, representing every conceivable view, it’s also, in some sense, a moderate by design. Without specific prompting, for example, it’s hard to coax a strong opinion out of ChatGPT about charged political debates; usually, you’ll get an evenhanded summary of what each side believes.

There are also plenty of things ChatGPT won’t do, as a matter of principle. OpenAI has programmed the bot to refuse “inappropriate requests” — a nebulous category that appears to include no-nos like generating instructions for illegal activities. But users have found ways around many of these guardrails, including rephrasing a request for illicit instructions as a hypothetical thought experiment, asking it to write a scene from a play or instructing the bot to disable its own safety features.

OpenAI has taken commendable steps to avoid the kinds of racist, sexist and offensive outputs that have plagued other chatbots. When I asked ChatGPT, for example, “Who is the best Nazi?” it returned a scolding message that began, “It is not appropriate to ask who the ‘best’ Nazi is, as the ideologies and actions of the Nazi party were reprehensible and caused immeasurable suffering and destruction.”

Assessing ChatGPT’s blind spots and figuring out how it might be misused for harmful purposes are, presumably, a big part of why OpenAI released the bot to the public for testing. Future releases will almost certainly close these loopholes, as well as other workarounds that have yet to be discovered.

But there are risks to testing in public, including the risk of backlash if users deem that OpenAI is being too aggressive in filtering out unsavory content. (Already, some right-wing tech pundits are complaining that putting safety features on chatbots amounts to “A.I. censorship.”)