A computer cannot write prose you would mistake for a human's
The old claim turned out to be wrong and was rejected by evidence.
TaughtComputers manipulate symbols; they don't understand language. A machine might autocomplete a word or churn out stilted, robotic sentences, but it could never write an essay, an article, or a story fluent enough to pass as the work of a person.
NowIn June 2020 OpenAI released GPT-3, a 175-billion-parameter language model that produced paragraphs of coherent, on-topic prose from a short prompt. In OpenAI's own tests, human readers could barely tell its short news articles from ones written by people, doing little better than a coin flip.
What actually happened
GPT-3 was not taught grammar or facts directly. It was trained to do one dull thing at enormous scale, predict the next word across a huge slice of the internet, and fluent writing emerged from that alone. The result surprised even its builders, who described it in a paper pointedly titled 'Language Models are Few-Shot Learners.'
The honest boundaries were clear from the start: the model has no understanding of what it writes, invents facts with total confidence, and echoes the biases of its training text. It is a spectacularly good mimic, not a mind.
But the confident classroom line, that a computer simply cannot produce writing you'd mistake for a person's, did not survive 2020. Everything that followed, ChatGPT included, is built on what GPT-3 showed was possible.
Newsletter
Get new weekly revisions
Sources
Who was taught this
Still standard through 2020, so anyone who finished school between 1950 and 2020 learned the earlier version.