Computers can't understand ordinary language
The old claim turned out to be wrong and was rejected by evidence.
TaughtMachines need exact commands. They cannot handle ambiguity, metaphor, or context, so you have to speak to them in code or in rigid menu choices. Natural conversation with a computer is science fiction.
NowLarge language models write, translate, summarize, explain, and code from plain conversational instructions. ChatGPT's launch in November 2022 reached an estimated 100 million users within two months, the fastest adoption of any consumer application to that point.
What actually happened
The claim was well-supported for decades. Rule-based natural language processing was brittle, and the classic demonstration was the ambiguity of a sentence like "time flies like an arrow." Machine translation was a punchline. Anyone who used a phone tree in 2010 had direct evidence that computers did not understand speech.
The hinge was the transformer architecture, introduced in the 2017 paper "Attention Is All You Need," and the discovery that scaling it produced abilities nobody explicitly programmed. GPT-3 in 2020 showed it inside the research community. ChatGPT in November 2022 showed everyone else, because it removed the last barrier: you no longer needed to know anything to use it.
The honest caveats matter here, because this entry is about a technology that will happily state falsehoods with confidence. These systems predict text rather than consult knowledge, they fabricate citations, and "understand" is a philosophically loaded verb that researchers argue about seriously. But the practical claim in the old curriculum, that you cannot get useful work out of a computer by simply asking in English, is finished.
Newsletter
Get new weekly revisions
Sources
- arXiv: Attention Is All You Need (Vaswani et al., 2017)arxiv.org
- Stanford HAI: AI Index Reporthai.stanford.edu
Who was taught this
Still standard through 2021, so anyone who finished school between 1950 and 2022 learned the earlier version.