On the dangers of stochastic parrots

The paper presented at the FAccT conference on 3 March 2021 made four arguments against the trajectory of the field. Training the largest models consumed energy on the scale of small towns, with the costs falling on people who did not benefit. Web-scale training data encoded the biases of whoever wrote the web, and was too large to audit. Research effort was being spent on scale rather than understanding. And a language model, the authors wrote, is a stochastic parrot: a system that stitches together sequences it has observed according to probabilistic information about how they combine, "without any reference to meaning".

The paper was famous before it was published. Google had asked its co-author Timnit Gebru, co-lead of the company's ethical AI team, to withdraw her name; she refused and was fired in December 2020, and her co-lead Margaret Mitchell was dismissed two months later. Thousands of employees and researchers protested.

The phrase entered the language as the standard sceptical position on large models, and the paper's specific concerns, energy, data provenance, bias, became regulatory issues. Whether the models refer to meaning remained the most contested question in the field.