On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

7 Paths Forward

1. Which development practices did Bender et al. advocate for language technology?
   Plan before building: weigh environmental and financial costs; curate and document data with frameworks like datasheets and model cards; identify direct and indirect stakeholders with value sensitive design; run pre-mortems imagining failures; consider dual use (e.g. watermarking generated text); and evaluate how models work, not just leaderboard scores.
   • Bender et al., 2021, On the Dangers of Stochastic Parrots (Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21)), §7, pp. 618–619
2. How did Bender et al. treat beneficial uses such as speech recognition for captioning?
   As a dual-use case: improved ASR helps Deaf and hard of hearing people, but large LMs aren’t the only path to better ASR (and can’t rely on terabytes of data for most languages); where LMs are critical, mitigate their harms as stochastic parrots while preserving their uses.
   • Bender et al., 2021, On the Dangers of Stochastic Parrots (Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21)), §7, pp. 618–619