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

6.2 Risks and Harms

1. What harms can follow when LMs reproduce biases in their training data?
   Generated text can spread and amplify stereotypes and abusive language, causing psychological harm to people targeted (microaggressions, stereotype threat), reinforcing others’ prejudices and discrimination, and feeding biased text into future training data; inside classifiers or search, biases cause allocational harms largely invisible to those affected.
   • Bender et al., 2021, On the Dangers of Stochastic Parrots (Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21)), §6.2, pp. 617–618
2. Which harms from deliberate misuse and from fluency did Bender et al. describe?
   Cheap on-demand text for extremist recruitment (GPT-3 imitating a conspiracy theorist), term-paper mills and bots; fluent but wrong machine translation taken as the author’s intent (a Palestinian man was arrested after “good morning” in Arabic was translated as “attack them”); and extraction of personal information memorized by larger models.
   • Bender et al., 2021, On the Dangers of Stochastic Parrots (Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21)), §6.2, pp. 617–618