A third category of risk involves bad actors taking advantage of the ability of large LMs to produce large quantities of seemingly coherent texts on specific topics on demand in cases where those deploying the LM have no investment in the truth of the generated text. These include prosaic cases, such as services set up to ‘automatically’ write term papers or interact on social media,23 as well as use cases connected to promoting extremism. For example, McGuffie and Newhouse [80] show how GPT-3 could be used to generate text in the persona of a conspiracy theorist, which in turn could be used to populate extremist recruitment message boards. This would give such groups a cheap way to boost recruitment by making human targets feel like they were among many like-minded people.