There is growing recognition of bias in generative artificial intelligence (AI) and large language models (LLMs). There are cases of chatGPT incorrectly parsing prompts because it fails to recognize women can be doctors; proposing an algorithm to determine whether someone is a good scientist solely based on whether they are white and male; and suggesting country of origin as a criterion for whether people should be tortured. Bias in this framework is the tendency to reflect and promote facts, ideologies, and beliefs that either discriminate against protected identities or inaccurately reflect the true underlying distribution of those facts, ideologies, and beliefs. This is concerning because generative AI has capacity to translate that biased information into seemingly objective text.

Governments with control over the information environment are able to produce and disseminate propagandist texts through print and digital news, which then are utilized as training data for LLM development. Discerning propaganda from non-propaganda content at scale remains challenging, and thus language models may be inept at removing such biased content from their training data and predictions.

A project from Joshua Tucker’s lab at NYU will consider propaganda bias in AI. They will test the claim that when strategic actors influence global discourse online, they also shape the training data for LLMs, which in turn can shape the models’ output in ways favorable to the actor’s cause. China will be used as a case study because of preliminary work the lab has done.

In addition to the full analysis of propaganda bias in the Chinese context, the lab will also audit Russian propaganda bias for terms related to Russia and Russia’s invasion of Ukraine; Arabic propaganda bias for terms related to Saudi Arabia; Spanish language propaganda bias for terms related to Venezuela; and Vietnamese propaganda bias for terms related to the Communist Party of Vietnam and its leaders.