U.S. FEDResearch
SSRN: Tam proposes training-cutoff disclosure and a pre-election partisanship slider for chatbots
In an October 13, 2026 SSRN preprint, Wendy Tam argues that large language models answering political questions act as political intermediaries, heirs to newspaper editors and broadcasters, because they compose and frame answers rather than retrieve them. She proposes requiring disclosure of training-corpus cutoff dates and, in the 60 days before a federal election, a partisanship slider in chat windows.
Key points
- Tam posted the article on SSRN on October 13, 2026.SSRN
- She notes voters have begun consulting language models for political information and that studies find answers can change minds.SSRN
- The paper reports interpretability research finding models encode partisanship as a single internal direction matched to users' inferred politics.SSRN
- She argues law has long regulated intermediaries such as broadcasters and cable operators, yet nothing regulates language models' political influence.SSRN
- The first proposed obligation dates each model's framing by disclosing when its training corpus closed.SSRN
- The second would show voters what the same model produces when steered otherwise; both proposals are described as content neutral.SSRN