MIT's Deb Roy Calls for Human Accountability Behind Every AI-Generated Word in Civic Life

Every sentence that enters public discourse should have a person standing behind it, according to MIT professor Deb Roy, even when a language model produced the words. Roy, who leads work on civic media at the MIT Media Lab, makes the case that accountability cannot be optional as generative AI moves deeper into political speech, journalism, and public comment. His position is blunt: if no one can be named as answerable for a claim, that claim has no place in civic conversation.

The problem Roy points to is structural rather than technical. AI systems can produce fluent, confident text at a volume no newsroom, legislature, or election office can manually review, and the output often carries no traceable author. Attribution tools that watermark or fingerprint model output exist, but they identify the machine, not the person who chose to publish the result. Roy's framing pushes past detection and asks a harder question: who signs off, and who takes the hit when the text turns out to be wrong.

That gap has already shown up in practice. Campaigns have circulated AI-generated audio and images of candidates, and municipal governments have been forced to respond to floods of machine-written public comments during rulemaking periods, a tactic that has stalled proceedings in several US cities. In each case, the dispute centers less on whether AI was involved and more on the absence of a responsible party. Roy's answer treats the accountability requirement as a design constraint, not a cleanup step after publication.

His argument lands at an awkward moment for the industry. Model developers have spent the past two years adding provenance metadata, content credentials, and usage policies, but those measures stop at the API boundary. Once text reaches a platform or a campaign, the chain usually breaks. Roy's proposal implies obligations that fall on publishers and platforms as much as on labs, and it would require some form of disclosure that names a person or institution rather than a model version.

Expect the idea to run into familiar objections about anonymity, whistleblowing, and speech protections, since requiring a named backer for all public text would collide with legitimate uses of pseudonyms. The harder fight is over enforcement, whether through platform policy, disclosure law, or professional norms in journalism and politics. What to watch is whether any legislature or major platform adopts an answerability rule, and how they define who counts as the responsible party when the drafting was done by a model and the publishing by a committee.