MIT’s Deb Roy Urges Answerability for AI-Assisted Public Speech
MIT researcher calls for accountability behind AI-assisted text
Deb Roy, professor of media arts and sciences at MIT and director of the MIT Center for Constructive Communication, is calling for every public-facing piece of text to have a person who can answer for it, including material drafted with artificial intelligence. In an essay for the Bulletin of the Atomic Scientists, “Who stands behind the words?”, Roy argues that AI can assist communication but cannot be the endpoint of responsibility.
Roy’s concern is especially acute in civic participation. He cites the US Federal Communications Commission’s 2017 proposal to repeal net-neutrality rules, which received more than 22 million public comments. The New York attorney general later found that nearly 18 million of those comments were fake. At the time, templated language and duplicated phrasing could leave evidence of coordinated submissions; large language models can instead produce fluent, varied text that is harder for reviewers to identify as inauthentic.
Standing and answerability in public processes
Roy distinguishes between a system producing words and a person having standing to make them. In his framing, the right to be heard should carry an obligation to answer for a statement, particularly when it makes a specific accusation. A software system, he argues, cannot anchor that chain of accountability because it has nothing at stake and cannot be cross-examined.
That approach does not require public disclosure of every speaker’s identity. Roy says privacy and answerability can coexist when an institution verifies that an eligible, distinct person is speaking and stands behind the communication without exposing that individual. He proposes that disclosure obligations should increase with power, while identity protection should increase with vulnerability.
AI remains useful within that framework. A person with a stake in a proposed rule can use AI to draft a comment, adopt it as their own, and remain prepared to answer for it. Roy’s distinction is that AI may support drafting, summarisation, and pattern-finding, but it cannot be the party ultimately responsible for the words.
Designing systems around traceable responsibility
Roy identifies four principles for systems that use AI in consequential communication: attributable authorship, a decomposed and inspectable process, provenance as infrastructure, and widening standing rather than concentrating it. The guiding rule, he writes, is that autonomy can be permitted while unaccountability cannot.
The MIT Center for Constructive Communication and its affiliated nonprofit, Cortico, use technology to help identify themes in group discussions while giving participants opportunities to inspect and override technology-based outputs. The centre’s VoiceTrace prototype, developed by PhD student Cassandra Overney, links summaries and recommendations back to the specific voices in the underlying listening process.
Roy also describes a coordination issue emerging in workplaces. Through work connected to the Google Experience Institute, the group has observed people using AI to develop ideas individually before returning to teams with increasingly divergent perspectives. The resulting challenge is not only technical efficiency but how groups maintain trust, listen to one another, and make decisions together.
Business implication
For organisations deploying AI in policy consultation, employee dialogue, customer communications, or internal decision-making, the practical implication is to retain an accountable human owner for each consequential output and preserve a traceable path from claims and recommendations to the people and discussions behind them.