In a 13 October 2026 MIT News interview, Deb Roy, director of the Center for Constructive Communication, argues for an accountable person behind AI-assisted civic text. A trusted institution may protect their identity. This is an argument, not a demonstrated technical intervention.
Mass-produced speech is not a new problem
In May 2021, New York’s attorney general reported that nearly 18 million of more than 22 million comments in the FCC’s 2017 proceeding were fake. This historical case of automation and impersonated participation predates today’s generative models. We do not describe it as an attack performed with a current AI product.
The context raises a question separate from writing quality. Even a persuasive comment does not demonstrate that a real participant has adopted it as their position. In a public consultation, the number of well-written messages is not automatically the number of people. That distinction is our editorial interpretation of why responsibility matters alongside fluency.
Provenance helps without settling every question
C2PA describes Content Credentials as a cryptographically signed history of a digital asset’s creation and changes. Signatures and hashes can reveal a broken link between a file and its provenance record. They provide information about origin, not automatic proof that a statement is true or that its speaker represents genuine civic participation.
For a newsroom, these are separate layers. One records which version was published and where its ingredients came from. Another takes responsibility for source selection, the meaning of claims and corrections. A valid technical record does not eliminate the need for somebody to read and understand the result before publication.
This distinction matters when a system displays a confidence or verification badge. The badge should describe an actual check. It should not quietly combine file integrity, source credibility and editorial judgment into one undifferentiated promise. Those checks can reach different conclusions about the same item.
A practical illustration
Imagine a local organisation using AI to summarise voluntarily recorded discussions. A useful question is whether participants can challenge a misrepresented position. If the summary says everyone supports a proposal, an editor should be able to locate the actual basis for that claim, including dissent. This is our hypothetical example, not a reported MIT result.
A participant may want to be heard without having their name preserved in a public archive. Accountability should therefore not automatically become compelled disclosure. The organisation should explain who can access the record, how mistakes are corrected and who resolves disputes over the representation of somebody’s words. Merely attaching a name to a generated summary would not answer all those questions.
It is equally important not to invent human approval. If a piece has passed only automated checks, that should be apparent. A person adopting a particular version and being available to answer questions represents a different state. A “verified” label is useful only when the reader can understand the process behind it.
What today’s discussion adds
Our editorial conclusion is specific: a system should preserve who adopts a claim and how it can be corrected. Technology can support that organisational task, but cannot simply declare it solved.
Editorial verification: 13 October 2026, 2026-10-13T16:16:52.923011+00:00 (UTC). The cover is an AI illustration, not a photograph of an actual event.
