MIT's Deb Roy Calls for Human Accountability and Answerability in AI-Generated Speech
MIT professor Deb Roy argues that civic systems and workplace tools must incorporate provenance and human answerability to prevent AI from eroding public trust.
Public participation mechanisms, such as notice-and-comment periods for proposed government regulations, are facing an unprecedented challenge from synthetic text generation. While administrative agencies have long contended with automated form letters, large language models now make it trivial to produce millions of unique, articulate responses that conceal their synthetic origins. In a recent essay published in the Bulletin of the Atomic Scientists titled "Who stands behind the words?," Deb Roy, professor of media arts and sciences at MIT and director of the MIT Center for Constructive Communication (CCC), argues that this development risks dismantling a pillar of democratic governance unless systems are redesigned around human accountability and answerability.
Redefining standing and answerability in the LLM era
Roy notes that public feedback channels have previously survived automated spam campaigns because older techniques left detectable traces. During the Federal Communications Commission's 2017 proceeding on net neutrality, for example, more than 22 million comments were submitted, and an investigation by the New York attorney general later revealed that nearly 18 million were fabricated. Those submissions were identified through repetitive phrasing and boilerplate templates. Modern generative models, however, eliminate those identifiable fingerprints by producing fluent, varied prose at scale, leaving reviewers unable to distinguish sincere civic input from algorithmic noise.
According to Roy, who co-founded the nonprofit Cortico alongside his work at CCC, software systems cannot replace the human responsibility required in public discourse. Because AI models have no physical stake or social consequences to bear, Roy emphasizes two crucial criteria borrowed from legal traditions: "standing" and "answerability." Standing refers to whether an individual has a genuine stake in a decision, while answerability entails the obligation to defend or explain statements once made. In Roy's formulation, autonomy in software operations is permissible, but unaccountability is not. To resolve this, he outlines four core design tenets: attributable authorship, decomposed and inspectable workflows, provenance as infrastructure, and expanding rather than concentrating standing.
Provenance and human oversight in practice
Rather than excluding generative technology entirely, Roy advocates for structuring AI as an assistive layer where people retain final authority and audit capability. At the CCC, researchers design systems where participants inspect and override machine outputs at every stage. For instance, in an effort led by MIT PhD student Cassandra Overney, the center developed a prototype called VoiceTrace. The tool maps summaries and recommendations generated during public discussions back to the specific recordings and spoken statements of individual participants.
This architecture has been applied across various community dialogues, including the Real Talk initiative, on-campus programs supported by MIT's administration, and a deliberative assembly at a New York City high school addressing school cellphone policies. Roy's team is also collaborating with the Google Experience Institute on the New Social Forms project. This initiative examines workplace collaboration where employees experience "tunneling"—delving deep into isolated work using personal AI tools, only to find their concepts out of alignment with teammates. Roy argues that both civic assemblies and workplaces require shared, auditable processes that advance "at the speed of trust" rather than purely prioritizing technical execution speed.
What it means for developers
Roy's framework highlights an important architectural shift for software engineers integrating generative AI into applications that collect feedback, summarize deliberative input, or assist group decision-making. Simply deploying an endpoint to summarize bulk user submissions risks obscuring participant voices and opening workflows to sophisticated astroturfing.
To address these risks, technical teams should consider incorporating inspectable chains of provenance into their products. Systems like VoiceTrace demonstrate that AI summaries must cite back to verified, underlying contributions rather than operating as opaque black boxes. Furthermore, verification layers can confirm that a unique human stands behind an action without unnecessarily compromising user privacy.
When designing and testing these accountable workflows, developers can evaluate top AI models cheaply through one API at https://apixoai.online. By experimenting with leading LLMs across different auditing pipelines, engineering teams can build platforms that harness generative synthesis while preserving inspectability, attributable authorship, and stakeholder trust.
Source: Q&A: Building accountability behind our words — and AI’s — in civic conversations — MIT News AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
