Accountability in the AI Era: Textual Responsibility and Attribution in Public Discourse

Introduction: When AI Joins Public Discourse, Who Is Responsible for the Words?

In an age of information overload, every piece of text in public discourse can shape public opinion, decisions, and even the direction of society. Yet as AI begins generating content at scale, a question surfaces: who should be responsible for these words? Deb Roy, a professor at the MIT Media Lab, offers a clear answer: whether the words are written by a human or by AI, there must be an accountable subject—an "answerable" person.

This is no idle worry. Roy has long studied human-computer interaction and media communication, and his view strikes at the core contradiction of today's AI governance: technology has made content production cheap and anonymous, but public life demands that responsibility remain traceable.

Why Is Accountability More Urgent in the AI Era?

Mass Production of Text and the Responsibility Vacuum

In the past, public speech usually had a clear author, editor, or spokesperson behind it. Traditional media had editors-in-chief to gatekeep, academic papers had peer review, and political speeches had teams of speechwriters. But the rise of generative AI has broken this chain. Anyone can instantly produce articles, commentary, or reports that look professional, while no one vouches for their accuracy, intent, or consequences.

Professor Roy points out that this "responsibility vacuum" is eroding the foundation of trust in public discourse. When a passage of text may come from AI, may have been rewritten multiple times, or may have no clear source, readers cannot judge what to believe—and when something goes wrong, they cannot find anyone to hold accountable.

AI Is Not a Legal Subject, but Humans Can Serve as Anchors of Responsibility

A common misconception is that since AI generated the content, AI should be responsible. But Roy's view is clear—AI has no legal personality and cannot bear moral or legal responsibility. Accountability must fall on people: perhaps the organization deploying the AI, the author using it, or the engineer who designed the system.

This line of thinking echoes the "human oversight" principle in the EU AI Act, but Roy goes further: he demands that every piece of public text have an identifiable responsible party, regardless of how it was generated.

Legal personality refers to the legal recognition that an entity has the capacity for rights and the capacity to act, and can independently bear legal responsibilities and obligations. In the vast majority of jurisdictions, only natural persons and legally registered legal persons (such as companies and social organizations) possess legal personality; animals, machines, and algorithms do not. This means that even if an AI system autonomously generates illegal or harmful content, it cannot be sued, fined, or held accountable—it has neither property to compensate with nor liberty to be deprived of. For this reason, legislation across countries generally treats AI as a tool rather than a subject. The "risk-tiering plus human oversight" framework adopted by the EU AI Act is built on this premise: AI providers, deployers, and users each bear different levels of compliance obligations. Roy's view is consistent with this, but goes further in requiring that every piece of public text be traceable to a specific person, thereby filling the accountability vacuum with a "responsibility anchor" in the absence of a legal subject.

How Can Accountability Be Implemented? Three Practical Directions

1. Provenance and Labeling: Making AI Involvement Transparent

The first step toward accountability is awareness. Roy suggests that any public content involving AI generation should be clearly labeled, indicating the AI's role—whether it assisted with polishing, fully generated the content, or merely provided data. This transparency allows readers to assess credibility and also provides clues for later attribution of responsibility.

At present, some platforms have begun experimenting with "AI-generated" labels, but standards vary. Roy calls for a unified, cross-platform labeling framework so that provenance information becomes as standard as a food ingredient list.

Content provenance and digital watermarking are currently two mainstream technical approaches. Digital watermarking identifies AI generation by embedding hidden signals in text or images, but text watermarking is especially fragile—rewriting, translation, or simple editing can destroy the signal. Content provenance, by contrast, uses cryptographic signatures to record the origin and modification history of content. The representative scheme is the C2PA (Coalition for Content Provenance and Authenticity) standard, jointly promoted by organizations such as Adobe, Microsoft, and the BBC, and already supported by some cameras and editing software. The two approaches differ in philosophy: watermarking seeks to "detect" traces of AI, while provenance "records" the chain of creation. The cross-platform unified labeling framework Roy advocates is closer to the provenance approach—not relying on after-the-fact detection, but leaving verifiable responsibility information at the generation and publication stages. At present, standards in this field remain fragmented, and insufficient interoperability is the main obstacle.

2. Responsibility Chain Design: From Prompts to Publication

At the technical level, accountability needs to be embedded in workflows. For example, when a company uses AI to generate public statements, there should be clear reviewers, approvers, and publishers. Roy calls this a "responsibility chain"—at every stage, a specific person is responsible for the content.

This resembles the "code owner" system in software engineering: every line of code has an owner, so when problems arise they can be quickly located. Transplanting this idea to content production can avoid the dilemma in which "everyone is using AI, but no one is responsible."

CODEOWNERS is a mechanism on code hosting platforms such as GitHub: teams can designate owners for specific files or directories in a repository, and when someone submits changes, the system automatically requests those owners to review them. Its core logic is to refine responsibility to the file level, ensuring that every change has a clear approval subject. Roy's transfer of this idea to AI content production essentially introduces the "auditable workflow" of software development into the field of public communication. In practice, this means designating responsible parties separately for prompt writing, model selection, output review, and final publication, while retaining operation logs. Unlike code review, content production is faster-paced and involves more dispersed participants, so responsibility chain design must balance efficiency and avoid losing the productivity advantages brought by AI due to too many approval layers.

3. Institutions and Culture: Cultivating the Norm That "Byline Means Responsibility"

Beyond technical means, accountability is also a culture. Roy emphasizes that participants in public discourse—journalists, scholars, policymakers, and opinion leaders—need to re-establish the norm that "byline means responsibility." Even if content is assisted by AI, the person whose name appears must still be responsible for the facts and views.

This requires the joint efforts of the education system, professional ethics, and platform rules. For example, journalism schools can teach ethical guidelines for AI-assisted reporting, and platforms can downrank misleading content that fails to label AI sources.

Challenges and Controversies: Will Accountability Stifle Innovation?

Dissenting Voices: Excessive Accountability May Suppress Expression

Some critics argue that requiring every piece of text to have a clear responsible party may make users afraid to use AI tools, or lead platforms to over-censor. Especially in scenarios such as anonymous whistleblowing or discussion of sensitive topics, mandatory accountability may create a chilling effect.

Roy's response is that accountability does not equal public identity disclosure. Through third-party custody, encrypted signatures, and other methods, it is possible to protect privacy while ensuring a traceable responsible subject. The key is that an accountable person exists, not that they must be disclosed to the public.

Technical Feasibility: Provenance Tools Are Still in Their Early Days

At present, AI content detection and provenance technologies are not yet mature. Watermarks can be removed, and detectors can misjudge. Roy acknowledges that a complete technical solution does not exist, but institutional design can come first—just as the copyright system does not rely on perfect technology, but on law and social norms.

Conclusion: Rebuilding Trust for Public Discourse

Deb Roy's appeal is essentially about rebuilding the trust infrastructure for public discourse in the digital age. When everyone—regardless of the tools they use—takes responsibility for their own words, public discussion can return to rationality and credibility.

AI is not an exception, but a catalyst. It forces us to rethink an ancient question: who is speaking, and who should be responsible? The answer must be: people.