In April, McClatchy, the newspaper chain that owns the Miami Herald and The Sacramento Bee, introduced an AI-powered “content scaling agent” (CSA). While AI summaries are now ubiquitous, McClatchy’s tool went a step further: It takes journalists’ original reporting (and even in some cases, reporters’ notes), reshapes them into new formats and targeted summaries, then publishes them as new, separate articles just as if they were written by fellow staff writers.

If it were up to McClatchy, CSA-generated articles would carry human reporters’ bylines. Kathy Vetter, the chief of staff, said during a March 17 meeting that “If [journalists] don’t have the ability in their contract to remove their byline, we’re going to use their name.” Why does McClatchy want human journalist bylines instead of running the CSA stories under a generic credit? Because they want the legitimacy that comes from marking such content as human. Eric Nelson, vice president of local news, said human bylines on the CSA stories was a way to show “authority” on Google so the articles would be ranked higher in the search engine’s results. Now, journalists at several McClatchy newspapers are engaged in a byline strike. They argue that the use of the CSA material amounts to an ethical breach that undermines the trust between local newsrooms and the communities they serve.

The McClatchy fight illustrates how AI accelerates and combines two challenges that the information ecosystem was already facing.

The first is what I call the Cylon problem. Simply put, one can never be quite sure whether a person or a piece of content one encounters online (and increasingly offline) is real or fake. The persistent possibility that anything or anyone one encounters could be fake cultivates a constant state of interpretive suspicion, a paranoid style of online (a)sociality. It also creates a second order dynamic where the real is constantly at risk of being undermined by the fake. The real is falsely accused of or negligently labeled as fake, it is imitated to death by the fake, and the real is drowned out by the fake. The Cylon problem predates AI, but AI intensifies the problem along both qualitative and quantitative dimensions. High quality fakery can now be produced at scale and on the cheap. Unscrupulous actors can mass-produce slop and engagement bait. They can also level (at least facially) credible accusations of fakery to discredit opponents. The Cylon problem accelerates the downward swirl of epistemic destabilization that the shift from traditional to online media had already kicked off. Left unchecked it produces what Jason Koebler calls the “zombie internet.”

The second challenge is concentrated informational power. AI doesn’t just empower fakers and fakery in horizontal competition with journalists, artists, and other knowledge producers. Cheap, high-quality fakes waiting in the wings threaten workers in their vertical relationships with bosses in newsrooms, studios, or universities. Like prior forms of automation, AI replaces labor with capital in production and thus offers a technological means of empowering capital at labor’s expense. The Miami Herald might employ a journalist, but McClatchy cannot own that journalist’s tacit knowledge of when to push a source, when to track a lead, and what cluster of facts form the heart of a story. The space of a journalist’s judgement, experience, instinct, and skill is also the space of her freedom and her power. An editor can impose some control but cannot tune how she wields her perspective and her skill in how a story gets told. But McClatchy can “own” the output of its CSA, and its AI can fine-tune its outputs along varying shades of emotional tone and perspective to precisely match the preferences of an editor (or an owner).

McClatchy is not alone in its push to supplement worker-made media with AI offerings. Universities have begun experimenting with feeding videos of course lectures from Canvas into AI systems to produce AI-built personalized courses without faculty knowledge or consent. Movie studios have pushed actors to consent to full-body scans and fights over digital replication and likeness were a major feature of the 2023 SAG-AFTRA negotiations. Startups have unveiled AI-generated podcasts.

In a general sense, the introduction of the CSA, and other AI tools like it, follows a well-worn playbook of automation as wage suppression. Step one: Rip off the knowledge labor encoded via stores of human content by scraping the internet and engaging in intensive worker surveillance. Step two: Volun-tell workers to engage with and refine AI products to make them of sufficient human-seeming quality. Step three: Once the AI is sufficiently good at mimicking worker produced content, fire the workers or wield the threat of AI replacement to weaken worker standing. To a greater degree than past forms of automation, producing AI good enough to replace or disempower workers requires cooperation from those same workers, or at least workers in the same industry. The transformation of knowledge labor into AI capital is mediated via data that workers themselves generate while doing their jobs. Indeed, steps one and two describe how data encodes the worker’s own knowledge, rendering it legible to AI systems that can then replicate the worker’s expertise.

**

Evidence suggests that many people do not like navigating the Cylon problem. Take music for example. A poll by The Hollywood Reporter and the Frost School of Music found that 66 percent of people have never knowingly listened to music generated by AI. This dislike of AI music also shows up in revealed preference. At the streaming platform Deezer, while there are 75,000 daily uploads of AI music, a number that “threaten[s] to overtake actual human-made music,” Deezer’s director of research notes that on the listener side, growth is mostly in fraudulent streams, and that “consumption after fraud removal is not gaining much traction and is still very concentrated on a few viral tracks.” Indeed, fraudulent streams of AI music make up an increasing proportion of AI music streams overall, suggesting that growth in AI music streams comes from fraudulent streams that have succeeded in beating listeners at the Cylon problem. Beyond music, when a production company debuted Tilly Norwood, the first AI-generated actress, it faced overwhelming backlash and condemnation across the film industry. A recent survey found that among professional visual artists, 99 percent of artists “disliked” generative AI. Familiarity seems to breed contempt; 85 percent of respondents said they abstained from using generative AI in their work, even though a majority encountered generated AI images at least weekly in their practice.

Indeed, a recent Pew Research Center poll found Americans are overwhelmingly more concerned than excited about the increased use of AI in daily life. Opinion tracks a distinction in how AI is used. As an information synthesis and detection technology embedded in finance, weather modeling, and fraud detection, people are broadly in favor of at least some AI use. People are overwhelmingly opposed to AI when it replaces human advice regarding issues people consider personal and more subjective, such as advising people about faith or relationships. To be clear, this is not to say that all people are against engaging with AI. But it is to suggest that many people are engaging with AI more than they would prefer, especially when such engagement is with AI agents or content that actively presents as human or human-made.

**

A range of possible interventions could alleviate key elements of the Cylon problem and its accelerating effect on informational power consolidation.

First, jurisdictions could extend existing fraud and misrepresentation doctrines to include corporate practices that intentionally or negligently mislead customers regarding a basic assumption underlying the commercial interaction—that customers are spending their valuable and finite time, attention, or money on human-produced content or speaking with a human agent.

Second, extending the same justifications that ground affirmative disclosure requirements under contract and consumer protection law, jurisdictions can pass legislation to require clear labeling of AI-generated content and disclosure of AI agents. Some entities, like Spotify and The New York Times, are already voluntarily developing practices to label AI content and disclose how AI was used in the creation process. These voluntary schemes are laudable. They help consumers make informed choices about what content they choose to consume. However, absent market-wide requirements for disclosure, companies may face structural incentives to underinvest in voluntary standards.

Underinvestment seems likely due to a combination of several factors. First, there is a large market for genuinely human content and human interaction that is (in theory) more expensive to produce. Given how cheap it is to imitate humans and human content, and how desirable human content or interactions are, there is a sizable incentive to pass off AI content as human. This, in effect, encourages business practices that deceive consumers into paying (either financially or with their time and attention) for content they have expressed a wish not to consume. In theory, such behavior would be disciplined by consumer demand and choice—customers, faced with options between human content and services and Cylon content and services, can reward the companies that invest in producing the content and services they like. But the very nature of the Cylon problem is that the risk of customer detection of AI is low. Thus, even well-meaning companies may face market incentives that reduce labeling practices into “humanwashing.” Companies can get a reputational boost from supporting human creators by announcing labeling practices—but will not face sufficient market discipline from consumers to adequately invest in the ongoing enforcement required to substantially prevent consumers from engaging with fraudulent AI.

In contract settings, many jurisdictions impose affirmative disclosure obligations on the party best positioned to efficiently identify a latent defect that is considerably more difficult (i.e., more costly) or impossible for the other party to discover. This maps nicely onto the Cylon problem. It is difficult if not impossible for consumers to detect if the content they are consuming or agent they are speaking to is AI but readily known by the company generating the content or hosting the agent. Rather than forcing consumers to engage in expensive (and often inaccurate) AI detection to obtain information the company already has, affirmative disclosure ensures efficient information sharing between the parties.

One common counterargument to the Cylon problem is to reject that it describes a problem at all. If consumers can’t tell the difference, the argument goes, then why should it matter if content is AI or human? This objection is odd given the prevailing tendency of commercial legal regimes to center consumer sovereignty and demand-led market incentives. Ours is not to question why consumers want to know if food contains GMOs, or if supply chains contain child labor, or if products are American made. If these are salient and material consumer demands, then we ought to facilitate market conditions that reward companies that invest in honestly and fully meeting those demands. The same goes for human content.

Of course, AI disclosure obligations—and even more so, claims that holding out AI-generated work as human work constitute fraudulent or negligent misrepresentation—exist in tension with the maximalist free expression positions that newsrooms have historically adopted regarding other disclosure obligations. In The Washington Post v. McManus for example, the U.S. Court of Appeals for the Fourth Circuit struck down a Maryland law requiring newspapers to disclose political advertising on First Amendment grounds. However, that law was content based—it regulated campaign related speech only—and it targeted political speech. Maryland’s law also implicated The Washington Post’s editorial discretion, since media outlets’ advertising choices receive First Amendment protections. Disclosing that an article was generated in whole or part by AI is not specific to the content of that article, nor does it target political speech. Insofar as it restrains editorial decisions, it does so narrowly, to disallow the deliberate misrepresentation of the nature of journalistic content to audiences.

To be clear, disclosure obligations may not resolve, or even apply to, every kind of AI Cylon problem. Overly ambitious AI misrepresentation or disclosure laws may run afoul of First Amendment challenges. Misrepresentation and affirmative disclosure frameworks govern speech within commercial exchanges; the burdens they place on speech are grounded in deep justifications regarding the centrality of commercial speech to contracting activity. AI content that does not implicate or involve a contractual relationship with the entity providing the speech, especially if such speech is political or entertainment, will have stronger First Amendment protections against mandatory labeling or disclosure.

Disclosure proposals focus on the demand side. The basic idea is that clearly demarking human content will relieve both horizontal and vertical pressure on human creators. If newsrooms are adequately rewarded for not replacing humans, they will face less market pressure to do so. If it becomes expensive and risky to lie to your customers about your “live booking agent” being AI, the cost savings of replacing humans with AI are reduced.

But reform proposals on the supply side exist too. Indeed, such proposals may offer a more enduring pathway to broader reform. Data association rights for content and knowledge producers would give data producers greater control over if, when, and how their data mediates the transformation of knowledge labor into AI capital. It would give journalists, influencers, and workers not just a negative stake against the privacy invasions of surveillance, or a limited copyright claim against certain forms of expropriation. Instead, systematic data entitlement rights would give data producing communities meaningful control—over whether their knowledge is commodified for AI use at all, and if so, how they may fairly benefit from its value. To provide effective supply side counters to concentrated informational power, broad data entitlement rights are best understood as rights to form associations and bargaining units with similarly situated data producers, rather than more traditional individual entitlements over one’s data itself. As I and many others have argued, data governance rights work best when collectively exercised via associations, rather than via individuals. Granting creative or workers associations greater collective power to determine if, when, and how their data is used to produce AI capital would allow creators and knowledge producers to assert control over how AI is produced and how its value is shared. Proposals for greater data entitlement have the additional benefit of largely avoiding (or at least minimizing) First Amendment challenges. Such proposals do not restrict or impose conditions on institutional speech; their intervention is upstream of any instance of expression. Instead, they expand the set of rightsholders that have a say in how corporate automated expression is produced to begin with, and who can claim a share of its value. Disclosure obligations and other demand-side reforms may provide a stop gap measure against an unchecked ecosystem of Cylons. But supply side interventions that distribute informational power more broadly are, in my view, necessary to achieve more enduring and significant reform.