The Centre Daily Times in State College, PA, has voted to unionize after months of pushback against its parent company’s AI tool—a move that, according to The NewsGuild-CWA, makes it the first newsroom in the union to cite AI adoption concerns as a primary reason for organizing.

As Nieman Lab reported, the Centre Daily Times staff voted to join The NewsGuild of Greater Philadelphia last month. All eligible editorial staff signed authorization cards, and McClatchy voluntarily recognized the union. The catalyst, reporters told Nieman Lab, was McClatchy’s Content Scaling Agent (CSA) tool—an AI system that repackages existing articles into short-form summaries for publication or video scripts—and a March internal meeting where Kathy Vetter, McClatchy’s chief of staff for local news, told staff the company would use their bylines on AI-generated content unless union contracts prohibited it.

Josh Moyer, a senior reporter at the Centre Daily Times, took that as a signal. “It was essentially like, if you’re not in a union, your byline gets used; if you are in a union, we’ll follow what the union says,” Moyer told Nieman Lab. “If we want to control what happens to our byline, that’s the company telling us that we need to form a union.”

McClatchy introduced the CSA tool at the paper earlier this year. Reporters initially published at least one CSA-assisted story per week under a generic byline noting AI assistance. But in late February, management changed the policy: AI-generated content would now carry the reporter’s actual name. Reporters objected that it misrepresented their work to readers.

“When our names go on a thing, it says that this article or video is from that person, but that is just not true in this case,” said Trebor Maitin, a service reporter. Maitin was the first reporter at the paper to have his byline changed to reflect AI assistance.

The NewsGuild-CWA’s president, Jon Schleuss, said unionized newsrooms have had more success keeping AI content clearly labeled: “Unionized newsrooms are the ones where McClatchy’s AI slop gets a clear label. In non-union newsrooms, the AI slop may be carrying a real human reporter’s byline.”

Multiple McClatchy publications have seen byline strikes over the CSA tool, and some have taken labor actions over the tool and related workplace issues. For the Centre Daily Times, the union opens the door to formal collective bargaining and the ability to join coordinated actions at sister publications.

“Some of us use AI a lot more and are okay with it,” Maitin said. “But there is an overall understanding that we need to be able to have a say in this, and that unionizing at least gives us a seat at the table.”

The post A newspaper unionized because McClatchy put reporters’ names on AI content appeared first on The Media Copilot.

YouTube is moving from voluntary disclosure to automatic detection when it comes to AI-generated content.

The platform announced Wednesday that it will begin automatically applying AI-generated content labels to videos that feature “significant photorealistic AI use”—even if the creator never disclosed it. As noted by Variety, the change marks a notable escalation of YouTube’s approach to AI transparency, which previously relied entirely on creators self-reporting their use of generative AI tools.

Previously, YouTube labeled AI-generated content only when creators voluntarily disclosed it in their video settings. Now the company is rolling out an internal detection system that will flag videos even without creator admission. Creators can dispute incorrect labels through YouTube Studio, but YouTube says the labels will “remain permanent” in certain cases, including content created using YouTube’s own AI tools like Veo or Dream Screen, and any video carrying C2PA metadata (standards from the Coalition for Content Provenance and Authenticity) indicating full AI generation.

The labels are also getting a more prominent placement. Previously buried in expanded descriptions, AI labels on long-form videos will now appear directly below the video player, above the description. For YouTube Shorts, the label will appear as an overlay directly on the video itself.

“The goal here is context at a glance,” said Rene Ritchie, YouTube head of editorial and creator liaison, in a video explaining the changes. “If it looks real but was made with AI, viewers will know immediately.” Ritchie emphasized that the labels do not affect monetization or recommendation algorithms — “This is purely about giving viewers the right information at the right time.”

This push for better AI disclosure follows a broader problem: why AI content labels keep failing the people who need them most. Content Credentials, the metadata-based standard designed to track an image’s AI origins, has existed for years, but social platforms have been inconsistent about adopting it, often stripping out the very metadata that makes the system work. YouTube’s move toward automatic detection is an attempt to close that gap, even if the underlying standards remain patchily implemented.

The move comes alongside YouTube’s expanded likeness-detection program, now available to all creators 18 and older, which helps users identify and request removal of AI-altered facial likeness content.

The post YouTube Will Auto-Label AI-Generated Videos, and Make Those Labels Harder to Miss appeared first on The Media Copilot.

A former startup CEO who pleaded guilty to defrauding investors has been linked to a network of AI-generated fake local news outlets — built, by his own account, to manipulate how both people and search engines perceive people like him.

Drew Chapin, who in 2021 pleaded guilty to investor fraud in connection with a failed tech startup, founded The Discoverability Company, a Philadelphia-based online reputation management firm. According to an investigation by The Florida Trib in partnership with KCRW’s Question Everything podcast, that firm is behind the South Florida Standard, a website that presented itself as a local news outlet but was, in fact, entirely AI-generated.

The South Florida Standard featured fake reporters with AI-created headshots and fabricated biographies. Stories were lifted from legitimate news organizations, run through AI, and republished as original content. The site published three stories a day, seven days a week—including Easter Sunday—under bylines that had no professional history or digital footprint outside the site.

The problem of AI-generated content disguised as legitimate journalism is one our coverage has tracked closely, and it’s getting harder to solve. Even as platforms roll out content credentials and provenance tools, Wikipedia has moved to ban AI-generated text entirely from its 7.1 million articles, citing hallucinations and fabricated citations. The South Florida Standard is a case study in why those bans exist.

Frechette’s analysis traced the South Florida Standard to at least two sister sites: the Charleston Sentinel in South Carolina and the San Francisco Download in California. All three were built from the same source code and controlled by the same entity. Across the network, The Florida Trib identified at least nine “reporters” who share names with people accused of or convicted of fraud or conspiracy.

Chapin acknowledged responsibility for the network, describing it as a six-month experiment to build what he called “geographic topical authority,” testing whether AI-generated news sites could rank alongside legitimate outlets in search results. He said he stood up 17 similar sites across the country, producing more than 3,500 URLs drawing more than 44,000 visitors.

“I don’t know whose job it is to make sure that people are represented fairly and wholly online,” Chapin told the outlet.

The experiment, by Chapin’s own account, didn’t work particularly well. Search engines, he said, could tell the difference between The New York Times and a fake outlet. His sites weren’t breaking through.

But the broader damage is harder to quantify. Known in academic research as “pink slime” journalism—named for the cheap meat by-product used as a filler—these fake news sites now outnumber local daily newspapers in the United States, according to data analysis firm NewsGuard. As of June 2024, NewsGuard identified 1,265 such outlets nationwide, surpassing the 1,213 daily newspapers still operating.

Florida, with the lowest number of news outlets per capita in the continental US, may be especially vulnerable. Researchers have found the state is already home to dozens of pink slime outlets, part of a national network of more than 1,000 sites backed by conservative think tanks, donors, and political operatives.

“Stuff like this has zero value to the public,” said Kelly McBride, a senior vice president at The Poynter Institute. “And in fact it has a negative impact on the news ecosystem, because it clutters the environment.”

The findings arrive as traditional local news continues its historic collapse. Real newsrooms are already experimenting with AI-generated content, and their own workers are pushing back. Since 2005, the country has lost almost 2,900 newspapers and roughly two-thirds of its newspaper journalists—43,000 positions—according to Northwestern University’s Medill Local News Initiative. As real local news disappears, the vacuum doesn’t stay empty for long.

Edited by Pete Pachal

The post A fraudster built a network of fake AI news sites to manipulate search results appeared first on The Media Copilot.

How can you tell something’s AI-generated? When it comes to writing, there are common tells: the excessive use of em dashes, sentences that are too rhythmically clean, and a general smoothness that feels overly engineered.

It’s hardly a perfect science, though, and most humans’ AI detection skills are based on vibes.

If humans are just relying on instinct, what are AI detectors relying on? Here, Zapier shares everything you need to know about how AI detectors work.

What is an AI detector?

An AI detector is a tool that analyzes content like text, images, or videos, and estimates the likelihood that it was generated by an AI model. Instead of giving a definitive yes-or-no answer, most AI detectors will give you:

• A probability score (for example, “74% likely AI-generated”)
• A confidence rating
• Highlighted passages that appear machine-written, if it’s text

Their goal isn’t to “catch” AI with certainty, but to flag content that statistically resembles AI-generated patterns.

How do AI detectors work?

The specifics of how AI detectors work vary depending on what type of content they’re analyzing. For simplicity, this article will focus on AI text detectors. But other types—like AI image detectors—work similarly.

Large language models (LLMs) generate text by predicting the most likely next word based on probability. It’s more nuanced than that, but that’s the idea. AI detectors reverse-engineer that idea: They look at a finished piece of writing and measure how closely it matches those probability patterns. Here are the main techniques they use.

1. Perplexity

Perplexity (not to be confused with the AI-powered search engine) measures how unpredictable a piece of text is to a language model. The lower the perplexity, the more the wording follows patterns the model expects to see.

AI-generated text often has lower perplexity because it’s built from highly common word sequences. It gravitates toward phrasing that’s safe, common, and structurally sound. Which is kind of the point. AI models are trained to predict the most probable next word, not the most chaotic or idiosyncratic—just the most likely.

Human writing, on the other hand, tends to raise the perplexity score because it’s usually less predictable. Unless you have a ruthless editor who’ll set you straight, humans use words that technically work, even if they’re not the exact right ones. They go off on tangents and litter their work with comma splices because those pauses just feel right to them.

2. Burstiness

Burstiness looks at sentence length distribution and structural variation to identify patterns that appear overly consistent.

Humans rarely write in perfect cadence. They mix short sentences with longer ones, occasionally go on tangents, and vary pacing without thinking about it. Earlier AI models, by contrast, tended to produce writing that felt evenly spaced and neatly balanced. Nothing was outright bad, just … suspiciously consistent.

That “too rhythmic” quality is often what sets off our internal AI radar. AI detectors try to quantify that instinct by measuring variation in sentence length, punctuation, and structure. If the tempo barely changes from start to finish, that uniformity can raise a flag.

3. Classifiers

A classifier is a machine learning system trained to categorize text as likely human- or AI-generated. Unlike perplexity or burstiness, which are individual signals, a classifier looks at many features at once and weighs them together.

Developers train their LLMs on large datasets of labeled human and AI text. Through that training, classifiers learn statistical patterns that tend to separate the two categories. Those patterns can include predictability scores, sentence variation, word frequency distributions, and other structural signals.

When you paste new text into an AI detector, the classifier evaluates how multiple signals interact and then produces a probability score. The final output reflects whether the writing, on average, more closely resembles patterns associated with AI-generated text or human-written text.

4. Stylometric analysis

Stylometric analysis is the study of writing style, including vocabulary richness, repetition, and sentence complexity. Think of it as your linguistic fingerprint.

The idea is that humans tend to develop quirks over time. For example, the author Fredrik Backman typically writes stories with a sort of progressive repetition that’s hard to describe, but is uniquely him. It’s what makes his writing so easily distinguishable.

AI writing, by contrast, often clusters around high-probability patterns, generating phrasing that reflects widely represented patterns rather than highly idiosyncratic ones. That’s also what makes much of AI writing feel technically solid but vaguely same-y.

5. Watermark detection

Watermark detection is a way of identifying AI-generated text by looking for a hidden signature baked into the writing itself.

Not all AI models use watermarking, and there isn’t one standard way to do it. But when watermarking is enabled, the model slightly nudges its word choices in a consistent, trackable way. The shifts are subtle enough that you wouldn’t notice anything while reading, but an AI detector that knows what to look for can spot the pattern.

In theory, that makes AI-generated content easier to trace. In reality, even light editing or paraphrasing can blur or erase the signal. So while watermarking sounds like a clean solution, it’s not foolproof.

How accurate are AI detectors?

AI detectors are probabilistic tools, not lie detectors. A detection score reflects how closely writing matches certain patterns. It doesn’t prove who or what actually wrote the text.

Here’s why accuracy gets complicated.

• False positives happen. Some human writing naturally resembles AI-generated text. If you refuse to give up the em dash and sprinkle them liberally throughout your writing, an AI detector may flag it as machine-written, even if it wasn’t.
• False negatives happen. AI models are improving at an alarming speed and learning to mimic human variability more effectively. Humans, for their part, are learning to refine their AI prompts to inject human signals—for example, telling their AI writing generator to mix up sentence patterns or intentionally include errors. As AI writing and human prompting become more nuanced, detection becomes harder.
• Hybrid content blurs the line. Most writing today isn’t purely human or AI. AI detectors struggle in this gray area because the final text contains both human and machine signals.
• Results vary across tools. Different AI detectors use different training data and different models. The same paragraph can receive dramatically different scores depending on the platform. That inconsistency makes it risky to rely on a single detection result for high-stakes decisions.

The bottom line on AI detectors

We’re no longer living in a binary world of purely human or purely AI-generated writing. A lot of content now sits somewhere in between. A draft may start with AI, a human reshapes it, AI tightens a paragraph, a human adds a lived example—the lines blur. And AI detectors have to make probabilistic guesses in that gray space.

This story was produced by Zapier and reviewed and distributed by Stacker.

The post How do AI detectors work? appeared first on The Media Copilot.

ArXiv, the open preprint repository that has become a cornerstone of scientific publishing, is escalating its fight against low-quality, AI-generated research with a new enforcement policy: one year-long ban for authors who fail to take responsibility for LLM output.

As reported by TechCrunch, the rule, announced by Thomas Dietterich, chair of arXiv’s computer science section, targets submissions that contain “incontrovertible evidence” that authors did not check AI-generated content. That evidence includes hallucinated references—fabricated citations—a problem research shows is on the rise in scientific literature. It also covers direct LLM comments or instructions left inside submitted manuscripts.

“If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can’t trust anything in the paper,” Dietterich wrote on X.

The penalty is steep. Offending authors face a one-year suspension from arXiv followed by a requirement that all future submissions first be accepted through a peer-reviewed venue before being posted. Dietterich described it as a “one-strike” rule, though moderators must flag violations and section chairs must confirm the evidence before a ban is imposed. Authors retain the right to appeal.

Crucially, arXiv is not banning LLM use outright. Researchers may use AI tools in their workflow, but they bear full responsibility for the final content. That means if an author copies AI-generated text containing plagiarized material, biased content, fabricated references, or factual errors, they face consequences regardless of the tool that produced it.

“Authors take full responsibility for the content of their submission, irrespective of how the contents are generated,” Dietterich said.

The move comes as arXiv navigates a broader transition. After more than 20 years under Cornell University, the organization is becoming an independent nonprofit, a shift expected to give it more resources to address challenges like AI slop. ArXiv has already required first-time posters to obtain endorsements from established researchers, a barrier aimed at weeding out low-effort submissions.

Fabricated citations have become a flashpoint. Peer-reviewed research published in The Lancet found that fake citations are increasing in biomedical literature, a trend researchers link partly to LLM use. Scientists are not alone—lawyers, firms, and other professionals have also been caught submitting AI-hallucinated citations to courts and other authorities.

The policy puts arXiv ahead of most academic platforms in explicitly addressing AI authorship accountability. Whether it deters misuse or simply shifts it elsewhere remains to be seen.

The post ArXiv will ban researchers for a year if they let AI write their papers appeared first on The Media Copilot.

Fake accounts are as old as social media itself. So when it came to light that a “hot girl” MAGA personality named Emily Hart was actually a 22-year-old male medical student in India, it could have been dismissed as just another internet deception story. Just another catfisher, another sock puppet, another scammer—the internet is full of them.

But this case was different. This one had photos. And videos. And thousands of followers across multiple networks with some posts getting millions of views. Emily Hart was a full-on influencer, not just some anonymous egg. The person who created Emily confessed to Wired that while the account was active, he was making thousands of dollars every month from posting softcore videos to an OnlyFans competitor and merchandising.

Emily’s creator is not a developer. He’s just a cash-strapped student with a good sense of American political culture and a Google Gemini account. Yet the Emily Hart story has done more than expose one fraud. It’s put a spotlight on how thoroughly AI has lowered the barrier for almost anyone to produce convincing content and manipulate social media’s engagement systems.

That reality raises a set of urgent questions. Is anyone looking out for us out there? How can you tell what’s real and what’s not anymore? And who is responsible for alerting social media users that the images they’re looking at might have come from AI?

How cheap AI tools made fake influencers scalable

The real significance of the Emily Hart story has little to do with a single fake account. The major implication is that this is the tip of the iceberg. AI has made creating online personas like Emily so easy that it’s enabled deception at scale. The Wired story points to other pro-Trump fake influencers like Jessica Foster, but you don’t have to look very far in your Instagram Explore page before you spot something AI-generated, and it’s rarely disclosed. The Emily Hart case proves that the template is cheap, fast, lucrative, and easy to copy.

Every major social network has policies that address AI-generated content. While they vary in detail, the gist is generally the same: Synthetic images must be disclosed—especially if it could be construed as real and the subject matter involves sensitive subjects like politics, health, finance, and current news. If the account doesn’t identify AI content, it could be frozen, demonetized, or banned.

In practice, those consequences almost never materialize. Enforcement is difficult, partly because detecting AI content is getting more difficult by the day. Most state-of-the-art image generators are light-years ahead of the models that created the first “Will Smith eating spaghetti” video, and telltale artifacts like extra fingers and disappearing background characters have largely become a thing of the past. Without watermarks, even automated systems have a difficult time parsing AI images from real ones just by looking at them.

Content Credentials and the AI labeling problem

A new standard was supposed to fix this. Content Credentials are a way to track how an image was created and modified throughout its life cycle. That provenance data can live in the image’s metadata, so the site displaying it can more easily tell whether it’s AI-generated, potentially passing on a label or warning to the user. The idea is that, as you scroll your social feed, any image would have a tiny icon next to it that would reveal its history when clicked.

However, even though this technology has existed for years and ostensibly has the support of major tech companies such as Adobe, Google, and Nvidia, social platforms haven’t adopted it consistently. Seeing the label is rare, and a Washington Post report found that social networks often strip out the metadata that enables Content Credentials. The stripping isn’t necessarily a deliberate act of sabotage — it follows a best practice from the early days of the web when every byte was precious. But the fact that it’s still happening shows there is little enthusiasm to make the system work.

Does labeling even change behavior? Emily’s creator says he believes many of his followers didn’t care whether the images he was posting were AI or not. That may be true for some, but data suggest labels can alter people’s propensity to engage with AI content. A 2024 study found that labels on AI-manipulated media reduced belief in the claims. The study also found that wording matters: “manipulated” or “false” were more impactful than process-based labels alone.

Put another way: labels work, but toothless labels work poorly. A buried “AI info” tag is not the same as a clear warning that an image might depict a person who does not exist.

The technical capacity to do better clearly exists. Platforms like Facebook, Instagram, YouTube, and TikTok already process and modify content at scale. They’ve spent two decades building the art of detecting copyright violations, nudity, spam, and engagement signals. It is hard to believe they are incapable of building a clearer label for AI-generated people.

Why platforms have reason to keep AI labels weak

The question then becomes: why haven’t they? The uncomfortable answer is that the incentives point the other way. While platforms want to keep bad content out, they are more motivated to keep people posting, scrolling, sharing, and buying. AI-generated material fits neatly into that machine because it is cheap to make, easy to personalize and highly compatible with engagement-driven feeds.

Mark Zuckerberg has been unusually direct about this, describing AI-generated material as “a whole new category of content” that he sees as important for Facebook, Instagram and Threads. That framing doesn’t signal that Meta or any other platform actively wants deception — deception is a subcategory of AI content, not the whole thing. But it does mean the companies have a business reason to welcome more synthetic content, and making the labels too strong or too visible could dampen the engagement they’re trying to encourage.

External pressure could shift the math, though. Europe’s AI Act includes transparency obligations for deepfakes and certain AI-generated public-interest content, with related rules taking effect this year. Should platforms start to rack up major fines for poor labeling, things could change in a hurry. Advertiser pressure would help, too, since appearing next to deceptive content is bad for business. Finally, and crucially, there’s audience behavior: if users begin to feel like they can’t trust what they’re seeing on a network, they might, over time, stop engaging with that network.

The disclosure system failure

At the moment, detecting AI content has become largely the user’s problem, with social platforms not prioritizing the technical progress that might help, and regulators only beginning to act. And you might question what’s the point — many of Emily’s followers no doubt knew she was virtual but followed, engaged, and maybe even forked over some money anyway. But that calculus depends entirely on having information. The choice to engage or not with a virtual influencer is robbed from you if you don’t know it’s virtual in the first place.

The technology industry has spent years presenting provenance as a central answer to synthetic media. Adobe, Microsoft, Meta, OpenAI, Google and others have backed standards, joined coalitions, made public commitments and embedded Content Credentials into their tools. Fine. Then show it to people. Make it visible before the share, before the follow, before the subscription, before the merchandise purchase. Because if the only way to learn that an influencer is fake is to wait for a magazine investigation, the disclosure system has already failed.

A version of this column appears in Fast Company.

The post Why AI content labels keep failing the people who need them most appeared first on The Media Copilot.

A concept making the rounds in AI circles is something called “liquid content.” The shorthand describes the act of reshaping facts, ideas, and expressions across mediums. The most well-known example is a feature within Google’s NotebookLM: Once you’ve filled a folder with various kinds of data, it can whip up a podcast about that data, enlisting a couple of cheery AI-generated voices to give you an overview, analysis, or debate.

Push the idea to its limit and you arrive at a vision where any piece a media company produces can flow into every other format on demand. Making a podcast? With the right tools and prompting, in mere minutes, it can be reimagined as a series of clips, a feature article, or even an interactive presentation. And for a traditional news publisher, the same archive of articles can fuel videos that previously got benched as too costly to bother with.

This is no longer a thought experiment. I recently attended a couple of industry conferences—the NAB Show and Adobe Summit—and tools that can intelligently translate one format into another are showing up all over the floor. Just two examples: Amagi showed off an AI system that can scan a live newscast, understand the different stories covered, and create short-form videos for each one on the fly, populating a TikTok or Instagram feed almost as soon as the news is out. Stringr‘s Genna system, meanwhile, can take any news article and turn it into a video, pulling photos and licensed clips from repositories like Getty to assemble the footage.

Repurposing isn’t new, of course. But now that artificial intelligence can do most of the heavy lifting—interpreting the content, determining how it’s best expressed in a new form, and then pulling all the levers to do the actual work—the work moves faster, costs less, and scales in ways that weren’t possible before.

Automation doesn’t remove the hard part

If you sense a “but” coming, your instincts are right. AI can be a great catalyst in reimagining content, but it doesn’t solve every problem associated with pushing into new formats, and it tends to invent a few of its own. The opportunity is real, but treating liquid content like a magic growth engine is a mistake. It’s better understood as a new production layer that needs careful tending. As media companies turn to AI to expand their content footprint, here are a few realities worth keeping in mind.

1. Generative content produces diminishing returns. A small but important distinction first: there’s a difference between using AI to assemble content and using it to create content. It’s particularly relevant in visual media, where accuracy in the imagery matters greatly.

Setting aside the obvious ethical knots that come with generative video in news, there’s a quieter issue: Audiences don’t respond to it in the same way. Inception Media is a podcast company based on AI-generated scripts and synthetic voices. It does respectable numbers, but they’re far below what it might get from human-driven shows.

AI is a powerful accelerant, but listeners and viewers still reward authenticity. A publisher dipping a first toe into podcasting or short-form video with synthetic content may be disappointed by the audience numbers. The safer route is to stick with non-generative content and simply use AI to assemble existing footage and imagery. But that still requires you to either produce or acquire that material, which eats into the cost savings the pitch deck promised.

2. Good AI needs good data. For AI to understand and interpret content reliably, the surrounding data has to be accurate and thorough. That means things like tags, categorization, metadata, dates, and notes (e.g., exactly who appears in a video) should all be present and correct.

Most operations have improved on this front, but the older you go in the archive, the less reliable the metadata gets, especially across system migrations that scrambled or lost it entirely. Messy data is the rule rather than the exception in media, and it will keep many outlets from getting full value out of their back catalogs.

3. Humans still run the show. AI is a tool that gets better and more versatile every day, but it’s still far from perfect. It can hallucinate and misinterpret, and because it lacks experience with the real world, it sometimes makes mistakes humans never would (pointing out that volleyball is hard to play without a ball, for instance). Audiences have low tolerance for slop or poor quality.

The point is simple: AI can do plenty, but it still needs people. And not just to spot-check the output. Venturing into new platforms requires more strategic thinking than simply putting the content out there. To zero in on just one use case: AI can do competent translation, but launching into a new market is still a long, deliberate exercise in management and care.

The back catalog gets more valuable

All that said, if you crack the playbook, AI as a content-repurposing engine has serious upside.

1. Archives are a gold mine. Most outlets will reshare evergreen “hits” on social media, which can drive a decent amount of views. AI takes that idea further: not just resharing an article once, but extracting the best parts and turning each “nugget” into its own video, gallery, or social post. It can also reinvent the “this day in history” trick by spotting current trends and pulling forward older stories that map cleanly onto them.

2. A way into younger audiences. Many small and midsize outlets simply haven’t had enough content to really monetize on a platform like YouTube or Instagram Reels. Success is often a numbers game, demanding regular posting to even have a hope of showing up in someone’s feed. AI-assembled video won’t compete with MrBeast for attention, but it does give a brand a foothold with younger viewers, 63% of whom primarily get their news from these platforms.

3. Lean teams can do the work. Venturing into a new platform used to require weeks of study, hiring dedicated staff, and building out a strategy. Now AI can accelerate all of that—not just the remixing itself. As already mentioned, humans still need to manage the process and have the final judgment over whatever’s produced, but building a content-remixing department won’t be nearly as expensive as a pivot to video.

The harder question hangs over all of it: will the ROI justify the effort? As more media adopts remixing strategies and agentic systems, the inevitable result will be a large increase in supply of repurposed content—especially video. More supply tends to mean less demand per piece, which thins audiences further. In aggregate, the revenue lift from a remixing strategy may end up modest.

There’s a wrinkle, though. For niche publications with few competitors, there’s less of a danger of saturating their market, and making a move to a multimedia strategy—on a smaller budget—might improve audience growth and retention with readers who prefer formats like video and podcasts. Local and regional outlets fall into the same bucket.

The dream of a general-purpose content engine that can reliably spin out engaging stories in any format is getting less fictional by the day. But the engine is still just an engine. Building a successful strategy around it requires intention, careful curation, and a strong understanding of both the audience and the platform they’re on. Liquid content is a powerful idea. Pouring it well is still an art.

A version of this column appears in Fast Company.

The post Why liquid content is harder than it looks appeared first on The Media Copilot.

My usual focus is the cutting edge of AI in media, examining how journalists and media companies are using the technology to change the way they work, reach new audiences, and transform their organizations. But the reality is that a persistent stigma still hangs over artificial intelligence in the journalism world. In conversations I have with working reporters and editors, there’s clearly still a lot of reluctance, if not outright disdain, for using AI in almost any part of their work.

Recent media coverage, though, paints a different picture. The Wall Street Journal recently profiled how Fortune business editor Nick Lichtenberg uses AI to turbocharge his output, sometimes writing as many as seven stories in a single day. The same day, Wired highlighted how several prominent reporters—including independents like Alex Heath and Taylor Lorenz as well as The New York Times’ Kevin Roose—use AI in various editorial tasks, sometimes in the writing itself.

Taken together, it feels like a dam has finally burst. And I don’t think the timing is accidental—this shift is happening alongside the arrival of Claude Code and Cowork, which has put remarkably powerful agentic AI within reach of everyone and reshaped what people expect from these tools. (An interesting aside buried in all this coverage of journalists’ use of AI is that it appears Claude is rapidly becoming what the Mac became among media pros: the platform of choice for creatives who “know better.”)

A plagiarism scandal puts AI trust on ice

But just as the relationship between journalists and AI seemed to be thawing, a high-profile incident threw it back into doubt. Last week, The New York Times severed its relationship with a freelance writer who had submitted a book review that was at least partially AI-written. The review by Alex Preston, published in early January, included passages that were nearly identical to Christobel Kent’s review of the same book that was published in The Guardian months earlier.

Preston admitted he used AI to assist in writing his book review, saying that he had “made a serious mistake.”

The episode is a clear wake-up call for the Times—and not its first—about communicating AI policy to freelancers. But it also sends a warning signal to every newsroom that has been inching toward greater AI adoption. Here, suddenly, was an error that appeared to validate all the restrictive rules.

Confronting what happened directly matters. The incident steers us back into the dark cave of AI scandals in media—from CNET’s bot-authored service journalism to the made-up book titles in the Chicago Sun-Times’ “summer reading list” last year. It risks erasing the productivity and content optimization gains that many journalists and newsrooms have been making, and could push those just beginning to experiment with AI back toward the simplest possible rule: don’t use it at all.

That makes it essential to examine specifically how AI was deployed here, so we can draw a clearer line between responsible and irresponsible use. It’s easy to say there wasn’t enough “human in the loop” (an increasingly unhelpful term)—but where in the loop? With prompting, fact-checking, something else? The whole point of AI is to outsource some human decision-making to sophisticated machines, so rather than pointing out the obvious—that humans need to shape and monitor the process—it’s better to zero in on the specific decisions that AI was asked to make, and whether the human gave the right parameters and restrictions.

When you look at the details, the answer is clearly no. According to The Guardian story, the two reviews have eerily similar language—so close that it’s difficult to argue against outright plagiarism. Consider these side-by-side passages:

• Original review, published August 21, 2025: “most significantly a song of love to a country of contradictions, battered, war-torn, divided, misguided and miraculous: an Italy where life is costume and the performance of art, and where circuses spring up on wasteland.”
• Times review, published January 6, 2026: “populate what is ultimately a love song to a country of contradictions: battered, divided, misguided and miraculous. This is an Italy where life is performance, where circuses rise on wasteland.”

Given the dates and the undeniable overlap, a few things become clear. Preston evidently asked the AI—directly or indirectly—to generate text he planned to use in the piece, and not just from his own notes. The four-month gap between the two reviews (and likely an even longer lead time given the Times’ editing process) almost certainly means the AI’s training data didn’t include Kent’s review. That points to the AI tool pulling from web search (also known as RAG) to produce the copy.

This was the critical error. Giving Preston the benefit of the doubt, he may not have deliberately told the AI he was using to synthesize other reviews of the book, and perhaps it grabbed The Guardian review on its own. But he certainly didn’t tell the AI not to do that, which would seem to be an essential part of your prompt if you want to avoid the very plagiarized text he ended up including.

Moving from stigma to smart adoption

It bears repeating: in most cases, how you use AI matters far more than whether you use it. Getting there requires deep familiarity with these tools’ strengths and weaknesses, careful attention to prompt design, and a commitment to continuous adaptation. It’s an ongoing process, and it needs guardrails—such as “always” and “never” commands to avoid specific problems and (human) fact-checking. Without those safeguards, you’re handling a loaded weapon that can easily misfire.

Broader structural protections help, too. Whether you’re an independent writer or a full newsroom, it pays to have an AI policy. As a media AI trainer, I of course would encourage investing in training, but I think it’s still objectively a good idea. But most importantly, the trial-and-error that comes with figuring out the boundaries of “good AI” should be kept out of public view if you can avoid it.

When it comes to AI-assisted writing specifically, developing your prompts and safeguards in a private sandbox is critical. That might seem obvious, but one of AI’s most deceptive qualities is that it produces outputs that look indistinguishable from work that went through a rigorous human process. To someone without experience, that surface-level competence feels sufficient.

Truly making AI work as a writing and journalism partner means going beyond trusting the process—it means accepting responsibility for building, testing, and refining that process yourself. The more journalists do that, the more the stigma will fade.

A version of this column appears in Fast Company. It has been lightly “remixed” (alternate words and phrasings used) with AI assistance and human review.

The post Journalists are opening up about AI, but one mistake shows how fragile that progress is appeared first on The Media Copilot.

A single unreleased AI model has triggered emergency regulatory mobilization on both sides of the Atlantic. UK financial regulators are holding urgent talks with the government’s cybersecurity agency and major banks to assess risks posed by Anthropic’s Claude Mythos Preview — days after US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an emergency meeting with Wall Street’s top CEOs over the same concerns.

In the UK, officials from the Bank of England, Financial Conduct Authority, and Treasury are in talks with the National Cyber Security Centre. Representatives from major British banks, insurers, and exchanges are expected to be briefed on cybersecurity risks at a meeting with regulators within the next two weeks, according to Reuters. The BoE, FCA, and NCSC all declined to comment.

The US response was more public. White House national economic adviser Kevin Hassett confirmed on Fox News that Bessent and Powell had convened bank chiefs — including the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs — to warn of cyber risks from the model. JPMorgan CEO Jamie Dimon was unable to attend. The urgency of the meeting reflected the capabilities Mythos Preview has demonstrated in controlled testing: the ability to identify and exploit weaknesses across every major operating system and every major web browser.

Anthropic has stopped short of a broad release, citing concerns the model could expose previously unknown cybersecurity vulnerabilities at scale. The company has been navigating an increasingly complex relationship with the broader tech and media ecosystem as its models grow more capable.

What Mythos Preview is — and who can use it

Despite not being publicly available, Claude Mythos Preview is already in active use — under strict controls. Under a program Anthropic calls Project Glasswing, select organizations have been granted access to the model for defensive cybersecurity work. Partners include Amazon, Microsoft, Apple, Google, Nvidia, CrowdStrike, and Palo Alto Networks. Access has since been extended to approximately 40 additional organizations responsible for critical software infrastructure.

Anthropic says Mythos Preview has already found “thousands” of major vulnerabilities in operating systems, web browsers, and other software. The company has committed up to $100 million in usage credits and $4 million in donations to open-source security groups as part of the program.

The framing is defensive. But the same capability that finds vulnerabilities can, by definition, be turned toward exploiting them — which is precisely what regulators appear to be stress-testing.

Why regulators are moving fast

The simultaneous and independent responses from UK and US financial regulators signal that Mythos Preview represents a qualitatively different kind of AI risk than those regulators have previously had to assess. Prior AI regulatory concerns have centered on bias, misinformation, and systemic market risks — as seen in ongoing debates around AI copyright policy and AI use certification. A model with demonstrated offensive capability against critical software infrastructure — in active use, even in a restricted form — is a different category of problem.

It is also a compressed timeline problem. The model exists. It is being used. The regulatory frameworks to manage it are still being assembled.

All three UK agencies — the BoE, FCA, and NCSC — declined to comment on the talks. Anthropic had not responded to a request for comment at the time of the Reuters report.

The post UK and US financial regulators hold emergency meetings over Anthropic’s Claude Mythos appeared first on The Media Copilot.

The Associated Press, founded in the mid-1800s to help New York newspapers share reporting costs, is offering buyouts to an unspecified number of U.S.-based journalists — the latest move in a long-running transformation from wire service to technology data company.

The News Media Guild, which represents AP journalists, said more than 120 staff members received buyout offers on Monday. AP executive editor and senior vice president Julie Pace said the goal is to reduce global headcount by less than 5%, though she acknowledged the cut among U.S. staff would likely exceed that figure depending on how many people accept.

“We’re not a newspaper company and we haven’t been for quite some time,” Pace said.

The numbers back her up. Over the past four years, AP’s newspaper revenue has fallen 25%. Big newspaper publishers, once the organization’s financial foundation, now account for just 10% of income. Gannett and McClatchy both dropped AP in 2024. Lee Enterprises — publisher of The Buffalo News, the St. Louis Post-Dispatch, and the Richmond Times-Dispatch — is now seeking an early exit from a contract due to expire at the end of 2026.

Where the growth is coming from

While the newspaper business contracts, AP’s technology revenue has grown 200% over the same four-year period. Kristin Heitmann, senior vice president and chief revenue officer, put it plainly: “If you can think of a large technology company, they are a customer of ours.”

AP was among the first news organizations to move aggressively into AI deals, agreeing in 2023 to lease part of its text archive to OpenAI. It has since launched on Snowflake Marketplace to license data directly to enterprises, stood up AP Intelligence to sell data to financial and advertising sectors, and last year secured a deal with Google to deliver news through the Gemini chatbot — Google’s first content deal with a news publisher.

Elections data is another growth vector. AP saw a 30% increase in election data customers between the 2020 and 2024 cycles, and last month agreed to sell U.S. elections data to Kalshi, the world’s largest predictions market. ABC, CBS, NBC, and CNN all signed on to the AP elections service last year.

What the restructuring looks like

Beyond the headcount reduction, AP is doubling down on video — it has already doubled the number of U.S. video journalists since 2022 — and deploying rapid-response teams that contribute to major stories regardless of geographic base. The organization says it will maintain a presence in all 50 states.

The union is pushing back. In a statement, the News Media Guild said AP “refuses to offer [staff] appropriate training and tools” and is “flirting with artificial intelligence — ignoring the opportunity to differentiate AP news stories as ones that are and always will be created by human journalists.” The union also said AP declined a request last week to bargain over AI use.

AP did not immediately comment on either claim.

Pace framed the restructuring as a strategic choice made from stability, not distress. “The AP is not in trouble,” she said. “We’re making these changes from a position of strength.”

The post AP offers buyouts as AI and tech companies now drive revenue growth appeared first on The Media Copilot.