Fiction written by artificial intelligence is easy to detect because it struggles with complex story structure and tends to moralize in clunky ways, according to a preprint study from researchers at University of Maryland, College Park and Google DeepMind. They found that AI fiction has tells that go beyond stereotypical overuse of em-dashes and other obvious AI tropes and have more to do with the formulaic nature of the text itself.

“AI stories over-explain themes and favor tidy, single-track plots while human stories frame protagonists’ choices as more morally ambiguous and have increased temporal complexity,” the study, which looked at more than 50,000 AI-generated short stories, found. “Claude produces notably flat event escalation, GPT over-indexes on dream sequences, and Gemini defaults to external character description. We find that AI-generated stories cluster in a shared region of narrative space, while human-authored stories exhibit greater diversity. More broadly, these results suggest that differences in underlying narrative construction, not just writing style, can be used to separate human-written original works from AI-generated fiction.”

Basically, AI-generated fiction sucks and at the moment is easy to detect. The typical method of detection involves looking for stylistic markers such as an abundance of em-dashes, the overuse of the word “delve,” or an obsession with goblins, but this project tried something different. “The idea for this project came because we are hoping to eventually move past plain text detection, into some sort of space where we can separate human ideas from AI-generated ideas,” Jenna Russell, a University of Maryland researcher and one of the study’s authors, told 404 Media. Russell is also an intern at the AI-detection company Pangram.

Russell and her team decided to attempt to detect what she called “narrative features” in AI- generated fiction. The detector is called StoryScope and it builds on NarraBench, a 2025 benchmark that suggested a taxonomy of narrative features in fiction. StoryScope looked at how fiction handled plot development, character descriptions, setting, and temporal structure to determine if something was written by a human or an AI.

“It was my first attempt at getting 'under the surface' and focusing more on ideas,” Russell said. “We wanted to see how close to typical AI-detection we could get by only relying on the narrative features, to understand if this sort of structural difference really even exists. This method also adds some interpretability to detection, which is an open question in the field. Using narrative features, we can point to certain tangible features (such as the number of subplots included in a story). I think this is why it's struck a chord recently, people can really say ‘ah these are some of the underlying traits of how AI writes fiction.’”

To test StoryScope, the researchers selected 10,272 human-written stories then reverse engineered them into writing prompts using Gemini 2.5. Then it took those thousands of prompts and fed them into Gemini 3 Flash, DeepSeek V3.2, Claude Sonnet 4.6, Kimi K2.5, and GPT 5.4. All of the data — including the prompts and the resulting AI stories — are available on Hugging Face.

To source the stories, the researchers used the Books3 dataset — a database of 183,000 books collected from pirated ebooks. The dataset is the subject of several lawsuits and has been used to train an unknown number of LLMs. The StoryScope study included more than 10,000 of some of the most famous short stories ever written, many of them pulled from popular anthologies. There’s Joyce Carol Oates, Stephen King, Louis L'Amour, Charlotte Perkins, and Harlan Ellison. All have been rendered down to their base elements by AI and then regurgitated into a different LLM to see if it can replicate them.

Russell told me the dataset was controversial. “Hence why we do not release it to the public,” she said.

The study itself contained a disclosure. “We acknowledge the copyright issues related to the Books3 dataset and do not endorse its use for model training or commercial text generation,” it said. “The use of the dataset in our paper is restricted to academic purposes only and is meant to understand the narrative differences in human-written and AI-generated text to help inform discussions on AI-detection, authorship, and copyright policy.”

The various AIs, of course, can’t possibly replicate the prose of O. Henry. So what, according to StoryScope, are the narrative quirks of LLM-written simulacra of English’s grand works of fiction? 

AI tools tend to over explain themes, for one. 

“Narrators explicitly explain the story’s theme 77% of the time, versus 52% for humans: a grieving character’s arc will typically end with the narrator stating the lesson learned. AI dialogue serves philosophical debate more often (59% vs. 34%), and references to other works tend to be vague allusions (72% vs. 50%) rather than specific, named references. The pattern is one of over-determination: AI spells out meaning rather than trusting the reader to infer,” the study said.

AI also more often avoids subplots and fails to play with time jumps and flashbacks. The systems overwrite passages about the body and senses. “Where a human author might write that a character ‘felt afraid,’ AI renders fear as a tightening chest, cold sweat, and dimming lamplight,” the study said. Humans also spin more complicated narratives involving more characters and locations than AI can handle. Humans also reference other works of fiction, specific people and places in a way that AI struggles with.

A disclosure caught my eye at the bottom of the StoryScope study. “Large language models and coding agents (Claude Code and Codex) are used to aid with and polish writing and generate some tables and plots,” it said.

“I believe it's important to disclose AI use (and ideally think it should be more in-depth than I wrote in the paper),” Russell told me. “Most researchers are using AI, a lot of it seemingly 'slop' [...] but a lot of it is high-effort, good research. Also, technically you are supposed to disclose AI use for conference submissions, but most people don't. I want to help change that norm!”

She also explained a bit more about how AI agents helped shape the project. “I use AI agents to help implement the code (using the claude code / codex interfaces). I also use them as an editor during the writing process! They have access to the project codebase and the paper latex, so the agents can implement graphics for me much more quickly than I could,” she said. “They write comments and add to the paper draft, but I keep it all in different colors so I can manually review and accept/reject/edit any suggestions from AI. I am a big believer that AI can help or hurt writing, but usually helps when not used to create more internet 'slop'.”

I kept thinking about Harlan Ellison and Robert Silverberg’s story “Ship-Shape Pay-Off” being turned into an AI prompt and then spit back out by an LLM. Ellison died in 2018 and was notoriously protective of his work to the point of violence. He successfully sued James Cameron for plagiarism over The Terminator. I have a hard time imagining he’d be happy to see his story pumped into a machine, no matter the results.

“A lot of people, like teachers or readers, don't really care if AI was used in the writing process, but do care if the human is the one behind the heart of it,” Russell said. “A teacher wants to know if their student understood the lesson, and a reader wants to know that the creativity behind a touching story was truly the work of the human author.”

This trip is also squarely in the “I didn’t realize that I don’t take many pictures” era of my life, so you’ll have to forgive me for not having too many pics, and trust that I’m going to take more in the future. The most important thing here is that I have an overabundance of pictures of the star of my trip, that sassy pelican.

Just look at that bird. That’s a bird with secrets.

Patreon announced on Thursday that it’s partnering with Cloudflare to block crawlers from stealing creators’ work to train AI models.

“I HAVE A KICKASS PRODUCT UPDATE FOR YOU ALL!” Jack Conte, the founder and CEO of Patreon, wrote in a post on Instagram with the superimposed text, “POV: you're CEO of one of these fucking tech companies, so you do what you want.” 

“Patreon has partnered with an internet infrastructure company called Cloudflare to block Al training crawlers from using the work you publish on your Patreon to train their Al models,” Conte wrote. “This is live and happening at the network level on all posts published on Patreon.”

"As AI agents become increasingly powerful and popular, creators deserve a meaningful say in how their work is used by AI companies. On most of the Internet, creators have to accept AI training on their work just to reach and grow an audience," Drew Rowny, SVP of Product at Patreon, said in a press release published by Cloudflare last week. "Patreon has a different vision: creators should be able to grow their audience and control how their work is used. That's why we're building on our existing work with Cloudflare to block known AI training crawlers at the network level across Patreon, while still allowing the crawlers that help creators get discovered and grow their businesses through search."

Last year, internet infrastructure company Cloudflare, which provides cybersecurity protection and content delivery services to websites, announced that it would start blocking AI crawlers from accessing content without website owners’ permission or compensation by default. And earlier this month, Cloudflare announced new options for website owners to control AI traffic based on whether bots are search, agent, or training crawlers. In September, according to the company’s blog, all new domains onboarding to Cloudflare will have training and agent bots blocked by default on pages that display ads, while search crawlers will remain allowed by default.

“Creators deserve credit, compensation, and consent. If that's not on the table, the crawlers can stay the fuck off Patreon. The free internet is alive and happening. The rebellion has already started,” Conte wrote in his post. 

In May, Conte posted a 43-minute video addressing how the AI industry fails to compensate creators. “Creators deserve consent, credit and compensation,” Conte said in the video. “Consent meaning, ‘Do I get to opt out of my work being used by these models as training data?’ Credit meaning, ‘If my work is used and you just replicate my whole vibe as an artist… do I get credit for that?’ And then compensation, meaning, ‘Do I get paid when that happens?’ Unfortunately, the answer to all three of these questions right now is a big fat ‘No.’”

AI-generated works are permitted on Patreon, as long as they comply with the platform's terms of use. In 2024, 404 Media reported that many creators of nonconsensual sexual images and videos monetized their content on Patreon. Last year, Patreon updated its content guidelines for AI content to state: “AI-generated depictions of people that are illustrated/animated are permitted; AI-generated hyperrealistic depictions of people are permitted only if the people are real and have documented their explicit consent.” 

Cloudflare did not immediately respond to a request for comment.

Updated 7/9 at 7:59 p.m. EDT to include Drew Rowny's statement.

Wednesday, John Deere agreed to give farmers broader access to repair their tractors and farm equipment under an antitrust settlement agreement with the Federal Trade Commission, one of the biggest wins in the long right to repair battle. The settlement is the latest and by far the most important development in several recent lawsuits against John Deere, and is finally an agreement that isn’t full of half measures and doesn’t have massive, obvious loopholes.

The FTC settlement is far better than a recent, highly controversial settlement in a separate class action lawsuit against Deere brought by farmers in Illinois, and it’s worth breaking down the differences. Two years ago, I wrote an article called “The Walls Are Closing in on John Deere’s Tractor Repair Monopoly,” which followed that Illinois case, in which several farmers brought a complex, class action antitrust lawsuit against Deere. The judge in that case, Iain Johnson, wrote several scathing opinions about Deere’s anti-repair practices that indicated that he was seemingly inclined to hit Deere with stiff penalties. 

But after years of litigation, the plaintiffs in that case decided to settle with Deere in April, earning a $99 million payout for farmers who paid for repairs over the last decade, and several right-to-repair protections that did not have much in the way of legal teeth.

This $99 million payout was roughly $79 million after legal fees and to be divided among more than 200,000 farmers; this means each farmer will receive roughly $395, or “less than the cost of a single authorized dealer service call for a typical 500-acre farm,” according to an analysis by Willie Cade, a longtime farm right to repair advocate.

“Bottom line is that farmers are getting $0.79 per acre for the eight years of Deere abuse,” Cade told me. “Bad settlement. The settlement is insufficient … the money is a small fraction of what the class could recover at trial, the claims process depends on labor-hour data only Deere holds, and the repair "fixes" are riddled with loopholes that leave Deere's monopoly intact.” 

Demand Is Booming for New No Tech, Repairable Tractor

“There is consumer pressure to back away from technology that is unnecessary to perform everyday tasks.”

The Illinois settlement would prohibit farmers covered by it from filing any future repair-related litigation against Deere, and only required Deere to provide parts and repair guides to farmers under poorly defined “fair and reasonable” terms, a loophole that other manufacturers have used to claim that their parts and tools are constantly out of stock or cost astronomic prices. 

“The ‘fair and reasonable terms’ standard is not price equality with dealers, nor is it a guaranteed price ceiling,” Cade wrote in his analysis. “Disputes about whether Deere’s pricing meets this standard are subject to Court oversight, but individual farmers may have limited practical ability to challenge pricing that does not obviously cross the line.”

The settlement in the Illinois case was so bad that one of the plaintiffs in the case, Wilson Farms, filed a 53 page formal objection to it two weeks ago, in part because it claims that there are many “unlitigated and uncompensated” cases in which farmers suffered under Deere’s monopoly. Under the settlement, farmers would no longer be able to sue Deere by “terminat[ing] Class members’ ability to collectively challenge Deere’s repair aftermarket monopolization for a generation.”

“Rather than provide any meaningful benefit to the Class, it appears that the proposed Settlement’s most important effect will be to give Deere its most powerful tool yet in its decades-long effort to block farmers from repairing their own equipment,” the objection says. “Extinguishment of farmers’ rights under the law.”

Other farmers called the Illinois settlement “disingenuous” and “unfair.”

The good news is that the wildly disappointing and seemingly unnecessary selling out of farmers’ rights in the Illinois case that Deere appeared to be losing very badly is greatly mitigated by the FTC’s settlement from this week. The FTC case was brought by Lina Khan under the Biden administration; to its credit, the Trump administration decided to continue litigating.

The FTC settlement does not have monetary damages for farmers, but it has far better right to repair protections for John Deere customers moving forward. In the FTC deal, the “fair and reasonable terms” are better defined and are based on the price that John Deere dealers actually pay for repair parts and tools. Deere and its dealers are not allowed to “discriminate or retaliate” against farmers who repair their own equipment (manufacturers have been known to brick devices that consumers fix themselves). The FTC settlement also includes access to farmers for “future repair resources,” meaning repair tools, guides, software, and parts that Deere creates in the future. 

Deere must also file “compliance reports” with the FTC, and the FTC will have oversight of the compliance. Crucially, the FTC settlement also does not affect farmers’ private grievances against Deere, meaning it is possible for farmers to sue Deere if the company’s repair practices have affected them. 

The FTC settlement is one that has actual legal teeth and enforcement mechanisms that Deere should at least theoretically have to comply with. Earlier agreements and right to repair “wins” for farmers were often half measures (though it’s worth mentioning that Colorado passed a good agriculture right to repair law in 2023 after years of struggle from farmers and advocates). Deere and various farmers’ public interest groups had previously agreed to right to repair “memorandums of understanding” in which Deere promised to make repair parts and tools available to farmers. In practice, however, these tools and parts were often not available, were not as good as what dealers and authorized service providers had access to, or were unreasonably expensive. These memorandums of understanding also had few or no enforcement mechanisms. 

Cade told 404 Media in an email that this settlement order “gives farmers real hope.” 

Nathan Proctor, senior right to repair campaign director for consumer rights group U.S. PIRG, said in a statement that the FTC settlement “is much better than the deal secured in [the Illinois] class action lawsuit.”

“Deere has now agreed to make available all materials needed to conduct repairs, including some which it has previously withheld,” Proctor said. “I want to thank the FTC for its work on this case. Our goal from the start of our campaign was to ensure that farmers and independent mechanics get everything they need to fix equipment. We will continue to monitor the situation and advocate to ensure that goal is a reality.” 

In other words, farmers finally have an actual, major win in the right to repair fight that goes far beyond earlier piecemeal and moral victories.

A shocking amount of the content that users encounter on popular social media websites is likely AI generated, according to data from a company that detects AI writing. As much as 41 percent of longform written content seen by users on LinkedIn is likely to be fully AI-generated and roughly a third of longer posts on X are AI-generated; roughly one-in-ten longer Reddit and Substack posts are AI, according to the data. 

The data was collected using a Chrome extension from Pangram, a company that detects AI-generated writing. Pangram’s Chrome extension scans writing that users encounter while browsing and determines if any given post is likely AI-generated or likely human written. Because Pangram works passively in the background while a user is browsing the internet, it only scans posts that its users actually see. This helps answer the question of whether AI slop is actually poisoning the internet that humans actually use, versus polluting the internet more broadly. The answer is unequivocal: AI slop writing is not just sequestered off on unpopular automated SEO farms or spam sites that no one reads; humans are regularly wading through AI dreck on hugely popular sites. 

“This isn’t something that had really been studied before—how much AI content people are actually seeing,” Max Spero, the CEO of Pangram, told me in a phone interview. “AI content is a tax on readers’ time.” 

(Pangram formerly advertised on 404 Media. I am covering this data because I have written many articles about how AI-generated content is taking over social media and is brute forcing social media algorithms, and I have not seen other data that attempts to measure the actual popularity of slop.)

For this research, Pangram specifically asked users of its Chrome extension to opt-in to share Pangram browsing results with the company. The company analyzed roughly a million posts that its users organically scroll through across LinkedIn, Medium, X, Reddit, and Substack over a two-month period. Pangram found that, universally, longer posts on all platforms are more likely to be AI-generated than shorter posts. The company split the content it analyzed into “shortform” (between 50 and 250 words) and “longform” (longer than 250 words). 

The data suggests, perhaps unsurprisingly, that a huge portion of longform posts on LinkedIn and X’s new article format are fully AI-generated or AI-assisted (meaning drafted, edited, or rewritten by AI with some human elements). Forty percent of longform LinkedIn posts analyzed in the data were fully AI-written; a quarter of X articles were fully AI written, but another 23 percent of X articles were AI-assisted, the company said. It intuitively makes sense that longer form content is more likely to be AI-generated, because people usually won’t bother to AI-generate a few word response or a pithy comment on a quote tweet, for example. AI is also famously verbose, meaning AI-generated content is more likely to show up in longer posts.

“Our data shows that AI-generated content is a problem across all platforms, and it is hitting longform content especially hard,” the company wrote in a blog post. “Contrary to what one might expect, people are overwhelmingly willing to use AI to speak on their behalf in professional settings that are associated with their real identity, and less likely to use it on casual and anonymous platforms.”

The study also found that top-level posts on LinkedIn and Reddit are far more likely to be AI-generated than the comments underneath an original post. 

I have been using the Pangram Chrome extension for several months now, after interviewing Spero for an article I wrote called “Your AI Use Is Breaking My Brain.” In that article, I wrote about the cognitive weight of the constant assessments I am doing when I’m browsing the internet, trying to determine whether a piece of writing is AI-generated or not. After writing that article, I decided to try the Pangram Chrome extension to see whether its assessments of likely AI-generated writing aligned with my own brain’s assessments. After using the extension for nearly two months, my experience has largely aligned with what Pangram’s data suggests: Many of the longform articles I see on X are obviously AI generated, and are detected by Pangram as such. A huge amount of the LinkedIn posts I see are obviously AI-generated.

Because of the way the study worked, by passively detecting AI generated content that people see in their normal browsing, the data is potentially more useful than other studies that have sought to estimate the raw percentage of AI-generated content on the internet, but not whether anyone was actually seeing that content. These prior studies, which found that as many as a third of new sites are AI, allowed for the possibility that AI-generated content was flooding the internet but that it was of such a low quality that actual people may not have been seeing it. 

The Pangram data raises questions about what platforms are doing to promote or disincentivize AI slop. LinkedIn, for example, had for years built AI writing tools into its platform meaning that it has been incredibly easy to post AI-generated content on the platform and that AI-generated content became incredibly common on the platform. In May, the company announced that it is trying to disincentivize AI content in the name of “keeping conversations real,” and the AI writing assistant is no longer built into the post button. Reddit, meanwhile, has become a vector for companies trying to game LLM tools by promoting their products on the site because AI search tools often scrape Reddit. But Reddit’s moderators are also overwhelmingly anti AI, and the company has worked to delete AI-generated posts and ban accounts that spam. On Monday, Reddit published a blog post saying that “in the age of AI, spam, bot activity, and inauthentic content are top of mind for people who love Reddit (and humans).” In the last few weeks, Reddit launched an ad campaign called “people are best” specifically highlighting that its users are human. A Reddit spokesperson referred us to the blog post when asked for comment.

As we have reported before, no AI detector is 100 percent foolproof, and Pangram certainly has both false positives (human content detected as AI) and false negatives (AI content detected as human). Spero said that the company is constantly working on minimizing both, and that it estimates its false positive rate at roughly one in 10,000. He said he believes the Pangram data is likely a “lower bound” and that the actual problem is likely worse, because people who are willing to install AI detectors on their browsers are likely trying to avoid AI-generated content.

“I think the data generalizes out [to non Pangram users], but that it’s a lower bound on AI content because someone with the Pangram extension probably cares more about seeing AI content than the average person and would be more likely to block or mute AI posters,” he said.

A LinkedIn spokesperson told 404 Media in a statement that “Professionals come to LinkedIn to hear from real people and their unique insights and perspectives. We actively work to reduce low quality, automated or generic content, and while AI can be used to beat the blank page problem, our focus is on surfacing professional conversations that help people advance their careers.” 

 Substack and X did not respond to a request for comment.

Meta has filed a patent for a system that records your voice and surroundings all day, then uses an AI to analyse your mood. The patent’s stated, theoretical goal is for Meta, a company that makes billions of dollars targeting ads at its users based on their data, is to sell users a wearable that tailors workouts for them based on whether they’re happy or sad. 

Patentlyze first noticed the patent which was published on July 2 after Meta filed it back in December of 2025. The filing described an “apparatus” that surveilled a user and their surroundings constantly to craft a better workout. “The audible communications may be associated with contextual factors such as time of day, location, user activity, or digital interaction,” the patent said. “The audible communications may be transcribed, and an emotional-state machine learning model may interpret verbal and nonverbal cues to determine emotional indicators.”

According to the filing, Meta needs to know when a user laughs or sighs, where they are physically, and what objects they’re surrounded by. It would even like to know when you’ve taken your meds. “The AI assistant may listen to a user(s) at predefined times to hear various types of communication, such as sighs, laughter, and/or the tone(s) of a voice(s),” the patent said. “The AI assistant may use these inputs to quantify the user's emotional state or generate other insights about the user [...] in another example, the AI assistant may take multiple inputs in in addition to audio inputs (e.g., of a user's voice) to provide a summary of emotional trends based on various inputs (e.g., a happier emotional state associated with a particular time of day or at a time when medication is taken, etc.).”

The more data it has, the patent explains, the better it could understand a user’s moods. “The system increases the precision and reliability of emotional inference by aligning multimodal sensor inputs on synchronized timelines, which creates a novel data structure that supports richer emotional analysis,” it said. “These combined features deliver a technical improvement in automated audio interpretation, enabling continuous emotional monitoring on everyday devices.”

The emotional-analyzing AI would need far more than just a user’s words to determine moods over time. A longer description of the hypothetical training data for the AI included “attributes of thousands of objects” such as a user’s books, personal messages, and newspapers. “In some examples, audible communications may include speech (e.g., voice data), sighs, laughter, or other nonverbal sounds associated with an expression(s), an emotion(s), or ideas. In some examples, the audible communications may include the tone(s) of a voice of a user while making the communication(s),” it said.

All this data, Meta says, would be in service of tailoring better workouts. Humans, the patent explained, are simply not as good as a machine for this. “Personal trainers cannot provide the level of precision in guidance, such as correcting a pose and/or body movement,” it said. “These challenges create a need for a practical approach that uses a single device to observe movement, recommend routines, and provide corrective guidance.”

AI lives and dies by its training data. Many of the leading LLMs have already scanned the entire internet and are still hungry for more. Meta’s patented system would give it unprecedented access to the movements, moods, and interactions of its users. Giving the user workout suggestions in return seems a paltry compensation.

A wearable device that records every sound you make and transcribes it for an LLM while monitoring your exact location is a privacy nightmare. It’s also a fear that underpins many people’s concerns about big tech. 

A wearable that records your every word and divines your emotions would also, necessarily, record your interactions with other people. Meta has pioneered non-consensual public recording with its smartglasses so it’s not shocking to see it file a patent that suggests it’ll move further into that space. 

The last time Meta explicitly pursued user’s emotional data, it horrified people. In 2012, the company then called Facebook conducted a study into “emotional contagion” using Facebook’s newsfeed. Meta altered the feeds of 700,000 users to see if it could make them happy or sad just by tweaking what they saw online. Meta found that it could, in fact, alter people’s moods if it wanted. It did this without informing users they’d been part of an experiment.

Now it’s patented a device that will record your laughter and play it back to you. All in service of crafting the perfect workout routine. “An implementation may show that the user laughs more often on certain days, shows improved mood after life events, or expresses more positive emotion during morning routines. The device may also provide citations to specific audio moments that support the emotional interpretation,” it said.

Meta wants to tell you how you feel and it’ll use your own voice to do it.

Meta did not immediately respond to 404 Media’s request for comment.