Why Medium’s AI Content Policy is Already Obsolete

Medium’s policy requiring disclosure of AI-generated content and banning it from the Partner Program is fundamentally unenforceable. Detection tools have failure rates as high as 74%, the company’s own CEO admits no AI detector works at scale, and the binary distinction between “human” and “AI” writing has collapsed. What was once a reasonable moderation approach has become theater, a policy that punishes honest disclosure while the real AI flood flows through undetected.

The evidence is damning: OpenAI shut down its own AI detector after achieving only 26% accuracy [1]. Stanford researchers found detection tools wrongly flag 61% of non-native English speakers as using AI [2]. And studies show humans themselves can identify AI text only 53% of the time, barely better than flipping a coin [3]. Meanwhile, independent analysis suggests nearly half of Medium’s content is already AI-generated [4], flowing past detection systems that Medium’s CEO admits don’t work [5].

The Detection Tools Don’t Work, And Mathematically, They Never Will

The foundational assumption behind Medium’s AI policy, that platforms can reliably identify AI-generated content, is provably false. Not just technically difficult: mathematically impossible.

OpenAI, the creator of ChatGPT, couldn’t build a working detector. On July 20, 2023, they shut down their AI Classifier after just six months. The official reason: “low rate of accuracy.” The actual numbers were devastating: the tool correctly identified AI-generated text only 26% of the time while falsely accusing human writing 9% of the time [1]. The company that built the most popular AI writing tool in history could not reliably detect its own output.

University of Maryland researchers proved this isn’t just an engineering challenge, it’s a fundamental limit. Professor Soheil Feizi’s team demonstrated mathematically that as AI models improve at mimicking human writing, detection accuracy approaches random chance [6]. “It’s very unlikely that we’ll be able to develop detectors that will reliably identify AI-generated content,” Feizi concluded. Their paper showed that simple paraphrasing tools reduce even the best detectors to coin-flip accuracy.

The tools that do exist are dangerously biased. Stanford’s Human-Centered AI Institute tested seven popular detectors on essays by non-native English speakers and found that 61% were falsely classified as AI-generated [2]. In 19% of cases, every single detector unanimously agreed the human-written essay was AI. Professor James Zou called the implications “serious questions about the objectivity of AI detectors” that could cause “foreign-born students and workers [to] be unfairly accused.”

Detection ToolKey FindingSourceOpenAI Classifier26% accuracy, discontinuedOpenAI [1]Turnitin50% false positive rate (Washington Post test)Washington Post [7]AI detectors on ESL writers61% false accusation rateStanford HAI [2]ZeroGPTPlateaued at 16.9% false positive rateRAID Benchmark Study [8]

Even Turnitin, whose AI detection is used by thousands of universities, admits their tool “should not be used as the sole basis for adverse actions.” The company deliberately accepts a 15% false negative rate, meaning one in seven AI texts passes through undetected, because they’re more afraid of false accusations than missed detections [9]. Vanderbilt University calculated that even at Turnitin’s claimed 1% false positive rate, 750 student papers would be wrongly flagged at their school alone. They disabled the tool entirely [10].

Medium’s CEO Admits Detection Doesn’t Work at Scale

Here’s the awkward truth Medium doesn’t advertise: the company’s own leadership knows AI detection is broken.

In a February 2024 interview with Semafor, CEO Tony Stubblebine was blunt: “Every couple of weeks, we test the latest AI detection tools. None of them are good enough for platform use.” [5] He described relying on human curators to spot AI content manually, an approach that works for editorial recommendations but cannot possibly police a platform processing tens of millions of posts.

Stubblebine’s strategy is essentially hope: “The whole point of having humans look at it is to find the stuff worth recommending… as soon as the AI-generated content started showing up, it was the humans that spotted it immediately.” But this means AI content isn’t being removed, it’s just being filtered from Medium’s recommendation engine. As Stubblebine admitted: “We have a lot of it on the platform.”

How much is “a lot”? Independent analysis suggests the answer is staggering. Pangram Labs analyzed 274,466 Medium posts over six weeks and estimated 47% or more were likely AI-generated [4]. CEO Max Spero noted this was “a couple orders of magnitude more than what I see on the rest of the internet.” Originality AI found similar results: roughly 40% of Medium posts in 2024 were likely AI-generated, compared to just 3.4% in 2018 [11].

Stubblebine disputed the methodology but not the underlying problem. “I am disputing the importance of the results,” he told Wired [12], an acknowledgment that Medium cannot prove the critics wrong.

The platform’s policy states that AI-generated writing “is NOT allowed behind the paywall” and violations may result in Partner Program enrollment being revoked [13]. But if nearly half the content is AI-generated and detection tools don’t work, who exactly is being punished? Only writers honest enough to disclose.

Economics Have Made Human-Only Policies Obsolete

Medium’s business model was built for a world where content was expensive and slow to produce. That world ended in November 2022.

AI-generated content is now 4.7 times cheaper than human-written content, $131 average per blog post versus $611, according to an Ahrefs survey of 879 marketers [14]. Production time has collapsed by more than 90%: what takes a human writer 4–6 hours takes AI under 60 seconds. GPT-4 can generate 200–500 words per second under standard conditions.

The flood is already here. According to Imperva’s 2025 Bad Bot Report, 51% of all web traffic in 2024 was generated by bots, the first time in a decade it exceeded human activity [15]. Gartner reports that 60% of newly indexed web pages as of Q4 2025 contain primarily synthetic content [16]. Reddit co-founder Alexis Ohanian, speaking at a Wall Street Journal panel, declared the “dead internet theory,” once dismissed as conspiracy, “a very real thing” [17].

Even Sam Altman acknowledged the shift. “I never took the dead internet theory that seriously,” the OpenAI CEO tweeted in September 2025, “but it seems like there are really a lot of LLM-run Twitter accounts now” [18].

The economic incentives are overwhelming. CybelAngel investigator Jean-Marc Manach uncovered 4,000+ AI-generated fake news websites operating primarily in French, with some operators reportedly becoming millionaires through ad revenue alone [19]. At those margins, no policy requiring disclosure can survive contact with reality. Writers who disclose are penalized; writers who don’t are rewarded. The honest are punished while the bad actors win.

The “Human vs. AI” Binary Has Collapsed

Medium’s policy draws a line between “AI-generated” and “AI-assisted” content. AI-generated writing must be disclosed and is banned from the paywall. AI-assisted content, defined as using AI for “outlining tools” or “spell, grammar, or fact-checkers” requires no disclosure [13]. This distinction has become meaningless.

Where exactly does “assistance” end and “generation” begin? Is a writer who uses AI to brainstorm 20 headline options, then picks one and writes the article themselves, “AI-assisted”? What about someone who has AI write a first draft, then rewrites 70% of it? Or someone who writes everything themselves but uses Grammarly’s AI-powered suggestions to restructure sentences?

73% of marketing professionals already use AI tools, but only 12% rely on them completely [20]. 72% of PR professionals use AI to write first drafts; 70% use it to edit drafts they wrote themselves [21]. The industry has moved to hybrid workflows where AI and human contributions are genuinely inseparable, not because writers are hiding AI use, but because that’s simply how modern writing works.

Research confirms the binary makes no sense. A study of 18 creative writers found they maintain “multiple, simultaneous relationships with AI that resist clear classification” [22]. Writers described AI variously as tool, assistant, junior writer, ghostwriter, collaborator, muse, and editor, often within the same writing session. One writer captured the absurdity: “I don’t want to say that the AI is alive… but I also don’t want to say that it’s this lifeless tool… so it’s a strange middle ground.”

Ted Chiang, writing in The New Yorker, identified the key distinction: “When you are writing fiction, you are, consciously or unconsciously, making a choice about almost every word… a ten-thousand-word short story requires something on the order of ten thousand choices” [23]. But modern hybrid workflows distribute those choices across human and AI in ways that defy categorization. The question isn’t “was AI used?” but “who made the meaningful creative choices?”and that’s not something any policy can measure.

Humans Cannot Tell the Difference Anyway

Even if platforms could perfectly detect AI at the technical level, there’s a deeper problem: humans themselves cannot reliably distinguish AI writing from human writing.

Penn State’s PIKE Lab tested human ability to identify AI text and found accuracy of just 53%, where random guessing achieves 50% [3]. “By and large, people cannot really distinguish AI-generated text well,” concluded Professor Dongwon Lee. Training didn’t help. Team-based detection didn’t help. Humans are essentially guessing.

A UC San Diego Turing Test study found that GPT-4 was judged human 54% of the time, compared to actual humans being judged human 67% of the time [24]. When GPT-4o was given persona prompts, it achieved a 77% pass rate, higher than the 71% rate for actual humans. The machines are more convincingly human than the humans.

German researchers tested 63 university lecturers on academic texts and found humans achieved only 57% recognition of AI content [25]. For professional-level AI writing, less than 20% was correctly identified. The better the AI, the less detectable it becomes.

If professional academics cannot identify AI writing with any reliability, if the humans reading Medium articles cannot tell the difference, and if detection tools produce false accusations at rates up to 61%, what exactly is Medium’s policy protecting?

Real People Are Being Harmed by Broken Detection

The consequences of unreliable AI detection aren’t theoretical. Real writers have lost livelihoods. Real students have faced academic tribunals. Real reputations have been destroyed.

Michael Berben, a freelance writer with three years of experience and approximately 200 published articles, had his career ended by AI detection [26]. A client ran his work through a detector that flagged articles as 65–95% AI-generated, including articles written before ChatGPT existed. Berben provided his complete Google Docs edit history proving human authorship. The client terminated his contract immediately based solely on the AI scores. “These AI detectors are literally destroying people’s livelihoods,” he said.

William Quarterman, a UC Davis senior, was falsely accused of using ChatGPT after his professor ran his midterm essay through GPTZero [27]. He received zero on the assignment and was referred to the university’s honor court for academic dishonesty. He suffered panic attacks while defending himself. To prove his innocence, he demonstrated that the same AI detectors flagged Martin Luther King’s “I Have a Dream” speech and the Book of Genesis as AI-generated. He was eventually exonerated but has since been contacted by “several students who have had similar experiences.”

Johns Hopkins professor Taylor Hahn noticed a pattern after Turnitin flagged a student’s paper as 90%+ AI-generated: “International students were disproportionately flagged” [28]. The Stanford research confirmed this systematically, detection tools mistake the clear, simple writing common among non-native English speakers for AI patterns [2].

These aren’t edge cases. They’re the predictable outcome of deploying fundamentally unreliable technology for high-stakes decisions.

Other Platforms Are Struggling Too, But Medium’s Approach Is Uniquely Vulnerable

Medium isn’t alone in facing the AI content crisis, but its business model makes it uniquely exposed.

Substack has no AI-specific policy at all, writers build their own subscriber bases, so there’s no shared revenue pool that AI content could dilute. WordPress encourages transparency but imposes no prohibition, their stance is that “quality matters more than the method” [29]. YouTube requires disclosure only for “realistic altered or synthetic content” in video, not for AI-assisted scripts [30]. These platforms have accepted that policing AI at scale is impossible and designed around that reality.

Medium’s Partner Program, by contrast, pools reader subscription revenue and distributes it based on engagement metrics: reading time, claps, highlights, responses [31]. This creates perverse incentives. If AI can generate engaging content more cheaply than humans, the economic pressure is enormous. If detection doesn’t work, honest disclosure becomes self-punishment.

The platform is “too big to police,” as Plagiarism Today observed “like Amazon, YouTube, Facebook and other large online platforms, [Medium has] to rely heavily on automated tools and algorithms, no matter how imperfect those systems are” [32]. But those systems don’t work, and Medium’s leadership knows it.

The company reached profitability in August 2024 after 13 years of losses, with over 1 million paid subscribers [33]. That’s a genuine achievement. But sustainability depends on the Partner Program maintaining value for human writers, and that requires either functional AI detection (impossible) or a different model entirely.

What Comes Next for Content Platforms

Medium’s AI policy represents a category error: treating AI content as a detection problem rather than an economics problem. The real question isn’t “how do we identify AI content?” but “how do we create value in a world where content is infinitely abundant?”

Researchers at Nature published evidence that AI models trained on AI-generated data eventually degrade into “model collapse” suggesting the open web could become, as one analyst put it, “a graveyard of synthetic noise” [34]. Dr. Elena Voss, a computational linguist, offered a stark metaphor: “We have polluted the information ecosystem with plastic. Just as microplastics are now in our bloodstream, AI ‘slop’ is now in our knowledge stream” [16].

Some platforms are experimenting with authenticity verification, proving human identity rather than detecting AI. Others are exploring economic models where human attention is the scarce resource to be compensated. Cloudflare, which powers 20% of the web, now blocks AI bots by default unless site owners explicitly allow them [35].

Professor Ambuj Tewari of the University of Michigan captured the fundamental dynamic: “AI text detection is part of an escalating arms race. Detection tools must be publicly available to be useful, but that same transparency enables evasion. As AI text generators grow more capable and evasion techniques more sophisticated, detectors are unlikely to gain a lasting upper hand” [6].

The arms race cannot be won. Disclosure policies reward deception. Detection tools harm innocent writers while missing the bad actors they target. The binary distinction between “AI” and “human” writing has dissolved into a spectrum that no policy can meaningfully categorize.

A Policy from a Pre-AI World

Medium’s AI content policy isn’t evil or stupid, it’s obsolete. It was designed for a moment when AI writing was detectably different from human writing, when the economics still favored human production, and when “AI-assisted” and “AI-generated” were meaningfully distinct categories. That moment passed in late 2022 and isn’t coming back.

The evidence is overwhelming: detection tools produce false accusations against non-native English speakers at rates exceeding 60% [2]. The best AI detector ever built, by the company that created ChatGPT, achieved 26% accuracy before being shut down [1]. Humans themselves cannot distinguish AI writing from human writing at rates better than random chance [3]. Nearly half of Medium’s content may already be AI-generated despite the policy [4]. And the company’s own CEO admits that “none of the AI detection tools are good enough for platform use” [5].

What Medium needs isn’t better detection, it’s a different model. One that doesn’t pretend binary distinctions exist where they don’t, that doesn’t rely on technology proven to be biased and unreliable, and that doesn’t punish honest disclosure while rewarding deception. The platform that figures out how to create value for human writers in the age of infinite AI content will thrive. The platforms still trying to detect their way out of this problem will fail, not because they lack will, but because they’re solving the wrong equation.

References

[1] OpenAI. “New AI classifier for indicating AI-written text.” OpenAI, January 2023.
https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/

[2] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. “AI-Detectors Biased Against Non-Native English Writers.” Stanford Human-Centered AI Institute, 2023.
https://hai.stanford.edu/news/ai-detectors-biased-against-non-native-english-writers

[3] Penn State University. “Q&A: The increasing difficulty of detecting AI- versus human-generated text.” Penn State News, 2024.
https://www.psu.edu/news/information-sciences-and-technology/story/qa-increasing-difficulty-detecting-ai-versus-human

[4] “Medium is being flooded with AI-generated content.” CO/AI, October 2024.
https://getcoai.com/news/medium-is-being-flooded-with-ai-generated-content/

[5] Stubblebine, T. Interview with Semafor, February 2024. Referenced in Fast Company.
https://www.fastcompany.com/91407562/medium-ceo-tony-stubblebine-interview

[6] University of Maryland. “Is AI-Generated Content Actually Detectable?” College of Computer, Mathematical, and Natural Sciences, 2023.
https://cmns.umd.edu/news-events/news/ai-generated-content-actually-detectable

[7] Weber, L., et al. “Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text.” International Journal for Educational Integrity, Springer, 2023.
https://link.springer.com/article/10.1007/s40979-023-00140-5

[8] RAID Benchmark Study on AI Detection Tools. Referenced in University of San Diego Legal Research Center guide.
https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367

[9] Turnitin AI Detection documentation and false positive rate disclosures. Referenced in Vanderbilt guidance.

[10] Vanderbilt University. “Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector.” Brightspace Support, August 2023.
https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/

[11] Originality AI. Analysis of AI content on Medium, 2024. Referenced in Slashdot.
https://slashdot.org/story/24/10/29/1733220/ai-slop-is-flooding-medium

[12] Stubblebine, T. Interview with Wired, October 2024. Referenced in CO/AI analysis.

[13] Medium Help Center. “Artificial Intelligence (AI) content policy.”
https://help.medium.com/hc/en-us/articles/22576852947223-Artificial-Intelligence-AI-content-policy

[14] Ahrefs. “AI Content Is 4.7x Cheaper Than Human Content [+ New Research Report].”
https://ahrefs.com/blog/ai-content-is-5x-cheaper-than-human-content/

[15] Imperva. “2025 Bad Bot Report.” Referenced in dead internet theory coverage.

[16] “Dead Internet Arrives: Gartner Reports 60% of Web is AI.” Editorialge, 2025.
https://editorialge.com/dead-internet-theory-2026/

[17] Fortune. “Reddit cofounder Alexis Ohanian says ‘so much of the internet is dead.’” October 2025.
https://fortune.com/2025/10/15/reddit-co-founder-alexis-ohanian-dead-internet-theory-ai-bots-linkedin-slop/

[18] Altman, S. Twitter/X post, September 2025. Referenced in BitcoinEthereumNews.
https://bitcoinethereumnews.com/finance/the-dead-internet-theory-noted-by-altman-and-ohanian-explained/

[19] CybelAngel. “Exactly how is AI Weaponizing SEO for profit?”
https://cybelangel.com/blog/ai-generated-fake-news-sites/

[20] Amra & Elma. “TOP 20 AI VS HUMAN COPYWRITER PERFORMANCE STATISTICS 2025.”
https://www.amraandelma.com/ai-vs-human-copywriter-performance-statistics/

[21] Siege Media. “50 AI Writing Statistics To Know in 2025.”
https://www.siegemedia.com/strategy/ai-writing-statistics

[22] Kim, J., et al. “From Pen to Prompt: How Creative Writers Integrate AI into their Writing Practice.” arXiv, November 2024.
https://arxiv.org/html/2411.03137v2

[23] Chiang, T. “Why A.I. Isn’t Going to Make Art.” The New Yorker, 2024.

[24] Jones, C. & Rathi, A. “People cannot distinguish GPT-4 from a human in a Turing test.” arXiv, May 2024.
https://arxiv.org/abs/2405.08007

[25] “Do humans identify AI-generated text better than machines? Evidence based on excerpts from German theses.” ScienceDirect, 2025.
https://www.sciencedirect.com/science/article/pii/S1477388025000131

[26] Authory. “How AI detectors can destroy innocent writers’ livelihoods.”
https://authory.com/blog/how-ai-detectors-are-destroying-livelihoods

[27] Tech & Learning. “He Was Falsely Accused of Using AI. Here’s What He Wishes His Professor Did Instead.”
https://www.techlearning.com/news/he-was-falsely-accused-of-using-ai-heres-what-he-wishes-his-professor-did-instead

[28] The Markup. “AI Detection Tools Falsely Accuse International Students of Cheating.” August 2023.
https://themarkup.org/machine-learning/2023/08/14/ai-detection-tools-falsely-accuse-international-students-of-cheating

[29] WPBeginner. “Is AI Content Bad for WordPress SEO? (Expert Insights & Tips).”
https://www.wpbeginner.com/opinion/is-ai-content-bad-for-wordpress-seo/

[30] Influencer Marketing Hub. “AI Disclosure Rules by Platform: YouTube, Instagram/Facebook, and TikTok Labeling Guide.”
https://influencermarketinghub.com/ai-disclosure-rules/

[31] Medium Help Center. “Calculating earnings in the Partner Program.”
https://help.medium.com/hc/en-us/articles/360036691193-Calculating-earnings-in-the-Partner-Program

[32] Plagiarism Today. “Medium Sets New Policies on AI-Generated Writing.” April 2024.
https://www.plagiarismtoday.com/2024/04/11/medium-sets-new-policies-on-ai-generated-writing/

[33] Fast Company. “Medium CEO Tony Stubblebine on launching a new app and turning around a 13-year losing streak.”
https://www.fastcompany.com/91407562/medium-ceo-tony-stubblebine-interview

[34] Nature research on model collapse. Referenced in dead internet theory analysis.

[35] Cloudflare AI bot blocking policies. Referenced in Windows Central.
https://www.windowscentral.com/software-apps/reddit-co-founder-says-much-of-the-internet-is-dead