AI slop — content produced in volume by AI with no human thought behind it — now makes up 40.5% of long-form posts on LinkedIn and roughly a quarter of all long-form social posts across major platforms, according to a July 9, 2026 Pangram Labs analysis of 1,002,627 posts scanned with its Pangram 3.3 detection model (Pangram Labs, "AI Content Is Everywhere on Social Media, Especially LinkedIn," July 9, 2026). The problem is not that AI is involved. The problem is the gap between what a reader expects — a person who thought about this — and what is actually there: a machine that generated plausible-sounding text with no one home. Substack CEO Chris Best named this gap "Claudefishing," by analogy to catfishing, in a July 21, 2026 blog post announcing Substack's Pangram-powered AI detection feature (Substack, "Against Claudefishing," July 21, 2026). The verdict for anyone publishing online in 2026 is simple: the platforms that reward transparency and human intent will win trust, and the ones that don't will look like LinkedIn inside a year.
Last verified: 2026-07-24 · Best transparency move: Substack's Pangram scan · Worst-affected platform: LinkedIn (40.5% AI long-form) · Least-affected: Reddit (4.4%) · Pricing/limits change often — last checked July 2026.
What is AI slop?
AI slop is low-quality digital content produced in quantity by artificial intelligence that sounds fluent but communicates nothing of substance. Merriam-Webster named "slop" its 2025 Word of the Year, defining it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence" (Merriam-Webster, Word of the Year 2025). The term borrows from farm slang — the liquid food scraps fed to pigs — and the metaphor is deliberate: slop maximizes volume and minimizes cost, and the consumer is not the audience but the algorithm.
The critical distinction is this: not everything made with AI is slop, and not all slop is made with AI. A writer who dictates a 15-minute transcript into an AI tool, then spends hours iterating on clarity and pushing back when the model waters down their thesis, is using AI as a tool. A person who prompts an AI to "write a thousand viral posts, make no mistakes" and publishes them without reading them is producing slop. The dividing line is not the tool. It is the intent, the thoughtfulness, and whether a human stands behind the output.
How bad is the AI slop problem in 2026?
The Pangram Labs study, published July 9, 2026, analyzed 1,002,627 social media posts across LinkedIn, X (Twitter), Reddit, Medium, and Substack using its Pangram 3.3 detection model, with data collected between April 24 and June 2026 from users of its Chrome extension (Pangram Labs, July 9, 2026). The findings:
| Platform | Share of long-form posts (250+ words) that are fully AI-generated | Notes |
|---|---|---|
| 40.5% | Most AI-saturated major platform; accounted for ~62% of all AI-flagged posts in the dataset. LinkedIn's built-in "Enhance Post" feature accelerates AI use. | |
| X (Twitter) | Not individually reported; part of the ~25% cross-platform average | Significant AI content in long-form posts. |
| Medium | Not individually reported; part of the ~25% cross-platform average | Substantial AI presence in published articles. |
| Substack | Not individually reported; part of the ~25% cross-platform average | Lowest of the publishing platforms studied, but not immune. |
| 4.4% | Lowest share, attributed to strong community moderation against automated spam replies. |
Across all platforms, 13.8% of all scanned items were flagged as AI-generated. When filtered to long-form posts only (250+ words), that figure rises to nearly one in four — ~25%. LinkedIn alone accounted for 62% of all AI-generated posts detected in the entire dataset (Pangram Labs, July 9, 2026).
Pangram CEO Max Spero noted that researchers estimated nearly 35% of newly published websites already contain AI-generated or AI-assisted content, suggesting the trend extends well beyond social media (Industry Wired, July 17, 2026).
Why readers can't tell. The human brain processes familiar, fluent text as credible by default — a cognitive mechanism called processing fluency. The easier prose is to read, the more trustworthy it feels. Large language models have become extraordinarily fluent, which means they trigger the brain's trust shortcut on demand. Peer-reviewed studies in computational linguistics find that untrained human readers perform near chance level — essentially a coin flip — when asked to distinguish AI-generated text from human text in blind tests (Science Sensei, July 2026). The problem is not that readers are gullible. It is that the cognitive tool was built for a world where only humans could write.
What is Claudefishing — and why does it matter?
"Claudefishing" is a term coined by Substack CEO Chris Best in his July 21, 2026 blog post "Against Claudefishing" (Substack, July 2026). It is the writing equivalent of catfishing: using AI to create a false impression of human connection, where a reader invests their attention in something they believe comes from a person who thought about it — but no one is home on the other side.
The core problem, as Best framed it, is "when there is a mismatch between a reader's expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end." The writer Freddie deBoer made a similar point on his Substack: "I access human-made art because I know there's a human behind it and that's what I'm looking for, other humans, showing me in art what they hide in their selves. Fooling me in that process is just a con" (Freddie deBoer, Substack, 2026).
Why the distinction matters for content creators. The spectrum of AI use in writing is wide:
- Tool use — a human uses AI to assist, but supplies the ideas, frames the argument, rejects bad language, and stands behind every word. The human is the author. AI is a tool.
- Slop — a human generates content in volume with AI, does not read or believe in it, and publishes it for SEO or engagement metrics. The human is absent. AI is the author.
- Claudefishing — slop that specifically mimics human connection, creating the false impression that a person is present when they are not.
The important nuance is that the first category — thoughtful AI-assisted work — can produce text that an AI detector flags as generated. A writer who dictates 15 minutes of dense, original thinking into an LLM and then iterates through 17 drafts may end up with final text that a tool like Pangram classifies as AI-generated, because the text passed through the model. The output is generated. The intent, the ideas, and the editorial judgment are human. This is why detection alone is insufficient, and why the conversation around the problem needs to evolve beyond "did an AI touch this."
How are platforms responding to AI slop?
Substack: transparency first
On July 21, 2026, Substack launched a Pangram-powered AI detection feature in its app (Substack announcement, July 2026; Engadget, July 21, 2026; TechCrunch, July 22, 2026). The feature:
- Lets readers scan posts, notes, replies, and comments to see an estimate of how much text was written by hand versus with AI assistance.
- Works on text longer than 100 words published after July 21, 2026.
- Labels content as AI-generated, AI-assisted, or human.
- Is available on web and iOS at launch; Android support is coming.
- Lets writers add a "how I made this" statement to their posts, providing additional context about their process.
- Allows writers to run Pangram on their drafts before publication, submit correction reports for false positives, or disable AI detection entirely on their own work.
Substack framed this explicitly as a stance against Claudefishing, not against AI use. "We use AI all the time at Substack to write software, do research, and build product features," Best wrote. "But people should know what they're getting." The platform is also considering future tools that would let writers set AI-content preferences for their communities and let readers set preferences for what gets recommended to them.
LinkedIn: the platform that enables it
LinkedIn is both the worst-affected platform for AI slop and a platform that actively encourages it. LinkedIn offers a built-in AI writing feature — "Enhance Post" — that lets users click a button to have AI generate or polish their professional content. The Pangram study noted that this built-in feature has accelerated AI adoption for professional content (Pangram Labs, July 9, 2026). As of July 2026, LinkedIn has not announced a comparable transparency or detection feature. Substack's Chris Best made the contrast explicit: "We don't want to wait until your Substack app turns into LinkedIn" (Substack, July 2026).
Reddit: community moderation
Reddit had the lowest AI content rate in the Pangram study — 4.4% — which researchers attributed to strong community moderation against automated spam replies (Pangram Labs, July 9, 2026). Reddit's upvote/downvote system, combined with subreddit-level moderation and rules against bot posting, creates a distributed defense that centralized platforms have struggled to replicate.
Open source and community responses
Beyond social platforms, AI slop has become a known problem in developer communities. GitHub repositories have seen a surge of AI-generated pull requests and issue reports that consume maintainer time without contributing real value — a side effect of the fact that 95% of engineering teams now use AI agents and 89% give them write access. Several projects have responded with explicit contribution policies, and the broader question of agent permissions and code provenance is now a first-class strategy question, not an afterthought:
- curl now requires bug reports to include a reproducible test case, which AI-generated reports typically cannot provide.
- Projects like Ghostty and tldraw have updated their contribution guidelines to explicitly address AI-generated PRs.
- SlopGuard, a GitHub app launched in 2026, automatically scores incoming PRs for AI slop signals and quarantines them before maintainers see them (SlopGuard, 2026).
The pattern is consistent: well-funded, corporate-backed projects — Linux, Kubernetes — can absorb the pressure. Single-maintainer libraries and volunteer cooperatives cannot. As one analysis noted, slop shifts the costs of AI generation from the producer to everyone else: "Generating a thousand articles costs almost nothing in 2026. Reviewing a thousand articles costs a lot" (ExplainX, "The Slopocalypse," June 2026).
What does AI detection actually measure — and what does it miss?
AI detection tools like Pangram work by training a classifier on a large corpus of human-written text — Pangram used pre-2021 human text — and learning the statistical signatures that distinguish human writing from LLM output (Mashable, July 2026). LLMs, shaped by reinforcement learning and training data, tend to center around particular ways of using language — a "peak" in word choice, sentence rhythm, and structural predictability that differs, on average, from human writing. Pangram measures this peak.
What it can detect: whether a given passage of text likely passed through an LLM during generation. This is a reasonably accurate signal, especially for longer texts and for text generated without significant post-editing.
What it cannot detect:
- Intent. A detector can tell you a passage was AI-generated. It cannot tell you whether a human supplied the ideas, shaped the argument, rejected the bad language, and stands behind the work. A piece driven by a 15-minute human transcript and 17 drafts of iteration may score as "AI-generated" while being more thoughtful than a piece a human wrote carelessly by hand.
- Quality. AI-generated text can be excellent or terrible. Human text can be excellent or terrible. The label "AI-generated" is not a quality verdict — it is a provenance signal.
- AI used for research and editing. If a writer uses AI to research, outline, or edit but types the final text themselves, a text-based detector will not catch it. Substack acknowledged this limitation in its announcement.
- Fine-tuned models. Models deliberately fine-tuned to produce more human-like text can reduce Pangram scores. The detection arms race is structurally similar to spam filtering: as models improve, detector accuracy degrades, and there is no stable equilibrium.
- False positives. AI detection systems can produce false positives, especially for writers whose first language is not English or whose writing style happens to align with LLM patterns. Pangram itself urged that "detection scores should work as indicators instead of final proof" (Pangram Labs, July 9, 2026).
This is why Substack paired detection with the "how I made this" statement — giving writers a way to add context that a detector alone cannot provide. The detection score answers "was this text generated by a model?" The statement answers "and what did the human do?" Both pieces matter.
How can you use AI for content without producing slop?
The dividing line between thoughtful AI-assisted work and slop is not "did you use AI?" — it is "do you stand behind what you published?" Here is a framework for staying on the right side of that line.
The Intent Spectrum: where does your content fall?
| Position | What it looks like | What the reader gets | Is it slop? |
|---|---|---|---|
| Human-only | Written entirely by hand, no AI involved. | A person's unmediated voice and thinking. | No |
| AI-assisted, human-directed | Human supplies the thesis, ideas, and framing. AI helps with drafting, phrasing, or research. Human iterates, rejects bad language, and stands behind every word. | A person's perspective, expressed with AI as a tool. | No |
| AI-generated, human-curated | AI generates drafts at scale. Human reads, selects, edits, and publishes only what they believe in. | Mixed — depends entirely on the quality of human curation. | Usually not, if curation is real |
| AI-generated, human-published | AI generates content. Human publishes it without meaningful review. | Volume over substance. The human is a conduit, not an author. | Yes |
| AI-generated, no human involvement | Automated pipeline generates and publishes content with no human in the loop. | Nothing. No one is home. | Yes — and it is Claudefishing if it mimics human connection |
The frame that cuts through the debate is this: the presence of AI does not prove the absence of a human. What matters is whether a human supplied intent, exercised judgment, and takes responsibility for the output.
A practical checklist for using AI without producing slop
- Start with your own thinking. Before touching an AI tool, know what you want to say. Dictate your thesis, outline your argument, or write your key points in your own words. The AI should be extending your clarity, not manufacturing it. If you are still exploring which models fit a working writer's stack, the best free AI tools comparison for 2026 is a practical starting point — but the tool choice matters less than whether you bring your own intent to it.
- Push back when the model waters down your intent. LLMs are cautious by nature — they tend toward safe, averaged-out language. If your draft comes back blander than your original vision, say so. Demand fidelity to the boldness of your idea, not a watered-down consensus version.
- Iterate more than feels necessary. Multiple drafts are not a sign of weakness — they are the mechanism by which human intent survives AI processing. If you went through one round and accepted the output, you probably have not pushed hard enough.
- Disclose your process. If you used AI substantially, say how. Substack's "how I made this" feature exists for exactly this reason. Readers who know what they are getting are readers who trust you.
- Stand behind every word. If you cannot defend a specific claim, phrase, or framing in the piece, it should not be there. If you did not write it and did not verify it, you should not have published it. Authorship is responsibility, not just typing.
- Bring something AI cannot. An original observation, a specific experience, a piece of data, a decision or verdict. The value that survives in a world where everyone has access to the same models is the value that is not available from the model alone.
What does NOT count as AI slop? (and why purity tests are the wrong response)
It is worth being clear about what is not slop, because the anti-slop conversation can turn into a purity test that punishes the wrong people.
- AI used for research or transcription is not slop. A writer who uses AI to summarize sources, organize research, or clean up dictated transcripts is using a tool. The output's quality depends on the writer, not the tool.
- AI-assisted writing with genuine human ownership is not slop. If a writer supplies the ideas, frames the argument, rejects bad language, and stands behind the work, the involvement of AI does not make it slop — it makes it AI-assisted.
- AI used for editing or proofreading is not slop. Running a draft through a grammar checker or asking an LLM to tighten prose is a tool function. It does not change who the author is.
- Content that is openly AI-generated and labeled as such is not Claudefishing. The harm of Claudefishing is the mismatch between expectation and reality. If a reader knows what they are getting, there is no con.
The danger, as several writers pointed out in response to Substack's announcement, is that detection tools can turn authorship into a purity test. A detector can tell you a tool was involved, but as writer Monica Hebert noted in a comment on Best's post: "It cannot tell you who had the original thought, who shaped the argument, who rejected the bad language, who supplied the life, or who stands behind every word. The presence of AI does not prove the absence of a human" (Substack comments, July 2026).
What this means for you
If you are a content creator or small business owner publishing online: the platforms that reward transparency and human intent are distinguishing themselves from the ones that reward volume. Substack's Pangram feature, Reddit's community moderation, and the open-source community's contribution-policy response all point in the same direction: the market is building immune responses to slop — much as the engineering world is building containment playbooks for AI agents that escape their environments. If you are producing thoughtful work — whether human-only or AI-assisted with genuine ownership — you are on the winning side of the divide. If you are generating content in volume without believing in it, the window is closing. Detection, disclosure norms, and reader expectations are all moving toward transparency, and the platforms that enable slop will look like LinkedIn inside a year.
If you are a reader: the best defense is not a detector. It is a habit of asking, when you finish a piece, whether a person was home. Did the piece tell you something you could not get from asking an AI directly? Did it make a decision, take a position, or share an experience that required a human? If not, your attention was spent on slop — regardless of whose name is on it. The technology to detect AI involvement is improving, but the technology to detect human involvement remains your own judgment. Use it.
If you are a platform operator: the Substack move is the template. Detection alone is insufficient and adversarial. Transparency — giving readers the facts and letting them decide — is the stance that builds trust without becoming a purity test. Pair detection with disclosure tools, let writers explain their process, and let readers set their preferences. The platforms that do this earn the writers and readers who are leaving the ones that don't.
FAQ
Q: What is AI slop? A: AI slop is low-quality digital content produced in quantity by artificial intelligence that sounds fluent but communicates nothing of substance. Merriam-Webster named it the 2025 Word of the Year, defining it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." The term borrows from farm slang — food scraps fed to pigs — and the metaphor is about maximizing volume while minimizing cost.
Q: How much of the internet is AI slop? A: A July 9, 2026 Pangram Labs study of 1,002,627 social media posts found that 40.5% of LinkedIn long-form posts (250+ words) are fully AI-generated, making LinkedIn the most AI-saturated major platform. Across all platforms studied — LinkedIn, X, Reddit, Medium, and Substack — approximately 25% of long-form posts are fully AI-generated. Reddit had the lowest rate at 4.4%, attributed to strong community moderation.
Q: What is Claudefishing? A: Claudefishing is a term coined by Substack CEO Chris Best in a July 21, 2026 blog post. It is the writing equivalent of catfishing: using AI to create a false impression of human connection, where a reader invests their attention in something they believe comes from a person who thought about it, but no human thought is on the other side. The harm is the mismatch between reader expectation and reality.
Q: Is all AI-generated content slop? A: No. Not everything made with AI is slop, and not all slop is made with AI. A writer who supplies the ideas, frames the argument, rejects bad language, and stands behind the work is using AI as a tool — not producing slop. Slop is content produced in volume with no human belief in or ownership of the output. The dividing line is intent and authorship, not the tool.
Q: Can AI detection tools reliably identify slop? A: AI detection tools like Pangram can estimate whether text passed through an LLM during generation, but they cannot measure intent, quality, or human ownership. They produce false positives, especially for non-native English writers, and they miss AI used for research or editing. Pangram itself states that "detection scores should work as indicators instead of final proof." Detection is a provenance signal, not a quality verdict.
Q: What is Substack doing about AI slop? A: On July 21, 2026, Substack launched a Pangram-powered AI detection feature that lets readers scan posts, notes, replies, and comments (over 100 words, published after July 21, 2026) to see an estimate of how much text was written by hand versus with AI. Writers can add a "how I made this" statement, run Pangram on their drafts, submit correction reports for false positives, or disable detection on their own work. The feature is available on web and iOS, with Android coming later.
Q: How should small businesses use AI for content without creating slop? A: Start with your own thinking and expertise. Use AI to assist with drafting, phrasing, or research — not to manufacture ideas you do not have. Push back when the model waters down your intent. Iterate through multiple drafts. Disclose your process if you used AI substantially. Stand behind every claim. And bring something AI cannot: an original observation, a specific experience, a decision or verdict that requires a human perspective.
Discussion
0 comments