You subscribe to a newsletter because you love the writer’s voice. Then you find out a fucking bot wrote most of it, and nobody told you.
Substack’s CEO gave that gut-punch a name this month: claudefishing. And the platform just released a tool to help readers catch it.
What Is “Claudefishing”
Chris Best, Substack’s CEO, coined the term in a July 2026 post called “Against Claudefishing.” He defined it as a mismatch between what a reader expects and what they actually get, especially when they invest real attention in something with no human thought behind it.
Best made his own position clear in the same post.
“Claude has a lot to offer. But when I want Claude’s opinion, I’ll ask Claude.”
The name riffs on catfishing, the old scam where someone fakes an identity to lure people in. Claudefishing does the same thing with words instead of photos and DMs. A post reads like a person poured themselves into it but in reality an AI actually wrote most of it, and no one disclosed that.
Best’s point isn’t that AI-assisted writing is bad. It’s that readers deserve to know what they’re getting, and right now, most can’t tell AI from human content.
Why Substack Built an AI Detector With Pangram
On July 21, 2026, Substack launched a detection feature built with Pangram, an AI-detection company. It scans posts, notes, comments, and replies and estimates how much of the text a human wrote versus how much came from AI.
A few things to know about how it’s scoped:
The scan only works on content published on or after July 21, 2026, according to Substack’s help center.
It can’t retroactively check your archive or someone else’s.
Substack’s own documentation confirms shorter notes show a “not enough text” message rather than a score; multiple outlets covering the launch report that the practical cutoff is around 100 words, though Substack hasn’t published that exact figure itself.
Results are visible only to the person who requests the scan.
Nobody gets a public AI score badge slapped on their profile. It launched on web and iOS, with Android support still on the way.
Best has been direct that this isn’t a purity test: “It is not perfect, but independent research suggests that it detects AI-generated text with a high degree of accuracy,” he said. That’s Substack and Pangram’s own claim, worth noting as a claim, not an independently audited result.
How to Scan Any Post or Note for AI Text
Here’s exactly how to run a scan yourself:
On the web:
Open a post from your Subscriptions in the Substack reader.
Click the three dots in the top right corner.
Select “Scan for AI text.”
On a Note:
Click the three dots on the Note.
Select “Scan for AI text.”
On iOS:
Tap the post or Note you want to check.
Tap the three dots in the upper right corner.
Tap “Scan for AI text.”
You can also scan individual comments and replies inside a Notes thread the same way.
A few formats are off limits for now. Video and audio posts, standalone Substack sites on custom domains, and emails can’t be scanned.
How Writers Can Disclose AI Use or Turn Off Scanning
If you publish on Substack and use AI anywhere in your process, you have three tools worth setting up today.
Add a “How I make this” statement. Go to your publication’s Settings, then Details. Click “Edit” next to your statement, describe your actual process, and save it. Readers see this when they scan your work, so it’s your chance to explain your process in your own words instead of letting a percentage speak for you.
Scan your own drafts before you publish. Run the same “Scan for AI text” check on a draft to see how it reads before your subscribers do.
Disable detection if you want to. On the Publish page of a draft, click “Disable AI detection.” On a Note, open the three-dot menu after a scan appears and select “Disable detection.” Readers will see an “AI detection unavailable” message instead of a score.
If a scan result looks wrong, you can push back. Select “Report detection error” on the report and submit feedback, which Substack says goes toward improving the tool.
The Real Limits of Pangram’s Scores, According to Critics
Here’s the part worth taking seriously before you treat any AI score as gospel, in either direction.
Matteo Wong, writing in The Atlantic, reported that Pangram’s tool “does make mistakes, perhaps to a greater extent than is currently understood,” and warned that AI accusations “could very quickly spiral into a witch hunt.”
Developer Perry Metzger has argued that any AI detector can be circumvented through fine-tuning, which makes detection an inherently temporary fix rather than a permanent solution.
Researcher Mor Naaman has gone further, framing detection as an unwinnable race: once writers know they’re being scanned, some will simply learn to write around the scanner.
Independent Substack writer Sam Illingworth, who publishes “Slow AI,” ran his own test. One of his posts scored as fully AI-generated, then scored as fully human after he ran it through a free “humanizer” tool. His conclusion: the tool measures how text is phrased, not who actually wrote it or how much care went into it.
None of that means the tool is useless. It means a percentage on a screen is a signal, not a verdict. Substack itself has acknowledged the tool isn’t perfect, and no detector, from any company, catches everything or nothing.
What This Means If You Use Claude to Write
If you use Claude, or any AI tool, anywhere in your writing process, claudefishing isn’t a threat to you. It’s a nudge to be upfront about something you were probably already comfortable with.
Write your “How I make this” statement now, before a reader scans your work and wonders “was this written by AI?” and be truthful. Say plainly what you draft yourself, what Claude or ChatGPT helps you edit or research, and what stays entirely yours. That single paragraph builds more trust than any score ever will.
Use the scan on your own drafts as a reality check, not a grade. If a heavily AI-assisted draft comes back scored as mostly AI, that’s accurate, not a bug. The fix isn’t hiding it better. It’s deciding how much of the final piece should carry your own judgment, examples, and voice before you hit publish.
And don’t let a false flag rattle you either. If Substack’s scan misreads your fully human writing as AI-generated, use the “Report detection error” option and keep publishing. The tool is new, imperfect, and still learning. Your voice isn’t. And remember you are what makes your work younique.