The line between human writing and artificial generation is blurring fast. Substack just made a move to help you tell the difference. The platform announced a new feature that lets you scan for AI-generated content directly within the app and on the web. It is powered by a detection engine from a company called Pangram. This tool now sits inside Substack posts, notes, replies, even comments.
The logic behind the update is simple enough. Not everything AI writes is garbage. Not all low-effort writing comes from a machine. But readers deserve to know what they are investing their time into. If you click on an article expecting a human voice, it feels wrong when you realize it’s just algorithmic noise. That mismatch is the problem Substack wants to fix.
How to use the AI detection feature
You do not need a new app to get this working. The tool is live right now on the web and in the iOS app. Android users will have to wait a bit longer. To use it, you simply look for the three-dot “…” menu on any given post. Tap it. Select “Scan for AI text.” A small popup appears instantly. It shows a percentage breakdown: how much is likely AI-generated, how much is AI-assisted, and how much is purely human.
There are rules to this. The text has to be long enough. If you are looking at a snippet shorter than 100 words, the tool shuts you down. It needs more data to give a reading. There is also a timeline restriction. Only content published after 11:30 a.m. Pacific on July 21 is eligible. Try scanning an older post, and you will get a popup telling you it is ineligible. Substack is being cautious about historical accuracy with this metric.
Why Substack cares about provenance
Chris Best, the co-founder and CEO, put it plainly in the announcement blog post. The issue is deception. When a reader spends twenty minutes on an essay, only to find zero human thought behind it, the trust breaks. The platform is trying to preserve that trust.
It is worth noting that the detection relies on user initiative. Pangram only scans text if you ask it to. The ball is in the reader’s court. You have to want to know the truth. The tool has limits, too. It cannot see if a human used AI for background research before picking up a pen (or keyboard). It only flags the final text structure.
Best even added a personal note on his own posts via the tool. “I’m still figuring all of this out,” he wrote. “In the meantime these words are mine,for better or worse.” A candid admission. He wants transparency, even if the science of detection is still young.
Features for authors and writers
It is not just a reader-facing gadget. Writers get tools, too. Authors can now add a custom statement explaining their writing process. If you used AI, say so. If you didn’t, clarify that too. This note appears when someone runs the detection tool on your work.
Authors also have power. If you believe Pangram mislabeled your work, you can run the detection yourself and submit a report. You can dispute the finding. You can even remove the AI detection flag entirely by disabling the feature on your content. It is your choice whether to keep the label visible, regardless of what the algorithm says.
How does it actually work under the hood? Pangram uses a classifier neural network. It is trained to distinguish human patterns from AI patterns. To avoid bias, they trained the model on human-authored text published before 2021. That was before generative AI flooded the internet. This ensures the baseline for “human” is pure. It keeps AI-generated samples out of the training data, which prevents false positives later.
The bigger picture of online writing
We are living in a moment where the volume of AI text is exploding. A recent study by digital marketing firm Graphito claims that online articles written primarily by AI now equal those written by humans. The landscape is shifting rapidly.
Several publications have faced backlash for relying on AI. Some of these articles were filled with hallucinations and errors. Others were attributed to fake authors created by AI. It creates a confusion spiral. Readers stop knowing who is real.
Substack sees this coming. They are considering more features. Imagine filtering your feed to show only human-written content. You could set preferences to exclude AI entirely. Best says they are listening to user feedback. If you want purity, the tool might soon give it to you.
But for now, you have to do the legwork. You have to tap that menu. You have to scan. The tool is a start. It is not a perfect balance. It is a signal in the noise. Will it catch everything? Probably not. But it asks the right question: who is actually talking to you? The answer matters.
