Author: Chris Best
Compiled by Deep潮 TechFlow
DeepChaio Summary: Up to 40% of content on social media is now AI-generated, yet readers are often unaware. Substack believes this "Claudefishing"—impersonating human writing—is eroding trust between readers and authors and threatening creators' livelihoods. Their response: let readers decide for themselves, rather than having the platform decide for them.
The internet, where truth and falsehood are hard to distinguish
It’s becoming increasingly difficult to distinguish what’s real on the internet.
This issue is relatively mild on Substack, and we hope to keep it that way. As a first step, we’re launching a feature today in the Substack app in partnership with Pangram. Going forward, you’ll be able to scan text to see how much of it may be handwritten versus completed with AI assistance. I’ll elaborate more shortly, but first let me explain why this matters.
There is water everywhere, but not a single drop is drinkable.
According to Pangram’s estimates, the proportion of AI-generated text on certain platforms is rising, reaching as high as 40%. This is a significant shift. Many people complain that their feeds are flooded with soulless junk content, while others aren’t even aware they’re reading something created by no human.
The core issue isn’t that people use AI, nor is it the quality of AI outputs. Not all content created with AI is garbage, and not all garbage is created by AI. The problem arises when readers’ expectations don’t match reality—especially when they unknowingly invest attention in content that lacks any human thought behind it. This is Claudefishing.
Freddie deBoer said:
I engage with human-created art because I know there is a person behind it—that’s exactly what I’m seeking: others revealing through art the hidden parts of themselves. To deceive me in this process would be nothing but a fraud.
We care about this issue on Substack because it gets to the heart of what we do: building economic engines for culture. This means helping real people—those who think, feel, observe, argue, experiment, reflect, struggle, advise, and care—earn money through the work they believe in. This network is built on trust between people, and that’s why it works.
When AI-generated content displaces human-created content on the internet, it pollutes the public sphere, making it difficult to discover and hear human voices. When readers are forced to question whether what they’re reading is authentic, it erodes trust in authorship and threatens the livelihoods of writers—including those who thoughtfully use AI tools to create work they genuinely believe in. Platforms that reward false content will trigger a race to the bottom.
What are we doing?
This is an emerging field, and we don’t claim to have all the answers. But one thing we do know is that we don’t want to wait until your Substack app becomes like LinkedIn before we start learning and making progress. So first, we’re introducing some new tools aimed at increasing transparency and laying the groundwork for your feedback.
We are partnering with the leading AI detection tool, Pangram. You will be able to scan notes, replies, comments, and posts to see an estimated percentage of text that is handwritten versus AI-assisted. This will apply to all texts over 100 words posted from today onward, and analysis results will only be visible to those who request to view them.
Image: Substack’s AI detection tool interface in collaboration with Pangram, showing the percentage of text written by humans versus AI-assisted. Source: Substack
Please note that Pangram can only detect whether AI was used to generate text; it cannot determine the amount of human effort invested in the creation process or whether AI tools were used as sources. While not perfect, independent studies show that it can detect AI-generated text with high accuracy. We believe it is beneficial for people to use such tools to assist in making informed judgments.
We’ve also introduced tools for creators, including the ability to add an “How I Created This” statement, where you can explain your process and set expectations for readers. You can run Pangram on your draft before publishing, and you can report and remove any scan results about your work that you believe are inaccurate.
These features are now available on the web and iOS, and will soon be coming to Android. This is just the beginning. Based on feedback and interest, we are considering the following tools:
Set your community's preferences for AI content in your reply rules.
Provide creators with more ways to express the personal value they offer, whether through human writing or thoughtful use of AI.
Allow readers to set their preferences for what content they would like to be recommended.
As AI accelerates these issues, continue improving our systems to combat spam, bots, and various scams.
We do not oppose people using AI to assist with their work; we believe individuals should be free to choose the tools they use to express themselves. On our platform, some are maximalists of human writing, while others openly explore the frontiers of AI tools, paying close attention to detail to create the content they support. We have long used AI on Substack to write software, conduct research, and build product features such as clipping and translation.
But people should know what they’re getting. For example, you might not mind reading someone’s sharp analysis that heavily relies on AI to present research findings, but you only want to spend time interacting with human-written responses. Either way, you should have the choice.
What would you like to see?
We’re all entering this strange new world together, and we can’t think of a better way to figure out the new tools and social norms we need than through open debate. We’d love to hear your thoughts: How can Substack help support valuable human creation in the age of AI? When is using AI “Claudefishing,” and when is it okay?
No matter how powerful AI becomes, the human perspective remains a fundamental element of culture—and it always will. In fact, the more you expect these tools to improve, the more important it becomes to understand the source of what you’re reading. Claude has many strengths! But when I want Claude’s opinion, I ask Claude directly.