Author: Chris Best

Translated by: Deep Tide TechFlow

Deep Tide Introduction: Up to 40% of content on social media is already generated by AI, but readers are often unaware. Substack believes this "Claudefishing" (impersonating human writing) is undermining the trust between readers and authors, threatening the livelihoods of content creators. Their response is to let readers make their own judgments, rather than have the platform decide for them.

The Hard-to-Distinguish Internet

It is becoming increasingly difficult to tell what is real content on the internet.

This issue is relatively mild on Substack, and we hope to keep it that way. As a first step, we are launching a feature today in collaboration with Pangram within the Substack app. From now on, you can scan text to see how much may be handwritten and how much is AI-assisted. I will elaborate on this later, but first let me explain why this is important.

Water Everywhere, Yet Not a Drop to Drink

According to estimates from Pangram, the proportion of text generated by AI on certain platforms is rising, reaching up to 40%. This is a significant change. Many people complain that their feeds are filled with soulless garbage content. Some even unaware that they are reading something created without human input.

The core issue is not that people are using AI, nor is it about the quality of AI output. Not all content produced with AI is garbage, and not all garbage is made by AI. The issue arises when readers’ expectations do not match reality, especially when they unknowingly invest attention into content that lacks any human thought. This is Claudefishing.

Freddie deBoer said:

I engage with art created by humans because I know there is a person behind it; that’s exactly what I’m looking for—others showing me what they hide within themselves through art. Deceiving me in this process is merely a scam.

We at Substack care about this issue because it touches the core of our work: building an economic engine for culture. This means helping real people—those who think, feel, observe, argue, experiment, ponder, struggle, advise, and care—to earn money from the work they believe in. This network is built on trust between people, which is why it works.

When content produced without human input occupies parts of the internet that should belong to humans, it pollutes public space, making it difficult to discover and hear human voices. When readers have to question whether the content they are reading is genuine, it undermines trust in author identity and threatens writers' livelihoods—including those who thoughtfully use AI tools to create the work they believe in. Platforms rewarding false content will trigger a race to the bottom.

What We Are Doing

This is an emerging field, and we do not claim to have all the answers. But one thing we do know is that we do not want to wait until your Substack app becomes LinkedIn before we start learning and making progress. So first, we are rolling out some new tools aimed at increasing transparency and laying the groundwork for us to gather your feedback.

We are collaborating with leading AI detection tool Pangram. You will be able to scan notes, replies, comments, and posts to see an estimate of how much of the text is handwritten or AI-assisted. This will apply to texts with more than 100 words published from today, and the analysis results will only be shown to those who request to view them.

Image: Substack's AI detection tool interface in collaboration with Pangram, showing the ratio of text written by humans versus AI assistance. Source: Substack

Please note that Pangram can only detect whether AI was used to produce the text and cannot assess whether a significant amount of human effort was invested in the creation process, nor can it detect if AI tools were used as sources. It is not perfect, but independent studies show it can detect AI-generated text with high accuracy. We believe that using such tools to help people make judgments is beneficial.

We are also launching tools for creators, including the ability to add a “How I Created This” statement, where you can explain your process and set expectations for your readers. You can run Pangram on drafts before publication, and you can report and delete any scan results about your work that you believe are inaccurate.

These features are now available on web and iOS and will soon be launched on Android. This is just the beginning. Depending on feedback and interest, we are considering the following tools:

Allowing you to set preferences about AI content in your community's reply rules

Providing creators with more ways to express the personal value they offer, whether through human writing or thoughtful AI use

Allowing readers to set preferences about what content is recommended to them

Continuing to improve our systems to combat spam, bots, and various scams as AI accelerates these issues

We do not oppose people using AI to assist their work; we believe people should have the freedom to choose which tools to express themselves. Some people on the platform are maximalists of human writing, while others openly explore the frontiers of using AI tools while paying attention to details to create the work they support. We have always used AI at Substack to write software, conduct research, and build features like clipping and translation.

But people should know what they are getting. For example, you may not mind reading someone’s sharp analysis in which they heavily rely on AI to present research results, but only want to spend time interacting with human-written replies. Whichever way it may be, you should have the choice.

What Do You Want to See?

We are all entering this strange new world together, and we can’t think of a better way to figure out what new tools and social norms we need than through open debate. We want to hear your thoughts: How can Substack help support valuable human creation in the age of AI? When is using AI Claudefishing? When is it acceptable?

No matter how powerful AI becomes, the human perspective is a fundamental element of culture, and that will always be true. In fact, the more you expect these tools to improve continuously, the more important it becomes to understand the origins of what you are reading. Claude has many advantages! But when I want Claude's opinion, I will ask Claude directly.