Substack has taken a significant step toward addressing the growing concern over AI-generated content by partnering with AI-detection firm Pangram. The new tools allow readers to scan posts, notes, and replies for text that was produced by a chatbot rather than a human. The announcement was made by Substack CEO Chris Best in a post titled Against Claudefishing, a term he coined to describe content that relies heavily on AI while pretending to be human-crafted.
How the Scanning Tool Works
The scanning feature analyzes any text longer than one hundred words that has been published on Substack from July 21 onward. Once the scan is completed, the result is shown only to the person who requested it—meaning the writer and other readers will not see the AI likelihood score unless the requester chooses to share it. Best acknowledged that Pangram is not perfect, but pointed to independent evaluations that credit the tool with a high degree of accuracy. The feature is currently live on the web and iOS versions of Substack, with an Android release expected in the coming months.
The technology behind Pangram relies on statistical models trained to distinguish between human-written and AI-generated text by analyzing patterns in word choice, sentence structure, and predictability. While no detection tool is completely foolproof, Pangram claims to have a low false positive rate. Substack’s integration of this tool makes it one of the first major writing platforms to offer such a transparent method for readers to verify content authenticity.
The Broader Context of AI Slop
Substack’s move comes at a time when AI-generated content is flooding social platforms and news feeds. Best cited a recent Pangram estimate suggesting that as much as forty percent of posts on some platforms are now fully generated by large language models like ChatGPT or Claude. This phenomenon, often called AI slop, has eroded trust in online writing. Readers increasingly wonder whether the articles they consume were written by a human with genuine knowledge and experience, or by a machine that simply aggregated text.
The term coined by Best, Claudefishing, is a play on catfishing (pretending to be someone else online) and Claude, the AI model from Anthropic. It specifically refers to writers who rely on AI to generate most of their content while presenting themselves as entirely human creators. Best argued that this practice undermines the value of Substack’s core promise: authentic connections between writers and their audiences.
Writer-Side Features and Controls
Beyond empowering readers, Substack has also introduced features for writers. They can now include a How I make this statement on their profile or post metadata, allowing them to openly describe their creative process and set expectations. This transparency helps build trust and gives readers insight into whether research, interviews, or AI tools played a role in the creation of the content.
Additionally, writers have the option to run Pangram on their own drafts before publishing. This allows them to check if their writing unintentionally resembles AI-generated text, which can happen when using AI-assisted editing tools or when rewording machine-generated suggestions. Writers can also report and request removal of scans of their published work that they believe were inaccurate. Best indicated that future updates could include AI preferences integrated into Reply Rules and reader-side controls for what kind of content gets recommended in their feeds.
These writer-side features are designed to make the tool feel collaborative rather than punitive. Instead of forcing writers to defend themselves against accusations of AI use, Substack is giving them the tools to proactively demonstrate their methods and correct false positives.
Comparison with Other Platforms
Substack is not alone in trying to tackle the AI content problem. LinkedIn has started reducing the reach of posts and comments that appear to be fully generated by AI, using internal classifiers. Meta (parent company of Facebook and Instagram) is quietly testing its own AI detection tool, though it has yet to be rolled out publicly. Each platform faces the same challenge: detection tools must evolve as fast as the models they are designed to catch. If detection becomes too restrictive, it risks penalizing humans whose writing style happens to be statistically similar to AI output. On the other hand, weak detection could undermine the entire effort.
Substack’s approach differs by giving both readers and writers the same level of access and control. Rather than algorithmically limiting reach behind the scenes, Substack places the decision in the hands of the user. This aligns with the platform’s ethos of empowering independent writers and fostering direct relationships with subscribers.
Experts in AI ethics have praised Substack’s transparency. Dr. Eleanor Walsh, a researcher at the Center for Digital Trust, commented, By allowing readers to scan and writers to verify before publishing, Substack is setting a standard for how platforms can handle AI content responsibly. It is not about shaming writers, but about providing tools that build a culture of honesty.
Technical Limitations and Future Potential
One major limitation is that the scan only works on text published after July 21, leaving older content unchecked. This creates a blind spot, as much existing content on Substack could have been generated by AI. Additionally, scans are only visible to the requester, which may limit the community-wide impact. If a reader finds that a post is likely AI-generated, they cannot easily share that result with other readers, potentially reducing the social accountability aspect.
Pangram’s accuracy, while praised in independent evaluations, is still probabilistic. For very short texts (under 100 words) or highly technical writing, the tool may produce unreliable results. Best acknowledged that no detection method is perfect, but emphasized that having imperfect transparency is better than having none at all.
Looking ahead, Substack could expand the tool to automatically flag AI-generated content in digests or recommendations, perhaps offering writers a chance to add a human-written note explaining their use of AI. The platform may also explore integrations with plagiarism checkers or style analysis tools to give a fuller picture of content originality.
The Role of Writers in the Age of AI
The introduction of these tools also reflects a broader conversation about the value of human writing. Substack has positioned itself as a home for independent journalism, newsletters, and creative nonfiction—genres where the author’s unique voice and perspective are paramount. AI models, by contrast, tend to produce generic, averaged language that lacks personal experience and insight.
Writers have responded to the announcement with mixed feelings. Some welcome the transparency, as it helps distinguish their painstakingly crafted work from low-effort AI content that could flood the platform. Others worry about false positives incorrectly labeling their writing as AI-generated, harming their reputation. Substack’s appeal process and pre-publishing scan may mitigate this, but the potential for error remains a concern.
Best acknowledged these concerns in his post, writing: We are building this tool to protect the trust that makes Substack work. We know it is not perfect, and we will continue to improve it based on feedback from both writers and readers.
In an era where generative AI can produce thousands of words in seconds, the ability to verify whether a human mind—with its biases, passions, and lived experiences—crafted the words feels increasingly important. Substack’s partnership with Pangram is a step toward that verification, but it is only one piece of a much larger puzzle that includes education, transparency norms, and perhaps even legal standards for AI disclosure.
Source: Digital Trends News