The initiative was unveiled by Substack CEO Chris Best in a detailed announcement titled "Against Claudefishing." The term, coined by Best, refers to the practice of creators using large language models—specifically naming Anthropic’s Claude, though applicable to all AI models—to generate content while presenting it as original, human-produced work. According to Best, the goal of these new tools is not necessarily to ban AI, but to restore the "human-to-human" trust that serves as the foundation of the Substack ecosystem. By providing a "check for AI" button, the platform aims to empower readers to make informed decisions about which writers they support with their time and subscription dollars.
The Mechanics of AI Detection on Substack
The new scanning tool is integrated directly into the Substack interface for web and iOS users, with an Android version currently in development. The functionality is straightforward but governed by specific parameters to ensure a baseline of reliability. Scans can be performed on any piece of text—be it a long-form article, a short-form "Note," or a reply—that exceeds 100 words in length. Furthermore, the tool is currently limited to content published from July 21 onward, reflecting the date when the platform began indexing metadata necessary for the Pangram integration.
One of the most notable aspects of the feature is its focus on individual user agency and privacy. When a reader chooses to scan a post, the result is visible only to them. This prevents the platform from becoming a public "shaming" ground where posts are automatically flagged with permanent scarlet letters of AI involvement. Instead, it serves as a private verification tool. If a scan returns a high probability of AI generation, the reader can choose to unfollow, unsubscribe, or simply adjust their perception of the author’s "voice."
Pangram, the technology partner behind the scanner, is an AI-detection firm that has gained traction by focusing on the linguistic patterns and statistical "fingerprints" left behind by models like GPT-4, Claude 3.5, and Gemini. While AI detection is notoriously difficult and prone to both false positives and false negatives, Best pointed to independent evaluations that credit Pangram with a high degree of accuracy compared to first-generation detection tools. The tool provides a probability score, acknowledging that in the realm of synthetic media, absolute certainty is rarely achievable.
Addressing the "Claudefishing" Phenomenon
The introduction of the term "Claudefishing" highlights a specific anxiety within the writing community. As AI models have become more sophisticated, they have moved beyond the robotic, repetitive prose of early iterations. Modern models are capable of mimicking specific tones, injecting simulated "personal anecdotes," and maintaining a conversational flow that can easily deceive a casual reader.
On Substack, where the primary value proposition is the unique perspective and "soul" of the writer, the use of AI to ghostwrite entire newsletters is seen by many as a fundamental breach of contract. "Substack is for writing that is worth paying for," Best noted in his announcement, "and part of that value is knowing there is a person on the other end." The phenomenon of "Claudefishing" is particularly insidious because it targets the emotional connection between creator and audience, using high-quality synthetic text to harvest engagement and subscription revenue without the labor of original thought.
Tools for Creators: Transparency and Defense
Recognizing that the relationship between writers and AI is complex, Substack is also rolling out features designed to protect and empower creators. Alongside the reader-facing scanner, the platform is introducing a "How I make this" disclosure section. This allows writers to proactively describe their creative process, including their use of AI for research, outlining, or editing, thereby setting clear expectations for their audience from the outset.
Furthermore, writers are being given the same tools as readers. Authors can run Pangram scans on their own drafts before they hit "publish" to see how the algorithms perceive their work. This is particularly useful for writers whose natural style might inadvertently trigger detection algorithms. In cases where a reader’s scan produces a "false positive"—flagging human work as AI—writers have the option to report and contest the result. This feedback loop is intended to help refine Pangram’s models and provide a recourse for authors who feel unfairly characterized by the software.
The Data: A Growing Tide of "AI Slop"
The timing of Substack’s rollout is driven by alarming data regarding the state of the broader internet. Pangram recently estimated that on certain high-traffic social platforms, as much as 40% of all posts are now fully AI-generated. This "AI slop"—a term used to describe low-quality, high-volume synthetic content designed to game algorithms—has begun to degrade the user experience on platforms like LinkedIn, X (formerly Twitter), and Facebook.
A recent report indicated that LinkedIn, in particular, has seen a massive surge in AI-generated "thought leadership" and comments, leading the platform to begin quietly reducing the reach of posts it identifies as synthetic. Similarly, Meta has been testing internal AI detection tools to manage the avalanche of AI-generated imagery and text across Instagram and Facebook. Substack’s approach differs by placing the detection tool directly in the hands of the user, rather than relying solely on a centralized, opaque algorithmic filter.
Chronology of the AI Detection Arms Race
The integration of Pangram into Substack is the latest chapter in a rapidly evolving timeline of AI governance in digital publishing:
- Late 2022: The launch of ChatGPT triggers a wave of AI-generated content across the web, leading to the "Dead Internet Theory" gaining mainstream traction.
- Early 2023: Early AI detectors (including one from OpenAI itself) are released but are widely panned for low accuracy and high false-positive rates.
- Mid 2023: Academic and professional publishers begin seeing a surge in AI-assisted submissions, leading to bans on AI-generated text in many scientific journals.
- Early 2024: AI models like Claude 3 and GPT-4o achieve a level of "human-like" nuance that makes manual detection nearly impossible for the average reader.
- July 2024: Substack begins the backend work for Pangram integration, setting the July 21 cutoff for scanable content.
- March 2025: Substack officially launches the "Against Claudefishing" initiative, providing the first major consumer-facing AI scanner on a social publishing platform.
Broader Impact and the Future of Trust
The implications of Substack’s move extend far beyond its own platform. By normalizing the "Check for AI" button, Substack is setting a new standard for transparency in the creator economy. If successful, this could pressure other platforms—such as Medium, Ghost, or even traditional news outlets—to provide similar verification tools.
However, the move also highlights the ongoing "cat-and-mouse" game between AI developers and detection firms. As detection algorithms improve, AI models are being trained specifically to evade them. This raises questions about the long-term viability of detection as a solution. Critics argue that detection tools may eventually become obsolete as synthetic text becomes indistinguishable from human writing at a statistical level.
For Substack, the gamble is that the platform’s specific culture—one built on direct financial support for individual voices—will prove resilient. Even if the technology is imperfect, the presence of the tool serves as a deterrent against "Claudefishing" and a signal to the community that human authenticity remains the platform’s north star. As the digital landscape becomes increasingly saturated with synthetic media, the ability to verify a "human in the loop" may become the most valuable commodity in the information economy.
In the coming months, the industry will be watching closely to see how Substack users engage with these tools. Will readers frequently "audit" their favorite writers, or will the tool be reserved for suspicious outliers? Will the "How I make this" disclosures lead to a more nuanced conversation about the role of AI in the creative process, or will it further polarize the community? Regardless of the outcome, Substack has taken a definitive stand in the battle for the "human" web, providing a rare moment of transparency in an era of increasing digital ambiguity.
