SubstackがAI生成ブログを検出するツールを導入Substack adds an AI detector to help spot blogs written by no one

Substack has integrated an AI detection tool from Pangram, giving the newsletter platform a way to flag posts that appear to have been generated by artificial intelligence rather than written by a human. The move matters because Substack has built its brand around personal, subscription-based writing and direct relationships between authors and readers, and the growing volume of machine-generated text across the web has raised questions about how platforms can preserve trust and authenticity.

According to the company, the feature works by analyzing published posts and estimating the likelihood that they were produced by AI systems. Rather than issuing a simple yes-or-no verdict, detection tools of this kind typically return a probability or confidence score, which platforms can then use to inform moderation decisions, surface labels, or guide further review. The stated goal is not necessarily to ban AI outright, but to help maintain a clear signal about what readers are actually consuming and who, or what, produced it.

Pangram is one of several companies specializing in AI text detection, a field that has expanded rapidly alongside the popularity of large language models such as OpenAI's GPT series, Anthropic's Claude, and Google's Gemini. These detectors generally look for statistical patterns and stylistic markers that tend to distinguish machine-generated prose from human writing. Pangram has positioned itself as offering higher accuracy and lower false-positive rates than earlier tools, though independent verification of any detector's real-world performance remains difficult, and results can vary significantly depending on the type of text being analyzed.

The reliability of AI detection is an important caveat. The broader industry has a mixed track record: OpenAI quietly retired its own AI text classifier in 2023, citing low accuracy, and academic studies have repeatedly warned that detectors can misclassify human writing as AI-generated. This is a particular concern for non-native English speakers, whose writing has sometimes been flagged at higher rates. Any system Substack deploys is therefore likely to be treated as a signal rather than definitive proof, and how the company handles borderline cases and potential appeals will shape whether writers view the tool as fair.

The change fits into a wider pattern across content platforms grappling with generative AI. Search engines, social networks, and publishing services have all been experimenting with disclosure requirements, labeling, and detection. Google has adjusted its search guidance to focus on content quality regardless of how it is produced, while sites such as Medium and various online marketplaces have moved to limit low-effort, mass-produced AI content. Provenance initiatives like the Coalition for Content Provenance and Authenticity, known as C2PA, take a complementary approach by attaching cryptographic metadata to media at the point of creation, though such standards are more established for images than for plain text.

For Substack specifically, the stakes are tied to its business model. The platform relies on readers paying for newsletters they trust, and the presence of undisclosed AI-written content could erode confidence in the value of a paid subscription. At the same time, many legitimate writers already use AI tools to help with research, editing, or drafting, which complicates any strict line between human and machine authorship. How Substack defines and communicates acceptable use will likely be as consequential as the detection technology itself.

It remains unclear from the initial announcement exactly how the detection results will be presented, whether to readers, to writers, or only internally to the company. Details such as whether posts will carry visible labels, how scores factor into recommendations or monetization, and what recourse authors have to contest a flag will determine the practical impact. The company has framed the effort as part of an ongoing commitment to content integrity rather than a finished product.

The introduction of the tool reflects a broader reckoning across the media and technology sectors, where the line between human and AI-generated work is becoming harder to see. Whether detection can keep pace with rapidly improving language models is an open question, and platforms may ultimately need to combine automated tools with transparency policies and human judgment to sustain reader trust over the long term.