Substack Adds Pangram AI Detector to Flag Bot Blogs

Substack is wading into the murky waters of content authenticity with a new feature designed to answer a question that increasingly haunts every text-based platform: was this written by a human, or by a machine? The newsletter company has integrated an AI-detection tool powered by Pangram, according to The Verge, giving readers a signal about whether the words they're consuming were produced by artificial intelligence.

What Substack Is Actually Shipping

The feature surfaces an indicator when Substack's system, drawing on Pangram's classifier, believes a post was likely generated with AI assistance. Rather than outright banning AI-written content, Substack's approach leans toward transparency — flagging suspected machine-generated posts so that readers can make their own judgments about what they're reading and, presumably, whether they want to keep subscribing.

This is a meaningful distinction. Many platforms have wrestled with whether to prohibit synthetic content entirely, an approach that is nearly impossible to enforce given how fluent modern large language models have become. By opting for labeling over prohibition, Substack sidesteps the enforcement quagmire while still giving its community a data point about provenance.

Why Pangram Matters Technically

Pangram is one of a growing class of AI-text detectors that have tried to distinguish themselves from the notoriously unreliable early tools like OpenAI's now-retired classifier and various academic detectors that produced high false-positive rates. Pangram markets itself on low false-positive performance, a critical metric because the real damage from AI detection isn't missing a machine-written post — it's falsely accusing a human writer of using AI.

Text detection is fundamentally harder than image or video deepfake detection. Where synthetic media leaves visual artifacts, compression fingerprints, or frequency-domain anomalies that classifiers can latch onto, AI-generated text can be statistically indistinguishable from human writing, especially after light editing. Modern detectors typically analyze token-level probability distributions, perplexity, and burstiness — the variation in sentence complexity that human writers naturally produce and that older models tended to smooth out. As LLMs improve, these signals erode, which is why detection remains an arms race with no permanent winner.

The Broader Authenticity Picture

Substack's move fits a pattern we've tracked across the media ecosystem. Music streamer Deezer has reported that more than half of its daily uploads are AI-generated. YouTube has been clarifying its policies around AI "slop" and low-effort synthetic content. TikTok is testing deepfake detection to help creators protect their likeness. Across audio, video, and now text, platforms are being forced to build provenance signals into their products because the volume of synthetic content is overwhelming human moderation.

What makes text detection particularly consequential for Substack is the platform's identity. Substack sells itself on the direct, personal relationship between a writer and their audience — a relationship built on the premise that a real person is doing the writing. Undisclosed AI generation corrodes that trust in a way that hits Substack's core value proposition harder than it might hit a video platform.

The Reliability Caveat

The obvious risk is accuracy. No AI-text detector is perfect, and even a low false-positive rate produces meaningful numbers of wrongly flagged writers at Substack's scale. A misfired detection could unfairly tarnish a legitimate author, particularly non-native English speakers, whose writing patterns have historically triggered detectors more often. Substack will need to be careful about how prominently it displays these signals and how much weight readers assign to them.

There's also the cat-and-mouse dynamic. Writers determined to disguise AI output can run text through paraphrasers, humanizer tools, or simply edit heavily — all of which degrade detector accuracy. This means Substack's tool is best understood as a friction mechanism and a transparency gesture rather than a definitive authenticity verdict.

Why It Matters

Substack's integration is a small but telling data point in the larger movement toward content provenance infrastructure. As generative tools become ubiquitous, expect more platforms to bolt on detection layers, watermarking schemes, and disclosure requirements. The industry is slowly converging on the idea that authenticity itself is a feature worth building — and that readers deserve to know whether there's a human on the other end of the words.

Stay informed on AI video and digital authenticity. Follow Skrew AI News.