Substack AI detector sparks debate over detection futility

The author opens with a blunt assertion: AI detection is futile and will not matter in a few years because the share of AI-generated articles has already surpassed human-created work. AI detectors themselves are unreliable, easily gamed, and produce too many false positives. The irony, the author notes, is that these detectors are trained on mountains of human-written content and then use that same body of work as a measuring stick to decide what is human or not—a logical trap that flags human-style writing as non-human.

The author acknowledges that most people outside writer bubbles do not care about the manufacturing process of content, only whether it is insightful, entertaining, or moving. In five years, the author predicts, nobody will care about AI detectors. The real questions that matter are whether content communicates something useful, valuable, and reliable, and whether it is enjoyable.

The author's stance centers on intention: if AI is part of a writer's process (helping workshop a title, edit repetitive sections, or refine an introduction), that is acceptable. If AI is the entire process (a user types "write me a personal essay" and publishes the result unchanged), that is not. The problem with Substack's new partnership with Pangram—an AI detector claiming almost 100% accuracy with a 1-in-10,000 false positive rate—is that it will enable users to hunt down any trace of AI use, regardless of how it was deployed. The author predicts "public lynchings" and a new form of cancel culture, where users crawl through a writer's back catalogue to find a single flagged post and weaponize the detection against them. Even if flagged for writing patterns like the "rule of three" (an ancient rhetorical device) or em dashes, writers could face pile-ons of outrage and righteousness, all based on a tool that may not be accurate.

The author concludes that the question should not be whether people use AI, but how they use it and whether the output delivers value. AI detectors measure only mechanical involvement, not thinking, effort, ideation, or editing. Substack's feature risks replacing genuine editorial judgment with a metric that is easy to measure but far less important than the one that actually matters: quality.

The article reflects a broader tension in online discourse around AI authenticity. The author acknowledges that AI-generated content already dominates platforms like Substack and that detection is a losing game—AI detectors are trained on human writing yet paradoxically use that same human writing as their baseline to flag non-human output, creating circular logic. The move to surface AI detection on every post, comment, and note appears to address legitimate user concerns about transparency and choice, but the author argues it inverts the priority: instead of measuring whether content is valuable, it measures only whether AI was mechanically involved, a distinction the detector cannot reliably make.

The author's core argument hinges on intention: using AI to workshop a title or tighten prose differs fundamentally from generating an entire essay from a prompt and publishing it unchanged. Yet Pangram's binary scoring ("98% human written") cannot capture that nuance. The risk, the author contends, is that Substack's feature will devolve into performative outrage, where readers flag ambiguous detections and pile on writers based on inaccurate signals, transforming the platform into a space policed by false positives rather than genuine editorial judgment.