A month after Substack announced its partnership with Pangram, the platform now surfaces whether a post appears to have been written with AI involvement. The stated aim is to deter low-effort, machine-generated content (“AI slop”) and to let readers decide whether they want to consume material produced with AI assistance.
Pangram’s claims and known limits
Pangram 4.0, the version cited in the discussion, reports a very low false-positive (FP) rate of about 0.0041%—roughly one false positive per 24,000 documents. The author emphasizes that while this rate is close to the practical minimum, it is not zero, and even a single unjustified flag is problematic from a fairness standpoint. The piece references the classical legal concern (better ten guilty go free than one innocent suffer) but concludes that, in this context, the risk of a few mistakes may be acceptable to protect the platform’s integrity.
According to Pangram’s published technical claims, the model can distinguish AI-edited text from interspersed human–AI authorship in a single pass, is robust to humanizers and adversarial prompts (identifying AI involvement in humanized output 98.83% of the time across 13 popular commercial tools), and generalizes across model families with false negative rates (FNR) below 0.7%. The author cautions these claims should be backed by independent third-party evaluations.
The trade-off between false positives and false negatives
The article explains the established trade-off: prioritizing minimal false positives (to avoid punishing honest human writers) tends to increase false negatives (AI texts labeled human). The author cites a prior figure for false negatives at about 0.34% (1 in ~300), arguing that while FP may be extremely low, FN at that scale could be unacceptable if widely deployed. The point is that detection design choices matter and reflect value judgments about which errors are more tolerable.
The partnership’s practical function: deterrence and transparency
Substack’s implementation is not an automatic punishment system: Pangram’s flags do not automatically ban users, and Substack users can opt out of Pangram checks. The feature functions primarily as an informational disclosure—readers can see whether content is likely AI-assisted and decide to engage or not. The author sees this as primarily a deterrent: if the initiative works on Substack, other platforms and institutions may feel institutional pressure to follow, making it harder for large-scale slop to thrive.
Social risks: the mob, nuance loss, and misinterpretation
A central worry is social dynamics. Even with a good detector, public reactions can be blunt and punitive: accusations based on single-dimensional signals invite mob-like responses. Substack’s guidance notes that Pangram detects AI involvement but not the amount of human care in a piece, nor whether AI tools were used as research or drafting aids. The article warns readers and writers to learn detector limits—short texts can be brittle, chunking can produce inconsistent scores—and not to weaponize the tool as a substitute for nuanced judgement.
Practical advice for writers and readers
The author offers practical recommendations: if you truly write your material yourself, don’t rely on Pangram; if you worry your draft shows machine fingerprints, use the detector to iterate and restore distinctiveness. The broader counsel is to cultivate idiosyncrasy in writing—being as "weird" on the page as you are in life protects against machine-like homogenization. Readers should avoid instant condemnation and learn how detection works before judging others.
Conclusion: a contested but potentially constructive step
The author concludes that Substack’s move to partner with Pangram is, on balance, a constructive attempt to reduce AI-generated low-quality content. It is neither perfect nor risk-free: detectors are imperfect, false positives matter, and social backlash is a real threat. Still, the partnership could provide meaningful deterrence and transparency, pushing platforms toward a cleaner information environment. The final trade-off is societal: are we willing to accept technical imperfections and some social frictions in return for less machine-generated dilution of public writing?
