Substack’s AI Detector and the Question of What’s Real Online

概览

This episode focuses on Substack’s new AI detection feature, powered by Pangram, and why the company says it added it: to help readers understand whether text was likely written by a person or generated by AI.

Substack CEO Chris Best argues that the goal is transparency rather than policing creators’ tools. He says AI can help people make work they stand behind, but it can also flood platforms with low-effort content that weakens trust.

The discussion also covers the limits of AI detection, especially false positives, and the risk that writers may start optimizing for detection tools instead of audiences. Best frames the broader question as one of emerging norms around when AI assistance feels acceptable and when it violates reader expectations.

分段落总结

[00:19] Substack Launches an AI Detector

[事实] The episode introduces Substack’s new AI detector, powered by Pangram. [事实] The feature lets users scan text on the website or iOS app and receive an estimate of how much was written by a human or by AI. [事实] The host says the feature has sparked discussion on Substack, a platform already known for discourse.

[00:58] Why Substack Added the Feature

[事实] Chris Best says it is becoming hard to tell what is real on the internet. [事实] He cites a Pangram study finding that on some platforms, including LinkedIn, as much as 40% of text may be generated by tools like ChatGPT or Claude. [事实] Best says readers can struggle when they are unsure whether they are reading something from a person or from AI. [推测] Substack sees AI disclosure as part of maintaining reader trust on a publishing platform built around personal voices.

[01:34] Creator Feedback and AI Disclosure

[事实] Best says feedback has ranged from annoyance about the feature to requests for stronger labeling or blocking tools. [事实] Some creators who use AI heavily were grateful that Substack was looking into disclosure. [事实] Best says thoughtful AI users can also be threatened by people producing large amounts of low-effort content. [推测] The feature is positioned as useful not only for AI skeptics, but also for creators who want to distinguish responsible AI use from spam-like output.

[02:33] The “How I Make This” Statement

[事实] Substack added a transparency note feature called a “how I make this” statement. [事实] Best says Substack is not trying to tell writers what tools they may use or tell readers what they may see. [事实] The feature gives writers space to explain their process, including whether and how they use AI tools. [事实] Best describes the problem as “Claude fishing,” when readers think something was written by a person but it was actually written by a machine.

[04:32] Accuracy Limits and Mislabeling Risks

[事实] The host notes that AI detection is itself a form of AI and is not perfect. [事实] Best says Substack takes mislabeling seriously and believes Pangram is probably the best tool on the market for this purpose. [事实] He says false negatives are a modest problem, while false positives are more serious because human-written work could be labeled as AI. [事实] Substack lets users report mistakes and remove a detection if it is clearly wrong.

[05:27] Risk of Public Shaming

[事实] The host discusses Derek Thompson’s post about rewriting an explanatory passage after Pangram labeled parts of it as AI-written. [事实] Thompson said he realized he was no longer writing for his audience, but writing to avoid being publicly shamed by AI detection. [事实] Best says Substack designed the feature to minimize that effect. [事实] He says the intended goal is transparency, so readers do not enter a piece with expectations different from reality. [推测] The exchange highlights a tension between disclosure tools and the possibility that writers may self-censor or change style to satisfy detectors.

[06:38] AI’s Place in Writing

[事实] Best says the role of AI-generated content on Substack is a deep philosophical question without a settled answer. [事实] He says the key thing readers subscribe for is a human point of view: what someone believes, feels, and wants to make. [事实] Best says AI tools can help people realize that point of view, like other tools. [事实] He also says AI can create new ways to be annoying online, such as asking ChatGPT to generate 10,000 viral posts. [推测] Best’s position is not anti-AI; it draws a line between AI as a tool for expression and AI as a substitute for human authorship.

[07:50] Emerging Norms Around AI Use

[事实] Best gives one example where AI assistance may be acceptable: a dense technical report where the analysis is strong and ChatGPT helps write up the text. [事实] He contrasts that with a story about someone using an AI agent to text their wife, which he says most people would not be okay with. [事实] He says much of the internet will fall between those extremes, and people will have to work out expectations together. [推测] The episode suggests that social norms around AI use may depend heavily on context, intimacy, and the reader’s expectations.

[08:26] LinkedIn and AI Slop

[事实] After the interview, the host says Pangram found more than 40% of LinkedIn content is likely AI-generated. [事实] The host says LinkedIn has rolled out a button to report content that seems like AI slop. [推测] The issue of AI-generated low-effort content is presented as broader than Substack and relevant across major online platforms.

[09:02] Post-Roll Promotion

[事实] The transcript ends with a promotion for another show, “This Is Uncomfortable.” [事实] The promoted episode features Rima Grace interviewing Abigail Disney about inherited wealth, guilt, giving money away, and conflict with family legacy.

播客点评/总结

The episode is valuable as a concise look at how a publishing platform is trying to respond to AI-generated writing without banning AI outright. Its strongest point is the distinction between tool use and reader deception.

A key limitation is that the episode relies mainly on Substack’s CEO to explain the feature, so it gives less space to critics, writers affected by false positives, or independent evaluation of Pangram’s accuracy.

[推测] This episode is best suited for listeners interested in AI disclosure, platform trust, creator tools, and the changing norms around authorship online.