Substack CEO on the platform's new AI detector

Summary

This Marketplace Tech episode interviews Chris Best about Substack’s new AI detector, powered by Pangram, and the platform’s attempt to make AI involvement in writing more legible to readers. Best frames the feature as transparency rather than an anti-AI rule: writers can use AI tools, but readers should not be misled about whether they are receiving a human point of view or machine-generated text.

The episode extends the wiki’s AI-authorship branch by connecting AI Writing Detection, AI Content Provenance, and AI Authorship Presence. Its main tension is that detection can reduce uncertainty, but false positives, public shaming, and detector-optimized writing can create their own trust failures.

Key Claims

• Substack’s detector lets users scan text on the website or iOS app and receive an estimate of how much was written by a person or generated by AI.
• Best cites a Pangram study suggesting that on some platforms, including LinkedIn, a large share of text may be generated by tools such as ChatGPT or Claude.
• Substack added a “how I make this” statement so writers can explain whether and how they use AI tools.
• Best says Substack is not trying to ban AI use or dictate what readers may see, but to make reader expectations clearer.
• False positives are treated as the more serious detector failure because a human-written piece can be mislabeled as AI-written.
• Substack lets users report mistakes and remove detections that are clearly wrong.
• The Derek Thompson example shows how public detector scores can push writers to optimize against the detector rather than write for their audience.
• Best argues that readers subscribe for a human point of view, while AI can be acceptable when it helps a person realize that point of view.
• The episode distinguishes AI as a writing tool from AI as a substitute for human authorship, especially in contexts where readers expect intimacy or personal judgment.
• The post-interview segment connects the same problem to AI Slop and AI Slop Detection on LinkedIn, where low-effort AI-generated posts can degrade platform trust.

Key Quotes

“Claude fishing” - Best’s term for the expectation violation when readers think they are reading a person but are actually reading machine-generated text.

“how I make this” - Substack’s creator-facing statement for disclosing process and AI assistance.

Connections

• Substack, Chris Best, and Pangram - platform, CEO, and detector-provider context.
• AI Writing Detection, AI Detector Bias, and Human Judgment Under AI - detector accuracy, false-positive, and review-boundary branch.
• AI Content Provenance, AI Authorship Presence, and Human Authorship Premium - disclosure, authorship expectation, and human-led creation branch.
• AI Information Pollution, AI Reality Verification Tax, AI Slop, and AI Slop Detection - platform-trust and low-effort AI-text branch.
• LinkedIn, ChatGPT, and Claude - platform and tool examples named in the source.

Contradictions

• No direct contradiction found with existing wiki content.
• The source qualifies Pangram and AI Writing Detection by showing a platform using detector output for transparency while still acknowledging false-positive and correction risks.
• The source qualifies AI Content Provenance by separating detector estimates and writer self-disclosure from stronger provenance mechanisms such as watermarks or content credentials.
• The source extends AI Authorship Presence by tying reader trust to whether AI helps express a human point of view or replaces the point of view the reader expected.