How AI detection actually works, with Pangram’s Max Spero
Summary
Patrick McKenzie interviews Max Spiro, CEO of Pangram, about AI detection and the societal impact of synthetic text. They discuss how AI saturates bureaucratic processes like grant applications and legal filings, overwhelming human reviewers. The conversation covers Pangram’s technical approach to distinguishing human from AI-generated content and the shifting value of human expertise in an automated world.
Topics discussed
Introduction and the motivation behind building Pangram
AI as a threat to friction-based proof of work systems
Why AI detection is harder than simple text reading
How Pangram's models detect AI-generated text and images
Training data, style transfer, and the 'smell' of LLMs
The increasing hostility of the internet towards web crawlers
Personal anecdotes: AI replicating author style for letters
Credit reporting agencies and the value of human effort
The 'ick factor' and low information density of AI text
The changing role of software engineers in the age of AI
Sponsors: Mercury and Granola
Feedback loops in coding and the future of programming languages
Mode collapse in AI recommendations and creative outputs
AI-generated fan blogs and the erosion of trust
Cognitive surplus, TikTok, and the attention economy
The cat-and-mouse game of AI detection and evasion
Reputation systems and the norms around AI disclosure
Model welfare and the ethics of training on human data
The diminishing amazement of AI and the value of technology
Listen ad-free on Castria