This Is How to Tell if Writing Was Made by AI

Odd Lots Podcast Recap

Published:

Duration: 48 min

Guests: Max Spiro

Summary

Tracy Alloway and Joe Wiesenthal explore the challenges of identifying AI-generated writing, featuring insights from Max Spiro of Pangram Labs. The episode discusses the subtle differences between human and AI writing and the implications of AI content proliferation.

What Happened

Fidelity Trader Plus, a new advanced trading platform, is now available on multiple digital platforms, marking a significant step in trading technology. Tracy Alloway and Joe Wiesenthal delve into the complexities of identifying AI-generated writing, noting that while AI writing often presents grammatical precision and remarkable turns of phrase, it typically lacks distinct style. Some books have begun including disclaimers to reassure readers that they were crafted without AI assistance, a testament to growing concerns about AI-authored content.

Max Spiro, CEO of Pangram Labs, joins the conversation, explaining how his company's service can effectively identify AI-generated text. Joe Wiesenthal recounts his testing of Pangram's service with translations, finding it notably effective. Pangram Labs' AI detection model boasts a false positive rate of about 1 in 10,000 and a false negative rate near 1%, leveraging deep learning trained on millions of examples to distinguish nuances between human and AI writing.

The model can detect writing patterns that are distributionally different from human-authored content, even if these differences are subtle. Pangram Labs estimates that around 40% of internet content is now AI-generated, with over half of new Medium articles being AI-generated as of a year and a half ago. Similarly, AI-produced content on Reddit has increased from 7% to over 10% in the past year.

AI-generated content is being used strategically by some companies to create organic mentions on platforms like Reddit, influencing model outputs by entering training data. Pangram Labs employs active learning to continually improve their model's accuracy by identifying and correcting errors. As AI-assisted writing becomes more prevalent, services like Google Docs and Grammarly are integrating AI to enhance user experience.

Pangram Labs differentiates AI-assisted from AI-generated writing by measuring the cosine difference in hyper multidimensional space. Their mission is to detect AI-generated content, providing value to various sectors including education, law, publishing, and individual internet users. Major internet platforms like Quora use Pangram's services to moderate content and identify AI bad actors.

As AI models become more sophisticated, their output becomes increasingly complex, necessitating larger models for accurate detection. Perplexity is used as a measure to determine how expected a piece of text appears to a language model, with low perplexity indicating less surprising or more predictable text. Early AI detectors encountered challenges with non-native English speakers due to the low perplexity of their writing.

Efforts to combat undisclosed AI-generated content include proposing norms against sending such outputs. Internet giants like Google are navigating the dual role of promoting AI-generated content while filtering out what is termed as 'AI slop' from search results. The severance of craft from output due to AI advances means that well-written content no longer reliably indicates the intelligence of the author.

Key Insights

View all Odd Lots recaps