AI loves negative parallelism

Key takeaways

Negative parallelism is a common indicator of AI-generated writing, characterized by statements that emphasize contrast, e.g., 'it's not X, it's Y.'
AI systems tend to favor specific language patterns, such as the overuse of em-dashes and negative parallel structures, due to how they predict the next words in a response.
Training on AI-generated content and synthetic data can reinforce these language patterns, making AI writing more predictable and easier to identify.
Humans also use negative parallelism, especially in rhetorical contexts, but AI may be adopting and amplifying these structures through its training process.
Detecting these linguistic markers can help distinguish AI writing from human writing, akin to a digital watermark, although it’s not foolproof.

Summary

AI's Repetitive Language Patterns and Negative Parallelism

The podcast discusses how AI-generated writing often exhibits a telltale pattern known as negative parallelism, characterized by statements like 'it's not X, it's Y.' This construction appears more frequently in AI outputs than human writing and has become a reliable indicator of AI-generated content. Examples include advertising slogans and political statements, where this structure emphasizes contrast. Detecting these patterns can help identify AI-written text without sophisticated software.

Evolving AI Detection Codes and Human Mimicry

Initially, AI detection focused on obvious cues, such as phrases like 'As an AI language model,' or overuse of certain punctuation like em-dashes. Over time, these cues have evolved, with negative parallelism becoming the most common tell. Interestingly, humans also use this structure, especially in rhetoric and literature, such as Shakespeare. Additionally, AI systems tend to favor this pattern because of how they predict the next words to generate responses, often choosing words that continue the negative parallel structure, leading to its dominance in AI outputs.

Impact of Training Data and AI Self-Generation on Language Patterns

The podcast highlights how large language models are increasingly trained on their own outputs and synthetic data, which may reinforce certain language patterns like negative parallelism. As AI models ingest more of their own generated text, these repetitive structures could become even more entrenched. Some efforts are underway to embed digital watermarks into AI-produced content, but verbal markers like negative parallelism could serve as emerging signs of AI authorship. Interestingly, human writers who frequently use AI start to adopt similar patterns, suggesting a convergence of AI and human language styles.

Show notes

Anthropic recently announced it's adding watermarks to all text outputted by its chatbot Claude. But often, you don't need fancy detection software to know; you just need to look for language that gets repeated over and over by AI chatbots, like “it's not X, it's Y.”

That's a rhetorical structure referred to as negative parallelism. And it's the topic for today's “Uncanny AI,” those times when it becomes clear AI doesn't think like us. Will Oremus recently wrote about this AI tic for The Atlantic.

More on this

“The Most Famous AI Writing Tic Is Also the Most Mysterious” from The Atlantic

“Move Over Em-Dash, There’s a New AI Tell in Town” from Ai-Ai-OH