How to spot AI science slop

A practical checklist explains how LLM-generated science content reliably produces convincing but error-prone material and why readers should treat it skeptically. Large language models tend to hallucinate, overstate confidence, and craft fluent-sounding arguments that exploit cognitive biases, so polished grammar or formal tone are poor indicators of accuracy. Rather than relying on single signals or AI-detection tools, the recommended approach is to look for many concurrent red flags and verify important claims against primary sources, because casual callouts can be counterproductive and public shaming has social harms.

Concrete tells to watch for include recurring linguistic quirks (overuse of em dashes, semicolons, dramatic short sentences, clichéd parallel constructions, lists of three, or emoji-laden bullet points), formatting typical of chatbots or automated infographic tools, and remnants of chatbot prompts or labels. Other warning signs are a generic, grandiose “voice” that inflates ordinary facts, unusually high output across diverse topics, sudden stylistic shifts, unquestioning enthusiasm for AI, implausible specifics (fake quotes or references, contradictions, “tortured phrases”), poor organizational choices, and dubious credentials or lack of sourcing. When many of these appear together, trustworthiness is low and independent fact-checking or simply unfollowing is advised.