Researchers from Waseda University have discovered that adding false publication dates to content online can drastically increase its visibility in artificial intelligence (AI) models, including tools like ChatGPT. This finding suggests that such models systematically prioritize the most recent content over older and relevant material, which could affect the visibility of high-quality content that is not regularly updated.
Lie, that something remains
The study involved adding fictitious dates to documents from standardized test collections. The researchers evaluated how seven AI models, including GPT-4, GPT-3.5, and LLaMA-3, classified these results. Surprisingly, each model preferred the text with more recent dates. This “swing effect” resulted in younger content occupying privileged positions, overshadowing even old and authoritative sources, such as academic research and detailed guides.
The configuration “use_freshness_scoring_profile: true” found in the ChatGPT configuration files evidences that OpenAI’s reordering system explicitly encourages this preference for recent content. This bias towards novelty could lead to a temporary arms race, where content creators are forced to constantly update their material, either substantially or artificially, to maintain their visibility in AI searches.
Industry experts, such as Chris Long and Rich Tatum, have commented on the implications of these findings. The conclusion is clear: in AI-driven searches, new content can surpass factual content, raising concerns about fairness and the quality of information available online. For now, if your content is not new, it may be at risk of becoming invisible to users who rely on these search tools.