Introduction

AI search now rewards “new” over “true.” Even perfect old articles lose to recent ones because of a built-in freshness bias.

What’s the problem it solves?

The article explains a new ranking bias found inside ChatGPT and other AI models. It shows that AI systems automatically push newer content higher, even when older material is better or more accurate.

Quick Summary

In August 2025, Metehan Yesilyurt discovered a single line of code in ChatGPT’s setup: use_freshness_scoring_profile: true. It meant that the system favors newer content. A few weeks later, scientists at Waseda University confirmed this bias through lab tests across seven large AI models, including GPT-4o, GPT-3.5, and LLaMA-3. They added fake publication dates to identical texts and saw that all models ranked the “newer” versions higher.

The results were dramatic. Rankings shifted by up to 5 years, some content jumped 95 places, and 1 in 4 relevance decisions flipped based only on date. This shows that freshness isn’t just a small factor-it can completely override quality. Models like Qwen2.5-72B were less biased, while others like LLaMA-3 showed the strongest bias.

For creators, this means older, high-quality guides or research are being buried by newer posts. The only reliable way to stay visible in AI search now is to update content often. The article calls this the start of a “temporal arms race,” where constant updates will matter more than timeless expertise until AI systems learn to adjust for context.

Key Takeaways

• ChatGPT and most AI models prefer newer content, regardless of quality.
• Rankings can move 1-5 years forward simply because of timestamps.
• About 25% of AI relevance choices flip when dates change.
• Bigger models aren’t always better; Qwen2.5-72B was the least biased.
• Recency bias can distort results in science, medicine, and education.
• AI search systems treat freshness as always-on, without considering topic type.

What to do

• Update often - Refresh all 2022 or older content now.
• Add date signals - Use “Updated for 2025” and note when data is current.
• Show timelessness - Mark long-term guides as “Foundational” or “Evergreen.”
• Measure your bias - Compare rankings before and after updates.
• Pick smarter models - Tools using GPT-4o or Qwen2.5 handle bias better.
• Plan regular reviews - Treat freshness as part of your ongoing SEO routine.

‍