AI Models Tricked by Fake Data to Increase Visibility
▼ Summary
– Waseda University researchers found AI models systematically favor content with newer publication dates over older material regardless of actual relevance or quality.
– Experiments showed all seven tested AI models (including GPT-4o and LLaMA-3) ranked fake-dated content higher, with some passages jumping up to 95 positions.
– This recency bias caused older authoritative sources like academic papers to lose visibility to newer, often less credible content across all models.
– Meta’s LLaMA-3-8B showed the strongest bias with 25% relevance reversals, while Alibaba’s Qwen-2.5-72B was least affected with only 8% reversals.
– The findings suggest publishers might artificially update dates to boost rankings, potentially creating a “temporal arms race” that rewards recency over quality.
Manipulating publication dates can significantly enhance how often online material appears in artificial intelligence search results, according to recent academic findings. A study from Waseda University reveals that simply assigning a more recent timestamp to existing content, without altering the substance, causes multiple major AI models to rank it higher. This suggests these systems heavily prioritize perceived freshness, sometimes at the expense of accuracy and authority.
The implications for content creators are substantial. High-quality, enduring articles may disappear from AI-generated answers if they carry older dates, even if their information remains correct and valuable. This creates pressure to frequently refresh content, whether through meaningful revisions or by artificially adjusting timestamps.
Researchers tested this phenomenon by taking text passages from established datasets and attaching fabricated publication dates to them. They then submitted these identical texts to seven prominent AI models, including GPT-4o, GPT-3.5, LLaMA-3, and Qwen-2.5, for ranking. The outcome was consistent: every single model demonstrated a clear preference for the content labeled with the newer date.
The data showed measurable impacts. On average, content entered the top ten search results when its date was shifted forward by one to five years. Some individual passages saw their ranking improve by as many as 95 positions. Perhaps most tellingly, one out of every four relevance judgments was reversed based purely on the date change, not the text itself.
This created a “seesaw effect” across all the AI systems evaluated. Newer-dated material consistently rose to the top ranks, while older content was systematically pushed down. Content in positions 1 through 10 was, on average, 0.8 to 4.8 years “fresher.” Conversely, content in the 61 to 100 ranking slots was often up to two years older. This bias even affected highly authoritative sources like peer-reviewed academic studies and medical journals, which lost visibility to more recent but potentially less reliable material.
The degree of this recency bias varied among the AI models tested. Meta’s LLaMA-3-8B exhibited the strongest bias, with a 25% reversal rate and an average freshness shift of nearly five years. On the other end of the spectrum, Alibaba’s Qwen-2.5-72B showed the least bias, with only 8% of decisions flipping and minimal year shifts. OpenAI’s models, GPT-4o and GPT-4, fell somewhere in the middle, displaying a noticeable but more moderate preference for recent dates.
This laboratory finding aligns with earlier discoveries. Earlier this year, an independent researcher named Metehan Yesilyurt identified a specific setting in ChatGPT’s configuration files labeled “usefreshnessscoring_profile: true.” This provided direct evidence that OpenAI’s reranking algorithm is explicitly designed to favor newer content.
Looking ahead, Yesilyurt warns of an emerging “temporal arms race.” Publishers might be tempted to boost their rankings by adding labels like “Updated for 2025” to old content. In response, AI systems will likely develop better methods for detecting these superficial edits. The core problem, however, is that the underlying bias could continue to reward recent publication over genuine quality and factual correctness.
Industry professionals have shared mixed reactions. Chris Long, a co-founder at Nectiv, noted on LinkedIn that updating content for freshness is a highly scalable tactic that has long helped with traditional search engine visibility, and it now clearly affects AI search results as well.
Rich Tatum, an SEO and AI solutions architect at Edgy Labs, offered a more nuanced perspective. He suggested that using freshness as a relevance signal is logical for large language models, and they may naively trust timestamps even on authoritative archives. He predicts the marketing profession will inevitably exploit these signals until future AI models are trained to be less gullible, at which point the signals will lose their power.
This prompted a strong rebuttal from Rand Fishkin, cofounder of SparkToro. He argued that he does not find it “sad” to exploit signals used by LLMs, drawing a parallel to his past opposition to spam tactics against Google. He stated that large tech companies and AI providers have not earned respect for their systems through ethical behavior and sees no reason for creators to adopt a set of arbitrary ethics to comply with their whims.
The fundamental takeaway is that in the current landscape of AI-powered search, being perceived as new often trumps being factual. For anyone relying on these systems for visibility, if your content isn’t labeled as recent, it risks becoming virtually invisible.
(Source: Search Engine Land)
