Generative-AI Evidence Ages Quickly: Audit of 40 Records Reveals Median Model Age of 281 Days
An arXiv paper recently examined 40 empirical studies on generative-AI evaluations released from July 2025 to July 2026. It revealed that the median age of the most recent model at the time of publication was 281 days. The analysis categorized aspects such as publication method, timing of execution, model identification, and supersession trends. The median age differed based on publication type: 395 days for journal articles, 56 days for preprints, and 49 days for lab reports. Each of the 40 studies featured an OpenAI system, although the authors clarify that this indicates design choices rather than frequency. Additionally, the paper presents a claim-currency framework and discusses a reflexive case involving frontier-model-assisted research.
Key facts
- Audit of 40 empirical records on generative-AI evaluations
- Records published between 18 July 2025 and 17 July 2026
- Median age of newest named model at appearance: 281 days
- Middle 50% range: 75-478 days; full range: 11-939 days
- Median age for journal articles: 395 days
- Median age for preprints: 56 days
- Median age for laboratory report: 49 days
- All 40 records included an OpenAI system
Entities
Institutions
- arXiv
- OpenAI