Original title: Stranded credentials: how a skill-signaling market absorbed generative AI
Song Yao · August 2026
Generative AI can now solve tasks that people use to prove their skills, raising doubts about whether old credentials still signal true ability. Researchers studied Kaggle, a major data science competition platform, to see if competition medals still predict future success. Kaggle ran two competition formats: one where participants simply uploaded predictions, and another where the platform ran participants' code on hidden data. The researchers found that medals still predict performance, but their value decays fast. Medals predict future leaderboard ranks almost entirely in the first year after someone earns them, across both formats. Fresh medals kept most of their predictive value during the rise of AI. While the total stock of upload medals lost 82% of its informativeness, this mostly happened because the platform stopped hosting upload contests for reasons predating AI, letting those credentials age out. An AI-like working style predicts performance similarly across both formats.
Imagine you collect stamps to show you are a good traveler. Each stamp proves you visited a place. But if a country stops issuing new stamps, your old stamp still looks impressive—until people realize it's from a place that no longer exists. Here, medals are like stamps: they prove skill, but only if the competition still runs. Once the competition ends, the medal becomes a relic, not a measure of current ability.
For employers hiring based on credentials, this means a medal from two years ago is nearly worthless. For workers, it means credentials must be continuously updated to signal ability. For credentialing bodies, it means the value of their certifications depends on how quickly they replace outdated assessments.
The abstract doesn't explicitly state limitations, but it focuses on Kaggle's two competition formats and does not generalize beyond that. The institutional stranding explanation accounts for half to three-quarters of the loss, leaving some variance unexplained. The proposed index was built on pre-AI data and its long-term validity is untested.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Stranded credentials: how a skill-signaling market absorbed generative AI — Song Yao
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.