Your Credential Has a One-Year Shelf Life. AI Isn't Why.

A credential you earned 14 months ago adds almost nothing to a prediction of what you can do today. That is not a claim about AI. It is what 444,698 competition entries on Kaggle show, across a window that starts six years before ChatGPT existed.

The finding comes from "Stranded credentials: how a skill-signaling market absorbed generative AI," a preprint posted to arXiv on August 17, 2026 by Song Yao. [Fact] It has not been peer reviewed, and I will come back to what that means for how much weight to put on it. But it is the first audit of a skill-signaling market large enough, and old enough, to separate two things that nearly every discussion of AI and credentials mashes together: whether AI degraded the signal, and whether the signal was degrading anyway.

The answer is mostly the second one.

The 82% number, and the trap inside it

Here is the headline that would make a good scary post. Kaggle ran two competition formats side by side. Upload-competitions score the predictions you submit, computed on data you already hold. Code-competitions score predictions by executing your code against data you never see. The stock of medals earned in upload-competitions lost 82% of its informativeness — 82% of its power to predict how the holder would later perform. [Fact]

Read that alone and the story writes itself: generative AI can produce a good prediction file, so the format that scored prediction files stopped meaning anything.

That story is wrong, and the paper's own data is what kills it. Between half and three quarters of the 82% loss is explained by what Yao calls institutional stranding. [Fact] Upload-competitions had already exited for reasons that predate generative AI. Once the format stopped running, nobody could earn a fresh upload medal. The existing stock simply aged out — under the same decay pattern that was already visible before AI arrived.

Do the arithmetic the abstract leaves implicit. If institutional stranding accounts for 50-75% of the loss, everything else combined — AI, shifting skill mixes, whatever else you want to blame — accounts for the remaining 25-50%. Against an 82% total, that leaves roughly 21 to 41 percentage points of lost informativeness for all non-institutional causes together. [Estimate] The scary number and the AI-attributable number are not the same number, and they are not close.

Credentials were always perishable

The deeper result is the one that should change how you think about your own record. Across both formats, medals predict later leaderboard performance almost entirely within the first year after they are earned. [Fact] Not mostly. Almost entirely.

That is a shelf life, and it existed before AI.

It also produces the paper's quietest, most damaging finding. Kaggle's official credential tiers — Grandmaster, Master, and the rest — are built on lifetime medal counts. A lifetime cumulative count is close to the worst possible way to summarize a signal whose value collapses after twelve months, because it weights a medal from 2015 identically to one from last quarter. The audit puts a number on the waste: the official tiers discard 13-16% of the information the underlying medals actually carry. [Fact] An index that simply weights recent medals more heavily — built on pre-AI-era data alone, with no knowledge of what was coming — beat the official tiers at predicting AI-era performance. [Fact]

The fix was available before the disruption. Nobody applied it.

The evidence that AI is not the culprit

The cleanest test in the paper is the comparison the two formats make possible. If generative AI were corroding these credentials by letting people fake competence, an AI-like working style should pay off differently depending on whether the format can be gamed with a good prediction file. It does not. An AI-like working style predicts performance similarly in both formats. [Fact] What Yao measures is institutional rather than personal — a property of which venues stayed open, not of who was using which tools.

There is a second, subtler correction. Old upload-competition medals look more valuable when you examine them in isolation. They are not. They are proxying for the rest of the holder's record — experience, mostly. [Fact] Strip that out and the apparent value goes with it. Anyone appraising a long-dormant credential in a vacuum is reading experience and calling it skill.

What this means if you work with data

For the occupations closest to this evidence, our tracker puts 2026 AI exposure for data scientists at 70% with automation risk at 43%, and machine learning engineers at 73% exposure and 45% risk. [Estimate] High exposure, moderate displacement risk — which is precisely the profile this paper describes. The work is being reshaped rather than deleted, and the signaling problem is about proving you can still do the reshaped version.

Three things follow.

Replenish instead of accumulate. If a credential's predictive value is concentrated in its first year, then "collect certifications, list them forever" is buying an asset that expires and never marking it down. One recent, verified piece of work beats four old ones.

Prefer formats that hide the test data. Code-competitions held up. That principle generalizes past Kaggle: a take-home project on a public dataset proves less than live problem-solving on inputs you did not prepare for. If you are choosing what to put in front of a hiring manager, choose the thing that could not have been prepared in advance.

Watch for stranded venues. The most damaging thing that happened to these medal holders was not AI. It was that the room where they earned their reputation closed. If the platform, community, or certification body you invest in shuts down, your record freezes and begins aging immediately, however good it was on the day it froze.

Where this evidence stops

The limits here are real and I would rather state them than bury them. This measures how well a credential predicts later performance on the same platform — leaderboard outcomes, not salaries, not offers, not promotions. Kaggle competitors are not a random sample of people who work with data. And this is a preprint: posted to arXiv on August 17, 2026, not yet through peer review, so the identification strategy has not been checked by outside referees. The one-year decay and the 82% figure are exactly the kind of results that can move under scrutiny.

The obvious objection is that Kaggle is a game, and games are not jobs. Fair. But it is a game with 444,698 recorded attempts, timestamped outcomes, and two scoring regimes running in parallel — which is more self-measurement than almost any professional credentialing system has ever produced. Most certification bodies cannot tell you whether their certificate predicts anything at all.

That is the uncomfortable part, and also the hopeful one. The finding is not that AI broke credentials. It is that credentials decay on a clock nobody was watching, and the institutions issuing them were discarding 13-16% of their own signal before AI showed up. A decay problem you can see is a problem you can plan around: earn recent, earn in formats that cannot be pre-cooked, and do not assume the room will still be there in five years.

Sources

• Song Yao, "Stranded credentials: how a skill-signaling market absorbed generative AI," arXiv:2608.17111 (econ.GN; cs.CY), submitted August 17, 2026. https://arxiv.org/abs/2608.17111 — preprint, not peer reviewed.
• 2026 AI exposure and automation risk figures for individual occupations are AI Changing Work model estimates, not from the paper above.

AI-assisted analysis: this article was drafted with AI assistance using the primary source listed above, and reviewed before publication. Every figure attributed to the arXiv preprint was checked against the paper's abstract as posted.