Michelle Yin

I care about this because I have sat across the table from workers whose career decisions depend on what researchers like me put into the world. If those numbers are not credible, we are failing the people we are supposed to serve.

AI companies are profit-driven and their models reflect their training data, and their design choices. Researchers who use these proprietary outputs as scientific instruments have an obligation to verify what they produce before families, communities, and governments act on it.

When I discovered the instability in AI exposure scores, I was working with workforce programs in Maine and Virginia trying to help real people navigate a changing labor market, and I realized the numbers we were relying on gave fundamentally different answers!

I want to share why this research is personal to me, not just professional. I came to this country as an immigrant and built my career as a labor economist because I believe how we measure work shapes how we value workers.

Then we asked: does this matter for the conclusions economists are actually drawing? We plugged each model’s scores into a standard labor economics analysis. With one model’s scores: significant job losses. With another’s: no detectable effect. The entire finding flipped.

Here is every occupation, all 95, sorted by how much the four models disagree. Top of the chart: 87 percentage points of disagreement for a single occupation. One AI sees the job as almost fully exposed. Another sees it as barely exposed. Find your job.

Then we asked: does this matter for the conclusions economists are actually drawing? We plugged each model’s scores into a standard labor economics analysis. With one model’s scores: significant job losses. With another’s: no detectable effect. The entire finding flipped.

Then we asked: does this matter for the conclusions economists are actually drawing? We plugged each model’s scores into a standard labor economics analysis. With one model’s scores: significant job losses. With another’s: no detectable effect. The entire finding flipped.

We replicated the most widely used AI exposure rubric (Eloundou et al. 2024) with four frontier models: GPT-4, ChatGPT-5, Gemini 2.5, and Claude 4.5. Same instructions. Same O*NET task data. Same pipeline. Mean exposure ranged from 14% to 51%. A 3.6x gap on identical jobs.