Sharper Research, Steadier Scores: Our Q2 2026 Research Refresh

Q2 2026 update · 9 min read

Three months ago, we published our last quarterly update and since then, new AI tools have launched, new research has come out, and the way people use AI at work has kept shifting. That pace is exactly why we built the AI Resilience Report as a living model that we rerun every quarter, not a one-time study.

This quarter, we didn't just refresh the data. We rebuilt a piece of the pipeline that does some of the heaviest lifting: the deep-research engine that reads current sources about every occupation and decides how exposed each one is to AI. That research feeds the “AI and automation exposure” dimension of our model. (For a refresher on how exposure fits into the overall score, see our previous post on the scoring methodology.)

The short version: the research got better, and the scores for most occupations barely budged. That sounds boring, but it's actually the result we wanted. Here's why.

What we changed under the hood

Our internal AI exposure score (the “AI Resilience Model”) is derived from a process that reads dozens of articles, reports, and industry publications for every occupation, then synthesizes what it finds into a short, sourced summary of how AI is affecting that work. That summary is what determines the AI Resilience Model score.

After some preliminary testing, we shifted to using Anthropic's Claude Opus 4.7 for the process this quarter, made some technical improvements, and rewrote the prompt that guides the research. Four improvements stand out:

1. Sources from inside the profession, not just tech media

The biggest change: before the research starts on any given career, a smaller helper model now drafts a list of the trade associations, professional societies, and career-specific publications that exist to serve people in that field. For dentists, that means the American Dental Association and the Journal of the American Dental Association. For petroleum engineers, the Society of Petroleum Engineers. For marketing managers, the American Marketing Association. And so on.

The main research model is then required to cite at least one source from that list and to search those websites directly. The original prompt just said “prioritize industry associations” without naming any, so in practice the research often defaulted to generic tech blogs. Now every occupation's analysis includes a voice from inside the profession in addition to other sources like news and consultant reports.

2. Stricter rules about how recent sources have to be

AI changes month to month, not year to year. A blog post from 2023 can be out of date in ways that matter. The new pipeline enforces a recency ladder: try sources from the last 3 months first, fall back to the last 12 months only if needed, and never cite anything older than 24 months. For occupations most exposed to AI right now, the rules are even stricter: only the last 3 months count.

Trade associations get a small exception: their reports come out quarterly or annually, so a 6-month-old report from the American Dental Association is more valuable than a 1-week-old tech blog post. The pipeline now handles that distinction automatically.

3. Citations that are actually citations

A surprisingly common failure mode in AI research is what we call “lazy citation”: the model writes prose that mentions a source by name without actually linking to it, or sticks a list of sources at the bottom that doesn't connect to specific claims. We made real progress on this. The new prompt asks for every named source to be wrapped in a real, clickable link in the body of the analysis, and an automated check after the research completes flags analyses where those inline citations are missing. Lazy citations do still slip through sometimes, but far less often than before. We'll continue to invest effort in ensuring accurate, linked source citation.

4. A full audit trail for every career

We now record every search query the research model ran, every page it pulled, and every source it actually used. That means we can answer the question “did the model actually look at the trade journal?” without guessing. For learners, this is invisible. For us, it means we can run audits on the research process and verify it's behaving the way we expect.

What the numbers say

We rebuild research from scratch each quarter, so by design every occupation's underlying analysis is brand new. The question is what that means for the score you see on the site.

For most occupations, the answer is: very little.

Across the 1,016 detailed occupations in our report, 79% landed in the exact same resilience label as last quarter. About 10% moved up to a more resilient label, and about 11% moved down. The label distribution barely shifted: every tier's share of careers changed by less than 2.5 percentage points.

That stability is intentional. The top-line resilience score is built from multiple sources averaged together. When one input changes (in this case, ours, a lot), the others act as a stabilizer. The published score is the rolling consensus across sources, not any single one.

You can see that in the size of the average movement:

Our internal CV deep-research score moved an average of 22 percentile points per occupation. The top-line resilience score — the one shown on every career page — moved an average of less than 4pp. Six times less volatility, even though one of the major inputs was completely rewritten.

The new research is more granular

Even though the top-line score was stable, the new deep-research pipeline did produce noticeably better-shaped data underneath. The old model tended to cluster careers near the middle of the scale — when in doubt, give it a 0.5. The new model spreads careers out more, separating “genuinely high-exposure” from “moderately high” from “moderately low”.

The 0.7–0.8 score range expanded from 13 careers to 129 — a sign that the new model is willing to call careers as clearly higher-exposure rather than hedging them toward the middle.

This is the kind of improvement we hoped to see. The new research pipeline isn't just rewriting the same opinions; it's making sharper distinctions, supported by more recent and more career-specific evidence.

The occupations that moved

So which occupations actually shifted? The biggest movers are a mix of stories: some are corrections (the old research was off, and the new research caught it), some reflect real-world changes in the last few months, and some show the new model drawing distinctions the old one missed.

A few patterns stand out. Healthcare shows up on both lists in interesting ways. Lab technologists and technicians climbed into more resilient territory, while specialist dentists, life scientists, and emergency medicine physicians moved the other way.

Skilled manufacturing careers, like foundry workers, metal fabricators, and welding machine setters, generally moved down in resilience, in our research. The new sources reflect a faster pace of automation in industrial settings than the old research had picked up. Whether that fully shows up in the published score depends on how the other AI exposure datasets evolve in the next quarter.

What this means if you're using the report

If you checked your career's resilience score last quarter, the chances are very high it's in the same tier today.

What did change, even for careers whose label stayed the same:

• The written analysis on every career page is brand new, with sources from the last few months instead of the last year or two.
• Citations now point to trade publications and professional bodies in the field: the kind of sources a careers counselor would actually recommend, not generic tech news.

What's next

We'll run the next quarterly update in late summer. Between now and then, we're focused on two things: bringing in additional AI exposure research as it's published (the field is producing new studies faster than ever), and improving how the report communicates why a career has the score it has, not just what the score is.

If you're using the report with students, in advising work, or for your own career thinking, we want to hear what's working and what isn't. Drop us a line at air@careervillage.org.