New Decade, Steady Scores: Our Q3 2026 Refresh

Q3 2026 update · August 30, 2026 · 8 min read · CareerVillage AIR Team

Every quarter, we re-run the AI Resilience Model from the ground up. Every task for every occupation is re-scored, and every occupation's written analysis is re-drafted using current evidence, so that what you read on a career page reflects AI's impact on the labor market as it stands this quarter.

This quarter's refresh replaced more of the model's inputs than usual. We rebuilt the entire task layer from O*NET's newest task catalog release and adopted the Bureau of Labor Statistics' new projections for the 2025–2035 decade. We also shipped a fix for a long-standing citation quality problem, and removed a group of occupations we could never score honestly.

After all of that, the average score moved 2 points and 87% of occupations kept their label. If that sounds familiar, it should: last quarter's refresh followed the same pattern, where a major input rebuild produced only modest topline movement.

Below, we explain what changed and what impacts those changes have.

A fresher picture of what each job involves

Every occupation's score starts from its task list: the specific things a person in that job does all day. Those lists come from O*NET, the Department of Labor's occupational database, and this quarter we rebuilt the entire task layer from the current release, O*NET 31.0 (August 2026).

The task catalog grew about 3%, to roughly 19,000 tasks, and 21 occupations got task data for the first time, including data scientists, penetration testers, information security engineers, and cardiologists. Until now, those occupations had been scored on occupational-level research evidence alone.

With the new tasks in place, deep research re-ran for every detailed occupation in the report, so every analysis paragraph on the site is grounded in this quarter's evidence.

The new BLS employment projections

On August 27, the Bureau of Labor Statistics published its 2025–2035 employment projections. These projections feed into our Long-term Employer Demand (LTED) sub-score.

The headline is slower growth. BLS projects that the economy adds 5.9 million jobs over the decade (+3.5%), down sharply from 10.9% growth over 2015–2025, with healthcare and social assistance accounting for about 37% of all new jobs and nurse practitioners the fastest-growing occupation at +41%.

What's new in this cycle is how directly BLS handles AI impacts. Four patterns stand out:

1. Computer and mathematical employment goes up. BLS projects growth for the occupations, like software developers, that are building and running AI. More on this one below, because our own sources are split on it.
2. The ‘picks-and-shovels’ build-out is real. BLS projects increased growth in careers connected to data centers, electricity production, and IT infrastructure.
3. Arts, design, entertainment, sports, and media decline, a drop BLS ties directly to generative AI absorbing creative and media production work.
4. The biggest hit is office and administrative support: −752,100 jobs (−4.0%), the largest decline of any major group, which BLS attributes to the continuing integration of automation tools, including AI-powered systems. Word processors and typists (−34.4%), telephone operators (−27.6%), and data entry keyers (−25.5%) top the declining list.

Ultimately, our LTED sub-score moved an average of 3.2 points per occupation with the new BLS employment projections. The shift in topline AI resilience scores due to new employment projections was even more modest–only about a point shift, on average.

This quarter's model improvement: citation enrichment

This quarter we also made progress on a quality problem we first described in our Q2 update. The specific failure is what we call a ghost link: the research model writes “according to a McKinsey analysis…” and has the McKinsey URL sitting right there in its research trail, but never connects the source to the analysis. The claim looks sourced but is not clickable, so our citation parser drops it. Roughly 21% of career pages had at least one ghost link.

The fix we landed on is a citation finalizer: one extra model pass at the end of deep research that links every named source to a URL already gathered during research, with a deterministic repair pass as backstop. We built the finalizer's instructions using GEPA, a prompt-evolution technique that tests candidate instructions against a real evaluation set and keeps what measurably works.

One policy governs this entire pipeline: link or leave. If we cannot verify a URL for a named source, the sentence stays exactly as written and gets flagged for review. We don't reword a claim to hide a missing source.

The new process has resulted in a dramatic decrease in ghost links: now just 3% of career pages contain them. The average number of linked citations per career page also went up.

Military occupations left the dataset

We also removed the 26 Military Specific Occupations (SOC major group 55). We should have done this earlier. O*NET does not collect task data for them, the BLS wage survey does not cover them, and all of our external AI-exposure sources exclude them. Their pages were effectively empty shells with a score attached. The report now covers 1,571 occupation entries, and removing the military group moved the all-careers average score by less than a tenth of a point.

What actually moved

With new tasks, fresh research, the new BLS decade, and the scope cleanup all landing in one cycle, here is how much the topline AI Resilience Score changed, measured across the 997 detailed occupations present in both:

Of the 132 occupations that changed labels, 78 moved up a tier and 54 moved down — there was no broad shift in either direction.

What this means if you're using the report

If you checked an occupation last quarter, odds are close to 9 in 10 that its label is the same today. What did change:

• The written analysis on every career page is completely new, grounded in this quarter's task data and evidence.
• Sources named in an analysis are now almost always clickable links, so you can follow any claim back to where it came from.
• 21 occupations, including Data Scientists and Cardiologists, are now scored on real task lists for the first time.
• Military occupation pages no longer appear in the report.

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.