The AI Salary Premium That Dissolved

We published a number. Then we made the data honest and the number went away.

On June 12, 2026 this post opened with two figures. The 687 AI-classified roles in our dataset carried a $168,000 median base. The other 2,758 carried $154,000. A $14,000 gap. A 9% premium.

Those numbers were correct. Run our method against the dataset as it stood that day and you get $168,000 and $154,000 to the dollar.

They are also gone.

On July 16 we re-anchored every US base salary in the dataset to its own federal occupation wage curve, placed by seniority tier, reproducible from the Bureau of Labor Statistics OEWS file. That was an improvement. Before it, a role's number came from our model of what the role pays. After it, the number traces to a federal wage survey.

The premium did not survive the improvement. Today the same 687 AI roles carry a $133,000 median and the same 2,758 non-AI roles carry $156,000.

The honest reading is not that AI roles started paying less. It is that our old numbers were measuring our own assumptions about AI titles, and the moment they stopped, the signal went with them.

Why the premium was never measurable at title level.

Here is the structural fact that explains it, and it is the thing worth taking away from this post.

Our 687 AI-classified roles map to 59 federal occupation codes. Fifty three of those 59 are also used by roles we do not classify as AI. 679 of the 687 AI roles, 99% of them, share an occupation code with a non-AI role.

So once every salary is anchored to an occupation code, comparing AI titles to non-AI titles compares two overlapping samples drawn from mostly the same wage curves. Whatever gap you compute is telling you which occupation codes each group happens to draw from, and in what seniority mix. It is not telling you what AI pays.

The compression is visible in the numbers. Our 687 AI titles are served by 79 distinct medians. All 3,445 roles are served by 202. Rank the 13 AI hubs by median pay today and you get 6 distinct values across 13 rows. Rank the top AI titles and four different vice president titles tie at exactly $303,000.

A ranking where a third of the rows are ties is not a ranking. It is a wage curve with job titles written next to it.

What we got wrong, and what to do about it.

I published the 9% figure and I would publish it again with the data I had. That is not a defence. A number computed correctly from a modelled dataset inherits the model, and I described it as a measurement of the market when it was a measurement of our own estimates. The dataset told me what I had told it.

If you are quoting an AI salary premium from anyone, including us, ask one question. Is the underlying number anchored to something outside the vendor, and if it is, do the AI and non-AI groups share the anchor? If they share it, the premium is an artifact of grouping.

What still holds: location, and it is the biggest lever in the data.

One finding survived the re-anchoring unchanged, because it never depended on the modelled layer.

San Jose, CA carries the highest cost-of-labor multiplier across all 81 cities in our data: 1.38. Charleston, WV carries the lowest: 0.78. San Jose pays 77% more than Charleston for the same role.

Worked example. A role with a $150,000 national median prices at $207,000 in San Jose and $117,000 in Charleston. Same title, same band, $90,000 apart.

The distribution matters more than the extremes. Of the 81 cities, 34 sit above the 1.0 national baseline, 1 sits exactly at it, and 46 sit below. Most tracked cities pay below the national number, so a national median quoted without a city is an overestimate for most of the map.

Location is the single biggest lever in this dataset that has nothing to do with your skills.

To price any of the 3,445 roles against any of the 81 cities, run the salary calculator.

AI does not own the top bracket, and now it is under-represented in it.

743 of the 3,445 roles, 22%, carry a median of $200,000 or more. Of those 743, 88 are AI-classified. That is 12% of the top bracket against AI's 19.9% share of the dataset.

Under-represented, not over. The June version of this post ran the same calculation on the modelled data and found AI mildly over-represented at 22%. The direction flipped for the same reason everything else did.

The reason underneath is unchanged and boring. The non-AI tail includes medicine, law, and executive roles, and those clear $200,000 on their own occupation curves.

The current distribution, for anyone who needs it: median $136,000, 10th percentile $101,000, 90th percentile $219,000, range $34,000 to $333,000, mean $156,139. The mean sits above the median, so the distribution is right-skewed. Quote medians.

Where these numbers come from.

Every figure here comes from our published salary dataset: 3,445 roles across 81 US cities, each role's base median anchored to its federal occupation wage curve by seniority tier and reproducible from the BLS OEWS file. The AI classification comes from our open role taxonomy, 687 roles across 13 hubs, CC BY 4.0.

The June figures are quoted from the dataset as it stood on June 12, 2026, which is why they no longer reproduce against today's data. The current figures were computed on August 29, 2026.

The limitations, and this post is mostly about one of them:

• Base salary only. No equity, bonus, or signing.
• US only.
• Medians, not individual offers.
• Titles collapse onto occupation codes. 3,445 roles resolve to 202 distinct medians, so many titles share a number with a title that is not much like them. Any comparison between groups of titles is affected by this, which is the finding above.
• AI classification is structural: taxonomy hub assignment, not job-description text.
• City figures are multiplier-based, and live role-by-city pages apply a deterministic per-page variance of up to plus or minus 4%.

If you cite this, cite the correction along with the number. The taxonomy is CC BY 4.0. Attribution is the whole ask.