Blog · Research · 6 September 2026 · 9 min read

We Wrote About a -23,000 Payroll Print. BLS Just Revised It to +21,000.

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

On 9 August 2026 this site published a post about a payroll figure. The Bureau of Labor Statistics had reported that US nonfarm payrolls fell by 23,000 in July, and the argument we made was that the number could not bear the weight being put on it, because BLS's own confidence interval for the monthly change is on the order of plus or minus 122,000.

On Friday 4 September BLS published the AugustEmployment Situation, USDL-26-1435. In it: "the change for July was revised up by 44,000, from -23,000 to +21,000."

What this does and does not do to the earlier post

The earlier post's specific claim, that a -23,000 print sat inside a plus or minus 122,000 band and therefore could not be distinguished from zero, was correct then and is still correct now. A +21,000 print sits inside the same band and also cannot be distinguished from zero.

What has changed is more uncomfortable than a wrong argument. The argument held; the data under it moved 44,000 in seven weeks. Anyone who read that post and took away "payrolls fell in July" is now carrying a fact that BLS has retracted, and no amount of correct reasoning about error bands prevents that.

So: July 2026 payrolls, as currently published, rose by 21,000. June rose by 31,000, revised up from +20,000. If you saw a headline in early August about the first negative payroll month in a while, it was describing an estimate that has since been withdrawn.

Why the first print is the weakest one

BLS states the mechanism in the release itself: "Monthly revisions result from additional reports received from businesses and government agencies since the last published estimates and from the recalculation of seasonal factors."

ItsTechnical Noteis blunter about what that means for the freshest numbers: "estimates for the most recent 2 months are based on incomplete returns; for this reason, these estimates are labeled preliminary in the tables." And: "the prior 2 months are routinely revised to incorporate additional sample reports and recalculated seasonal adjustment factors."

The rule this suggests is not that the data is untrustworthy. It is narrower and more useful: a payroll number less than two months old is a draft. If a decision, a forecast or an argument turns on the sign of a single recent month, it is turning on the part of the series most likely to move.

There is a second 23,000 in this release, and it is a different number

The August release contains its own -23,000, and it is not the July one. BLS reports: "Information employment declined by 23,000 in August, following losses that had averaged 8,000 per month over the prior 12 months."

That is roughly three times the sector's own recent monthly average, and BLS breaks it down:

The information supersector covers software, data processing, publishing, broadcasting and telecoms, which is close to the list anyone would write if asked which industries are most exposed to generative AI. So the inference is available, and it was made in coverage within hours.

The release does not make it. A direct search of the text returns no instance of "artificial intelligence", no "AI", and no "automation". This is not an omission. The establishment survey asks employers how many people are on the payroll. It does not ask why, and BLS does not speculate in a statistical release.

What would have to be true for the information figure to mean what it is being read to mean

Three things, none of which this release establishes.

First, the decline would have to persist. One month at three times the trailing average, in a supersector estimated from a smaller sample than the total, is the shape of both a real inflection and ordinary noise. The 12-month average BLS publishes alongside it exists precisely because the single month is not self-supporting.

Second, the losses would have to be concentrated in the roles AI actually substitutes for. "Publishing industries -7,000" is compatible with editorial jobs replaced by generated text. It is equally compatible with an advertising downturn, with consolidation, and with print contraction that started long before any of this.

Third, there would have to be no competing explanation with better evidence. There are several. Interest rates, the post-2021 correction in tech headcount, offshoring, and the tail of a hiring freeze all sit in the same window and none of them is ruled out by anything in the release.

None of this means AI is not affecting employment in these industries. It means this release is not the evidence, and treating it as evidence is how a real effect ends up being argued from a number that cannot support it.

What we are not claiming

We are not claiming the July revision proves the labour market is stronger than it looked. A +21,000 print is inside the same error band that made -23,000 uninterpretable. The honest reading of July, before and after, is that the survey could not tell it apart from no change.

We are not claiming the information decline is noise. We are claiming that one month does not distinguish signal from noise, and that the release itself provides the trailing average that makes the comparison possible.

We are not claiming BLS made an error. Revision is the documented, intended behaviour of a survey that publishes before its returns are complete. The error, if there is one, belongs to anyone who treated a preliminary estimate as final, and that includes the earlier version of this site.

What would change this

The September Employment Situation, due in early October, will revise August and July again. If information posts a second consecutive month well below its trailing average, the case for a real inflection strengthens considerably. If August moves back toward the average on revision, the story was a sample artefact, and it will have been repeated for a month by then.

An annual benchmark revision, which reconciles the survey against unemployment insurance tax records covering nearly all employment, would settle the level question for the whole period in a way no monthly print can.

Until one of those lands, the defensible summary is short. July did not fall. August's information decline is real, large relative to its own recent history, and unexplained by the document that reports it.