Opens in a new tab
needsahuman
Search ⌘KGet the data

Editorial policy

This page sets out the rules every page on NeedsAHuman follows. They exist so you can trust a page without having to trust us.

This page sets out the rules every page on NeedsAHuman follows. They exist so you can trust a page without having to trust us.

What we publish

We publish scores and short explanations for about 1,000 US occupations, with a UK section on each page. We also publish rankings, curated lists, job family pages, guides and an open dataset. Every score page is built from the same public method.

The rules for numbers

  1. Every number has a source and a date. Each section of a job page ends with a source line: the dataset, its version and when it was published.
  2. No estimate without its range. A replacement year is always shown with its 80% range. A quality score is always shown with its evidence grade. If we cannot give a range, we do not give the number.
  3. No invented statistics. Written summaries may only use numbers that are in that job’s data. Our automated checks reject any draft that contains a number not in the data.
  4. Imputed values are never shown as a job’s own result. When a job has no direct evidence, the model sometimes needs a stand-in value from the wider job family or a neutral default. We use it inside the calculation, mark it as imputed in the dataset, and never display it as if it had been measured for that job.
  5. Thin evidence is said out loud. If there is no direct test of AI against people in a job, the page says “Not yet measured” and explains what evidence would settle it.

The words we use, and the ones we don’t

  • We never say a job “will be eliminated”. The most we say is that a job “could be largely automated”, and always with a dated range.
  • Replacement years past 2060 are written as “Not foreseeable before 2060”. We do not guess further out.
  • The verdict words are fixed: Largely (0–19), Mostly (20–39), Partly (40–59), A little (60–79) and Nah (80–100). “Nah.” is only ever the verdict for the top band.
  • We avoid fear language, countdown clocks and robot imagery. We also avoid false comfort: “A little” still means AI can do a real share of the work.
  • Card labels are plain questions. Technical names such as QPI and ERY appear only in the methodology.

Sources we use, and how we rank them

We prefer, in order:

  1. Official statistics (US Bureau of Labor Statistics, Census Bureau, UK Office for National Statistics).
  2. Open research datasets with a clear licence (O*NET, the Anthropic Economic Index, Microsoft’s Working with AI data).
  3. Peer-reviewed or openly published research with a described method.
  4. Benchmarks that compare AI with qualified people on real work.
  5. Company and vendor reports, which we label as such and weight lower in quality scores.

For quality evidence we weight independent studies above academic studies above vendor studies, and newer results above older ones. The quality method page explains the weights.

We do not use Glassdoor, Indeed reviews, LinkedIn or Blind content, and we do not scrape sites whose terms forbid it.

How AI is used on this site

  • Scores come from a data pipeline. No chatbot decides a score.
  • Task ratings. Part of the “Can AI do it?” score uses a language model to rate each O*NET task against a published rubric. The rubric, the model name and the date of the run are published on the coverage method page.
  • Written summaries on job pages are drafted by software from that job’s own data. Automated checks reject drafts that are missing the verdict, use a number not in the data, state a single date without a range, or say a job will be eliminated.
  • Human checks. Each data release, a person reads a sample of job pages, weighted toward the most visited, against the data. Guides, lists, methodology and reports may be drafted with AI help, but a person edits them and checks every fact against its source before they go live.
  • Regeneration. We only rewrite a job’s summary when one of its scores moves past a set threshold, so pages do not churn every quarter for no reason.

Independence

No commercial relationship ever changes a score. Affiliate links are labeled where they appear and are confined to the tools module and a labeled block in “How to stay needed”. Nothing commercial appears above the evidence sections of a job page. Read more on the About page.

Corrections

If we get something wrong, we fix it, say what changed and log it with a date on the corrections page. Score changes from new data are not corrections; they are listed in the data changelog. Method changes get a new score version and a note on the methodology page.

Updates

  • Benchmarks and AI tool facts are checked monthly.
  • The Anthropic Economic Index and O*NET are updated quarterly.
  • Employment, pay and projections are updated yearly when the agencies publish.

Each job page shows the date it was last scored and the score version it uses.

Reporting a problem

Use the contact page or the “Report an issue” link at the foot of any job page. Tell us the page, what looks wrong and, if you can, a source. We read every report.