Career Atlas mixes published government data with numbers we model ourselves, and in a few places with our own judgement. This page says which is which for every number on the site, how each one is built, and where it can be wrong. We would rather be transparent than confident.
Every figure carries one of three provenance labels:
| Label | What it means |
|---|---|
| Sourced | Published data, used as-is |
| Modeled on sourced data | Computed by us from published data |
| Our judgement | Our opinion, with no external dataset to check it against |
The distinction that matters most is the last one. A number can be carefully built and still be an opinion, because no published dataset exists to check it against. Where that is true we say so on the number itself, not only here.
| Metric | What it measures | Provenance | Externally checked? |
|---|---|---|---|
| AI exposure | How much of the work is the kind of thing AI is used for | Modeled on sourced data | Yes, against Microsoft and Anthropic |
| Displacement risk | How much of that work AI may do instead of the person | Our judgement, over sourced inputs | No dataset exists to check it |
| Augmentation | How much of that work AI may do alongside the person | Our judgement, over sourced inputs | No dataset exists to check it |
| Legal barriers | Whether the law requires a particular human to do the work | Mostly derived from entry routes; 85 curated exceptions | Partly, by definition |
| Future-proof score | A weighted blend of growth, displacement, salary, skills, retention | Our judgement, over sourced inputs | No |
| Pay, outlook, openings | What the job pays and where it is heading | Sourced | Published by BLS |
| Skills, work context | What the work involves day to day | Sourced | Published by O*NET |
| Difficulty | How hard the role is to enter and to do | Our judgement | No |
A percentile rank from 0 to 100, grounded in real AI usage. We start from each occupation's O*NET task list and weight every task by its measured share of actual Claude conversations (the Anthropic Economic Index), counting core tasks more heavily than occasional ones. That signal is nudged by two work-context cues, how digital and desk-based the role is and how little hands-on physical work it involves, then ranked across all occupations. Roles with no usage data for any of their tasks fall back to a work-context and sector blend.
It is a rank, not a proportion. "93rd percentile" means the role is more exposed than 93% of occupations. It does not mean 93% of the job can be automated. The underlying distribution is uniform by construction, so ranks exaggerate differences at the extremes: two roles a fraction apart in raw score can sit ten percentile points apart.
Exposure is not elimination. A high score means more of the day-to-day work could change. In Anthropic's analysis of real AI usage, about 57% of use augmented people versus 43% that automated a task. That figure is a single global number covering all occupations. It cannot be resolved to an individual job, and we do not use it to split any job's score.
Exposure says how much of a job's work AI touches. It says nothing about whether a human is still required, which is why licensed, hands-on and relational roles can score high on it and yet be poor candidates for replacement. These two numbers add the missing term:
shield = 0.6 × gating + 0.4 × physical demand (capped at 0.85) displacement = task score × (1 − shield) AI does it instead of you augmentation = task score × shield AI does it with you
The two always add up to the task score, so they read as one sentence: of the part of this job AI can touch, this much is likely to be done with you and this much instead of you. Bands are Very high 0.60 and above · High 0.42–0.60 · Moderate 0.25–0.42 · Low 0.12–0.25 · Very low below 0.12.
The shield is our judgement. Its two inputs are defensible, legal barriers and physical demand from O*NET, but the weights are ours, and there is no ground-truth dataset of AI displacement anywhere to fit or check them against. The published benchmarks measure exposure, not displacement, so testing our discount against them would be circular. We validate what can be validated: the task score against published benchmarks, internal consistency, and a fixed panel of roles where the answer is not seriously contested.
The closest substitute we have found is BLS revising its own ten-year projections between cycles, which our model cannot see because BLS growth is not one of its inputs. We ran that test and published the result: displacement risk explains about 2% of the variance in those revisions, and of the occupations we score highest, BLS revised more of them up than down. It is a null, and it is reported as one. It does not vindicate the shield and it does not refute it, because the revision is not a measurement of AI either.
We looked for outside support and did not find it. The nearest available check is BLS Employment Projections: if legal barriers protect work, more gated occupations might be expected to hold employment better. They do not. The shield correlates with projected employment change at 0.09 across 960 occupations, and at -0.03 once healthcare is removed. Mean growth does rise across the barrier tiers, but only because healthcare occupations go from 3% of the ungated tier to 51% of the most gated one, and healthcare grows for demographic reasons that have nothing to do with AI. This is not evidence the shield is wrong, since BLS projections cover employment change from every cause and do not model AI at all, so no result there could confirm it either. It is a reason to treat the shield as the opinion this page already says it is.
Physical demand is used and remote-friendliness is not, because the two correlate at -0.96 across the 453 occupations carrying both readings; including both would count the same signal twice. Legal barriers are near-independent of physical demand, correlating 0.00 with it, which is what earns them a separate term. The 0.85 cap means no occupation is ever shown as immune, not because we know every job carries residual risk, but because we will not claim otherwise.
The shield models legal and physical barriers only. Where the law names a person, whether by licence, signature or fiduciary duty, it is on solid ground. Where a role is protected by organisational authority or relationship capital and nothing else, it has no term at all, and scores that role as exposed.
Chief executives used to be the clearest example. They are not any more, because Sarbanes-Oxley turns out to name them directly, and that is a legal barrier rather than a missing one. What remains are the roles with real influence and no legal duty whatsoever: chief marketing officers, chiefs of staff, technical program managers. They still score as highly exposed, and the model has nothing to say in their defence. We have chosen to leave them that way rather than shield seniority as a category, so read a high score on that kind of role knowing the model is blind here.
We tried twice to close this with O*NET's measures of decision-making and failed both times. One measures how much freedom someone feels they have, which is not authority: data entry keyers score 4.52 on it, above financial managers at 4.34 and registered nurses at 4.23. The other measures the impact of a person's decisions, and puts 101 occupations above chief executives, electricians and public safety dispatchers among them. Both are recorded in the code so a third attempt starts somewhere new. The underlying problem is that O*NET describes how a job feels to the person doing it, and authority is a fact about the organisation around them.
Whether the law requires a particular human to do the work. That is a different question from how hard the credential is to get. The law does this in two ways. It can require a licence, so only a credentialed person may practise. Or it can name an individual who must personally act and answer for it, with no licence anywhere in sight. Most values are derived from the entry route already recorded for each occupation ("plus licence", "plus bar", "plus residency"); 85 occupations carry a curated exception, each with its reason attached, because an easy licence can still be an absolute legal barrier and a punishing exam often is not one at all.
We checked these against CareerOneStop's licence data, which records 27,067 state and federal licences by occupation code. It could not be used as the source: it is keyed to occupation codes that are broader than our job list, so roles that merely share a code with a licensed profession inherit its licence, and it records that a licence exists for an occupation rather than that you must hold it to do the work. A professional engineer's licence is needed to stamp drawings, not to be employed as an engineer. Used instead as a check, it found 30 occupations we had wrongly left ungated, mostly licensed clinical roles, and those are fixed.
The second kind of barrier was missing entirely until recently, and it had one loud consequence: chief executives hold no licence and sit at a desk, so the model called them freely replaceable. Sarbanes-Oxley requires the chief executive and chief financial officer to personally certify every financial report, with criminal liability for signing a false one. Ten occupations now carry a barrier of this kind, each citing the instrument that creates it: Sarbanes-Oxley, ERISA for pension fiduciaries, the Bank Secrecy Act for compliance officers, the GDPR for data protection officers, NRC rules for radiation safety officers, and the perjury declaration on IRS Form 990 for nonprofit directors. Founders and franchise owners are included on a different basis: no statute names them, but somebody has to hold the downside, and software cannot be the party that is sued or loses the money.
Seniority is not the test, and we check that it isn't. Chief marketing officers, chiefs of staff, technical program managers and compensation managers are senior, well paid, and carry no personal legal duty, so they are untouched and score as exposed as before. Two guards run on every build: those roles must stay unshielded, and the shield as a whole must keep correlating negatively with pay, which it does. If shielding ever starts tracking salary rather than law, the build fails.
One honest limit. Sarbanes-Oxley binds the officers of public companies, but the occupation "chief executives" also covers the person running a twelve-person firm, who has no such duty. We cannot split an occupation that the underlying pay data does not split, so the barrier is applied to the whole of it. Read it as describing the role at the companies where the law bites.
| Level | Label | Meaning |
|---|---|---|
| 0 | no legal barrier | No credential or personal legal duty attaches to the work |
| 1 | certification expected | A credential is common or expected, but nothing in law stops an uncredentialed person doing the job |
| 2 | licence or named officer | A licence is legally required to practise, or the law requires a specific individual to hold the role and answer for it |
| 3 | a named human must sign | A licensed or legally designated human must personally perform or sign off; the authority cannot be delegated |
This could not be sourced: O*NET as distributed does not ship a licensing file, and its Job Zones measure preparation rather than legal necessity. A licensing dataset would override our values wherever it covers an occupation.
A single 0-100 blend of how a role is likely to hold up. The weights are ours:
growth 0.29 displacement 0.26 salary 0.24 skills 0.09 retention 0.12
Each input is scaled to 0-1 before the weights are applied, and where a scale has a boundary that boundary is ours. Growth is normalized over -10% to +20% projected 10-year change, a window holding 95% of the occupations we carry; the 26 outside it sit at one end or the other. It previously used a -2 to +4 window left over from a hand-authored rating, which gave 62% of the catalogue the same growth contribution and let the term take 44.5% of the score's spread while declaring 0.29.
The displacement term replaced a sector-wide automation constant, which gave every occupation in a sector the same contribution regardless of what the job actually involves. There is deliberately no remote-work term: remote-ability correlates -0.96 with physical demand, which is already inside the shield that displacement is built from, so including it would count one signal twice. The weight it used to carry was redistributed across the remaining terms in proportion, so removing the double count did not smuggle in a new opinion about what matters.
Because it folds displacement in, this score inherits everything said above about the shield being our judgement, and it is not externally checked by anything.
On every occupation we show our modeled exposure next to independent, published benchmarks, compared by percentile within each method's own distribution (the raw scores use different scales, so relative position is the fair comparison):
When our estimate and the published research disagree sharply, treat the number with extra caution.
Pay and outlook: U.S. Bureau of Labor Statistics (OEWS, Employment Projections), public domain. Skills and work activities: O*NET, U.S. Department of Labor / ETA, CC BY 4.0. AI-exposure benchmarks as listed above. Some occupations have no BLS entry and fall back to curated figures; those are labelled on the occupation itself. Career Atlas is independent and not affiliated with or endorsed by these organizations.
Career Atlas is an independent project. Pay and outlook data: U.S. Bureau of Labor Statistics (public domain). Skills and work activities: O*NET, U.S. Department of Labor / ETA, used under CC BY 4.0. AI-exposure and future-proof scores are modeled estimates, not professional or financial advice.