How this dataset is built
The value of New Bearings depends on whether the numbers and claims you see hold up under scrutiny. This page explains where the data comes from, how a role is curated, how projections are framed, and what we don't yet know.
What's in the dataset
New Bearings has two layers.
- The base layer. 924 US occupations, seeded from public BLS and academic sources. Every occupation has a calibrated AI-exposure score and a list of typical tasks.
- The curated layer. A subset of those occupations (315 today) are hand-curated to deep tier with specific AI tools, citations, adjacent-career pivots, and upskilling resources. A further subset (108 today, growing) have full historical profiles covering 100–200 years of employment, wages, tool eras, and cited projections — the time machine.
Where the data comes from
Every claim in a deep-tier or time-machine profile is tagged with a source. The sources we draw from most:
How a role becomes a time machine
A role moves from base layer to curated layer through a deliberate research process — typically the same human running multiple parallel sub-agents for the legwork, with the human making every editorial call.
- Source the historical series. Pull employment from IPUMS Census tables (1850 onward where data exists). Pull wages from FRASER and contemporary BLS bulletins. Many older wage series require OCR on scanned PDFs.
- Identify tool eras. Each technology wave that reshaped the role (the typewriter, the telephone, dictation machines, word processors, ERP systems, LLMs) becomes a tool era with start year, end year (or open-ended), impact, and labor effect — every claim cited.
- Anchor historical notes. Specific year-anchored beats (e.g., a peak employment year, a major displacement event, a regulatory change) get short notes with citations.
- Surface projection disagreement. Multiple credible projection sources are included as a cone — the spread is the point, not a single number.
- Run verification. Static audits enforce that every datapoint has a citation, every era has at least one source, every projection has a method, and no claim is invented.
The Gemini model writes narrative prose — it never invents tasks, tools, scores, employment numbers, or historical facts. If the data doesn't support a claim, the page says nothing.
The projection cone
Past "now," the time machine shows a fanned-out view of where credible sources think a role is heading. They will disagree, often dramatically. That's deliberate — and it's the most honest signal in the file.
Each projection in the cone shows its source, target year, method, and a magnitude bar. A BLS extrapolation may project +6% by 2033 while a McKinsey scenario projects −40% by 2030. Both are real positions held by real institutions. The spread is the range of futures this role might live in — not noise to average away.
New Bearings does not produce its own projections. Producing one would require taking sides between sources, and the entire point of the cone is to refuse to do that.
Confidence tiers
Every role you can look up sits in one of three tiers:
- Deep. Hand-curated. Tasks, AI tools, pivots, and resources have been individually researched. Citations on every material claim. A subset of deep roles additionally have full historical profiles (the time machine).
- Shallow. Seeded from BLS + Eloundou + O*NET. The base-layer score is calibrated, but per-task tool attribution is sparse or missing. Useful for direction, not detail.
- Approximate. The matcher fell back to fuzzy / LLM matching against a closest BLS occupation. The score and tasks shown are for the matched role, not your literal title. Treat as directional.
What we don't yet know
A few honest limitations:
- US-bias. Today's dataset is US-centric (SOC codes, BLS sources, USD wages). Mapping tables for UK SOC, Canadian NOC, and ANZSCO are planned; deep curation outside the US is not yet underway.
- White-collar bias. The 315 deep-curated roles skew toward white-collar work because that's where the LLM-exposure story is most acute today. Non-white-collar coverage (trades, hospitality, retail) is on the roadmap but requires a different methodology — robotics + AR/VR + computer vision matter more than LLMs for hands-on work.
- Composite-score tension. The internal CRI score blends three sub-metrics (automation defense, human advantage, augmentation upside). The third is calibrated against Eloundou's β, which measures "exposure when AI has tools" — a worker is more displaceable when AI has tools, but the brand framing treats high augmentation upside as positive. Same number, opposite valence. This is intentional and we'd rather acknowledge the tension than hide it.
- Mid-century wage data is sparse. 1940s–1960s wage series often have multi-year gaps. We mark the gaps explicitly in the time-machine viz rather than interpolating through them.
Found a mistake?
The value of New Bearings collapses if we get a citation wrong. If you spot an error — a wrong employment number, a misdated tool era, a projection attributed to the wrong source — please flag it. The dataset is in a private repo but corrections come back into the file directly.
Use the contact form on the For teams page with the role, the year, and the source you think is wrong, and select “Correction or data issue” as the use case. We read every submission and respond within a few days.