# Language models don't understand people.

_They are brilliant at language. People are not language. Why hiring needs a people world model, and what it has to learn._

Published 2026-11-08 · Updated 2026-11-13

Ask a language model whether someone will succeed on your platform team and you will get a polite paragraph. It will list skills from a CV, mention culture fit in the abstract, and sound confident. None of it is grounded in how your company actually works.

That is not a flaw in the model. It is a category error. Language models read text. People show up as text only after the fact: a CV, a LinkedIn headline, interview notes if someone typed them up. The work itself, the team around them, the decisions they owned and the pattern your best hires share, lives somewhere else.

A people world model is built to represent that elsewhere. It learns your teams as they are, not as a job description imagines them. It reads who reports to whom, what shipped, where the gaps are, and how your leaders describe the bet they are making. When it searches the market, it is not matching keywords to keywords. It is asking who would change outcomes here.

That is why the same model can answer different questions for different seats. A recruiter asks which companies breed people like your platform team. A hiring manager asks who completes the team they have today. A founder asks why you are losing people to a competitor, and whether you are staffed for the launch you promised. The language is plain English. The reasoning is company-specific.

Contrast that with AI sourcing. Those tools start from a shared catalogue of profiles. Everyone in your market sees the same names. The LLM on top can rewrite outreach and summarise a CV, but it cannot tell you why Hannah fits your mentoring gap and not your competitor's, because it has never seen your team map.

Hiring managers learn this quickly. They stop trusting ranked lists. They ask for the reason behind a name. When the reason is generic, the name gets ignored. When your own recruiters cannot defend the shortlist, the slate dies in Slack and the role stays open.

A people world model answers with evidence. Team mapping comes before the search: who is on the team, what they optimise for, what leadership wants more of. Culture profile compares how your people work with what your handbook claims. Signals are plain-English labels, not scores, each tied to something a sceptical VP can challenge.

Language models will keep improving at language. They will draft cleaner emails and summarise longer documents. They will not, on their own, learn what good looks like inside your organisation. That learning is private. It comes from your hires, rejections, interview outcomes and the strategy your exec team sets. It should not leak into a shared ranking everyone else queries next quarter.

Orbit is Halloway's first people world model product. It is built as a recruiter's right hand: agent mode when a role lands in your ATS, Scout when a leader describes a person in a sentence, and slates that arrive with reasons. The recruiter still runs the conversations, still reads the human, still closes. The model does the reading at scale.

If you are evaluating AI in hiring, ask one question first: does this tool understand our company, or does it understand English? If it is only the second, you will keep seeing the same five hundred names as everyone else, faster.

Start with definitions if helpful: see our glossary for people world model, team mapping and signals. Then see Orbit on a live search, or read the founding letter on why recruiters were never the bottleneck.
