A new digital divide: The AI you use can decide what jobs you get

AI screeners are 23–60% more likely to shortlist CVs written by their own model — and the model you use is quietly becoming a marker of where you land in society.

Applying for a job in 2026 is mostly an AI-to-AI conversation. You write your CV with a chatbot. The company reads it with another one. Humans appear at the ends of the pipeline — pressing send, signing the offer.

But the machine on the other side may have a favorite. And the favorite is itself.

The machines have a favorite

A UK law firm received 178 applications for one IT job — a year earlier, that ad drew about two dozen. (FT / Irish Times) The firm asked an internal AI agent to score a shortlist the humans had already picked. The agent’s favorites were the glossy ones — exceptional feats in record time, big projects on thin experience. Too good to be true, and smelling strongly of AI authorship. The head of HR had two theories: the agent read AI writing as evidence of AI capability, or it simply preferred CVs AI had a hand in. Same conclusion, different optics. The firm hired a human-chosen candidate — not the agent’s number one.

Notice what nobody did wrong here. Nobody programmed a preference. It emerged.

It is not a one-off. A working paper from Xu et al. simulated hiring for 24 jobs — teachers, consultants, chefs, engineers. Candidates whose CVs were written by the same model as the screener were 23–60% more likely to be shortlisted. In some configurations, measured self-preference hit 67–82%. The bias tracks the origin of the text, not its quality.

The AI watermark was built for slop

Today that preference is implicit — a style read, fuzzy and deniable. That changes this month.

In October, OpenAI published textGrain, an invisible statistical watermark woven into word choices, rolling out to ChatGPT and Codex output in the EU to satisfy the AI Act’s requirement that generated text be identifiable in machine-readable form. The same week, Google’s SynthID Detector went fully public — anyone can now upload a file and check it against watermarks from Google and its partners. Provenance is no longer a research demo. It is infrastructure.

One wrinkle: the watermark is going onto text now, but the tool that reads it back is still restricted to approved researchers. The labels are spreading faster than anyone outside the labs can read them. That gap is why the honest framing here is “not ready,” not “already here.”

And the label is both sticky and unreliable, which is the worst combination for fairness. textGrain detects 80% of AI-written passages at 200 tokens, at a 1% false-positive rate. Swap out 10% of the words and detection drops from 92% to 66%; at 25%, it collapses to 17%. OpenAI itself is careful to say a watermark “does not measure human contribution,” and its absence doesn’t prove a human wrote it. So the mark survives on genuinely AI-drafted CVs, misfires on formulaic human writing — and either way, it is a signal about the writer, attached to the file, readable at scale.

The new divide

Run the logic forward and you get three tiers:

Your CV, written with Model X
        ↓  (watermark + style signal rides along)
AI screener, running Model Y
        ↓
        ├── X = Y ──► "sounds like us" ──► boosted to the shortlist
        ├── X ≠ Y ──► "wrong accent" ───► quietly ranked a tier down
        └── no AI ──► "illegible" ──────► the new handicap

Same-model candidates — their CV reads like the screener’s native tongue, watermark or not. Boosted.
Other-model candidates — perfectly good CV, wrong accent. Quietly penalized.
No-AI humans — the new handicap. Not short on skills, just illegible to the reader.

This is no longer hypothetical. By mid-2026, AI chatbot web traffic ran roughly ChatGPT 53.9%, Gemini 27.9%, Claude ~9% (Similarweb data, May 2026). “The wrong model” usually means “the model your recruiter’s vendor didn’t buy.” And candidates cannot see which model is screening them — that information asymmetry is the divide.

A screener cannot measure competence. It measures what it can see. Provenance is the easiest thing in the world to see.

I build AI agents for a living and a few weeks ago I wrote about designing for AI as your audience — the AIx idea. I am exactly the person who would wire up a screening agent like this, with the best intentions: filter the spam, rank by quality, save the recruiters a week. The road to this divide is paved with the productivity gains.

What actually helps

The Xu paper tested mitigations, and the good news is the bias can be more than halved with two moves:

Deliberate blindness — instruct the screener to ignore origin and score content only
Model diversity — use several assessment models so no single model’s taste dominates

One more, on honesty: candidates deserve to know which model is reading them. They require disclosure from human screeners. The bar should not drop because the interviewer is an Artificially Intelligent Agent.

The first digital divide was about access to AI. The next one is about which AI you sound like.

P.S. The Xu paper is still under peer review — the numbers may shift, though the direction holds in follow-up analyses. And in the interest of the point this article makes: yes, this draft was written with AI, then edited by a human who took responsibility for it. If knowing that changed how you weighed the argument, you have just felt the divide.