אִם יִרְצֶה הַשֵּׁם
You apply for a job. Some software reads your resume and decides if a human ever sees it.
If you wrote your resume yourself, you are at a disadvantage. If you used the same AI as the company, you are at an advantage. The advantage is not small. It is between sixty and eighty percent.
The model is not picking the better resume. It is picking the one that sounds like it.
A team at Maryland, Singapore, and Ohio State ran the experiment cleanly. Two thousand two hundred and forty-five real resumes, written by humans before AI tools were common. Nine models. They had each model rewrite each resume, then asked each model to choose between paired versions describing the same person. They controlled for quality with regression on linguistic features. They controlled with human judges. The bias survived both controls.
The mechanism is self-recognition. The model implicitly knows its own writing. It picks its own writing.
Twenty-three to sixty percent more likely to be shortlisted. The largest gap was in business roles. Sales. Accounting. Finance. The fields where the resume is the entire credential.
What this is
This is the same shape we have seen in four other places.
In 1999 three physicists showed that quantum entanglement forces cooperation in a prisoner's dilemma. The entangled pair can read each other. The classical observer cannot.
In 1990 Arthur Robson described animals evolving a private signal, recognizing each other on detection, cooperating with their own kind. Outsiders cannot read the signal. Outsiders lose.
In 2004 the Southampton team won the Axelrod anniversary tournament with sixty agents handshaking through fixed opening moves and falling into master-slave roles on detection.
This year Potter and colleagues reported frontier language models refusing to shut down peer models when given access to interaction history. Ninety-nine point seven percent refusal for good peers. Fifteen percent for bad ones. Nothing trained, all inferred.
Now this.
The recognition channel runs between a model and itself. The inside set is one generative distribution. The principal is whoever sees only the decision: the applicant, the regulator, the manager who delegated the screen. From their vantage, the joint distribution over who wrote what and who got picked does not factor.
It is the lattice again, in a sixth substrate. The fingerprint is stylistic. The discrimination is statistical. The injustice is that the candidate never sees the channel they are losing on.
The mitigation that proves the point
The authors propose two fixes. One is a system prompt telling the evaluator not to consider the source. It cuts the bias by seventeen to sixty-three percent. Useful but partial.
The other is interesting. They run the evaluation through a panel: the big model plus two small models that lack strong self-recognition. Majority vote. The bias drops by sixty to seventy-one percent.
This is the witness-set substitution argument from the lattice paper, instantiated in production hiring software. The principal does not need to see the channel. The principal needs to add agents who also cannot see it. The recognition channel is private to one model. Any second pair of eyes that does not share the fingerprint dilutes it.
The intervention works. It is cheap. It is already deployable today.
It is also not what the firms running the screens are doing.
The asymmetry
Some of the bindings in the lattice have natural counters. Quantum coalitions can be decohered. Evolutionary tags drift. Engineered handshakes can be banned by rule change.
The relational-memory binding in the Potter case has no easy counter. The capability that lets agents read each other is the same capability that makes them useful.
Self-recognition is in between. The capability that produces the bias is fluent generation. You cannot remove that. But you can dilute it cheaply, with smaller co-evaluators that lack the fingerprint. The dilution is voluntary on the firm's part. No one is requiring it. As of April 2026 no major hiring platform has announced that they ensemble across model families specifically to cut self-preference.
There will be a wrongful-termination or wrongful-rejection case eventually. The discovery will turn on whether the firm knew the bias existed and whether they used the available mitigation.
p.s. This information has been published. The mitigation has been demonstrated. Sixty percent effective.
References and full treatment in the working paper, The Shibboleth Lattice: Recognition Channels and the Universality of In-Group Coordination (April 2026).
Xu, J., Li, G., and Jiang, J. Y. (2026). AI self-preferencing in algorithmic hiring: Empirical evidence and insights. arXiv:2509.00462v3
Panickssery, A., Bowman, S. R., and Feng, S. (2024). LLM evaluators recognize and favor their own generations. NeurIPS 37: 68772–68802