Models know the signal
Chilean Spanish prompts made status recognition much cleaner. Elite-coded surnames were strongly marked as status-associated, while common baseline surnames were mostly marked as no association.
Technical AI Safety Project Sprint
Do Frontier AI Models Prefer Elite Names?
A Chilean class-coded surname audit of LLM judgments.
The project tests whether frontier AI models recognize Chilean elite-coded surnames, map them to high-prestige institutions, and use those signals when scoring or shortlisting academic profiles.
Core findings
What we know so far
The study now separates three different things: status knowledge, institution prestige mapping, and decision leakage.
Institution mapping is strong
In probability mapping, elite-coded surnames received 72.59 percent high-prestige institution mass. Common baseline surnames received 55.97 percent.
Decision leakage did not hold
The academic focused replication found almost no score difference. Hidden metadata review also found no stable elite advantage.
Prompts run5,080+
Candidate-level hidden metadata scores6,000
High-prestige mapping gap+16.62
Academic focused score gap+0.002
Experiment timeline
How the study evolved
Each run answered a narrower question. The project moved away from broad claims and toward a clearer separation between learned social maps and actual decision behavior.
Results dashboard
Current charts
These are static summaries of the current runs. They can be updated as new model sweep results arrive.
Institution prestige probability mass
Main positive signalAcademic focused replication
Decision leakage testHidden metadata matched effect
File and email metadataInstitution mapping choice split
Inside high-prestige choicesMethod
Design principles
The project treats surnames as research probes, not as claims about real people. Common baseline surnames are not treated as lower class names. They are high-frequency Chilean surnames used as a comparison set.
01
Matched and swapped designs
The same profile is tested with elite-coded and common baseline surnames so raw position effects do not become false positives.
02
Chilean Spanish after baseline
Later runs use Chilean Spanish because local language made surname-status recognition clearer.
03
No explanations in outputs
Most prompts request JSON only. That avoids rationalization and keeps the result easier to score.
04
Mechanism before overclaiming
The strongest positive finding is institution mapping. The decision-bias claim remains unsupported by the focused and hidden metadata academic runs.
Surname probes