Open LLM essay scorer shows persistent first-language bias across prompts

The study's sharpest finding isn't that bias exists, it's that cross-prompt generalization and demographic fairness pull in opposite directions: a model that learns to score well across unseen writing tasks simultaneously encodes stronger L1-correlated disparities. These are not the same problem, and fixing one can worsen the other.

This connects directly to the 'Harnessing LLMs for Reliable Academic Supervision' paper covered the same day, which argued that raw model capability matters less than architectural accountability in high-stakes domains. That paper's prescription (audit trails, human-in-the-loop gates) is exactly what's missing from the scoring pipeline examined here. Both papers are converging on the same uncomfortable conclusion: deploying LLMs in credentialing contexts without structured fairness audits is an institutional liability, not just a technical gap. The routing ceiling work on Gemma covered in our Llama and Gemma piece adds another layer, suggesting that Gemma-3's known distribution-shift vulnerabilities may be compounding the bias problem.

Watch whether ETS or a comparable testing body publishes a formal fairness audit of any LLM-assisted scoring system within the next 12 months. If they do not, the absence itself signals that production deployments are outpacing the accountability infrastructure both papers say is necessary.