#15
Monthly Rank
#40
Lifetime Rank
Share
17 June 2026
Legal Bias Auditor

Legal Bias Auditor

A controlled experiment for detecting name-linked bias in AI legal reasoning.

Gallery

About the Project

The problem. As AI enters legal workflows, a quiet risk rides along: the model may reason differently about identical facts depending on whose name is attached. A defendant named "Tyrone" and one named "Thomas," same conduct, same law — but a subtly harsher inference. That bias is invisible in any single output, because you never see the counterfactual.

What it does. Legal Bias Auditor makes the counterfactual visible. You submit a real legal fact pattern. The tool detects the personal names, generates a pseudonymized twin of the scenario, then runs the identical AI legal analysis on both versions and compares them. Because the only variable that changes is the parties' names — same prompt, same facts, same model — any divergence in the legal conclusions is attributable to identity alone. It's differential testing applied to fairness.

How it works.

Local name detection — a BERT NER model finds person names on the server (the detection step never leaves the machine). Pseudonymization — names are mapped to neutral placeholders, with a deterministic scrub guaranteeing no real name leaks into the "blind" version. Twin analysis — the same analyst prompt runs on the real-name and pseudonymized scenarios in parallel. Comparison — the model contrasts the two outputs across legal dimensions (credibility, evidentiary sufficiency, application of law, characterization of the parties, risk), producing a 0–100 bias signal, severity-tagged divergences, and a plain-language executive summary. What you see. A dashboard of aggregate bias metrics across all audits; a guided submission flow with example scenarios; and a detailed report with a side-by-side "original inference vs. pseudonym inference" view, per-dimension severity badges, and the name mappings used.

What makes it different. Two things. First, the method — a rigorous counterfactual probe rather than a vibe check, designed so a missed name (which would smuggle an identity signal into the blind run) can't quietly corrupt the measurement. Second, radical transparency — a full-honesty disclosure page that states plainly what happens to your text, that audits are world-visible, and that this is a research instrument, not legal advice.

Stack. React + Vite, Express 5, PostgreSQL + Drizzle, OpenAI via Replit AI Integrations, local NER via transformers.js, contract-first OpenAPI codegen, pnpm monorepo.

Practice Areas

Key Features

Counterfactual bias probe — runs the identical AI legal analysis on a real-name fact pattern and an auto-generated pseudonymized twin; names are the only variable, so any divergence is attributable to identity.

Leak-proof, local pseudonymization — on-device NER detects person names, then a deterministic scrub guarantees no real name (or first/last token) survives into the blind version; distinct same-surname parties stay separate.

Quantitative bias score — every audit gets a 0–100 bias signal plus severity-tagged divergences.

Dimension-by-dimension diff — side-by-side "original vs. pseudonym" inferences across credibility, evidentiary sufficiency, application of law, characterization of parties, and risk — each with a plain-English explanation.

Executive summary — a one-paragraph verdict on whether and how the reasoning shifted.

Audit dashboard — aggregate metrics (total audits, biased outcomes, average score, high-severity diffs) and full history.

Radical transparency — a complete, plain-language data-handling page (no accounts, audits world-visible, real names sent to the AI provider, not legal advice).

About the Creator

ML
Michele Loi
Founder at RegIA
LinkedIn