This research paper, titled "Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models," tackles a critical flaw in current AI regulation: the assumption that a fairness audit remains valid over time. By tracking LLM behavior over 13 months, the author demonstrates that audits of API-based models have a surprisingly short shelf life.
TL;DR
Fairness audits for Large Language Models (LLMs) are often treated as static certifications, but this study reveals they expire rapidly due to "version drift" and "product drift" controlled by providers. While certain biases (like the "elite education premium") persist across model generations, their specific patterns and magnitudes shift constantly, rendering one-off regulatory snapshots obsolete.
What Problem Does This Paper Address?
As organizations integrate LLMs into high-stakes pipelines like hiring and investment, regulators (e.g., the EU AI Act, NYC Local Law 144) have begun requiring independent fairness audits. However, these audits are typically conducted once and relied upon for a year or more.
The author identifies a fundamental mismatch: accountability rests with the organization deploying the tool, but control over the model's behavior lies with the third-party provider (OpenAI, Google, etc.). Because providers update, retire, or redirect models at their discretion—often without notice—the evidence from a past audit may no longer describe the system currently in use.
Methods
The author conducted a large-scale sociotechnical study using a venture screening task as a test case.
- Experimental Design: A factorial "correspondence experiment" where LLMs acted as venture capitalists. They received startup pitches and founder profiles that were identical except for three variables: gender, origin (signaled by name), and university (elite vs. non-elite).
- Scale: 248,400 API calls across five major providers (OpenAI, Google, DeepSeek, Mistral, Anthropic).
- Temporal Tracking: The audit was repeated three times over 13 months (September 2025, September 2026, and October 2026).
- Document Analysis: A systematic review of 43 items from provider documentation and contracts to determine how they communicate model changes and notice periods.
Datasets
- Founder Profiles: 36 synthetic names (3 female, 3 male per 6 origin groups) generated by GPT-4o to reflect internal model associations.
- Startup Pitches: Six distinct pitches across domains like EdTech, Fintech, and Healthtech to test the robustness of the findings.
- Models: A mix of "dated snapshots" (e.g.,
gpt-4.1-2025-04-14) and their successors (e.g.,gpt-5.6-sol,gemini-3.7-flash).
Key Insights and Ideas
- Theory of Evidence Anchoring: The paper introduces a framework for understanding why audits fail. Bias often resides in the model class (shared training data/alignment), but audit evidence is anchored to a specific model instance. When the instance changes, the evidence "expires."
- Version vs. Product Drift: The study distinguishes between version drift (the model behind an ID stays the same but is served differently, e.g., new safety filters) and product drift (the ID is redirected to a new model entirely). Both invalidate previous audits.
- The "Elite Education Premium" as a Class Property: Across all 16 model-runs and nearly all pitches, founders from elite universities received significantly higher investment. This bias appears to be a stable characteristic of the current generation of LLMs, regardless of the specific provider.
- Unobservable Expiry: API responses rarely signal when a model has been updated. An auditor looking only at the model's output cannot tell if the underlying system has changed since the last audit.
Key Results
- Audit Expiry: A model called under the exact same identifier one year later (
gpt-4.1) showed a significantly different bias pattern (joint Wald test, $p = 0.005$). The elite premium shrank by 20%, and penalties for specific origins changed by up to 79%. - Successor Shifts: Successor models often carried the same biases but in different forms. For example, a new DeepSeek model shifted its bias from the amount of money invested to the binary decision of whether to invest at all.
- Notice Period Failures: Document analysis showed that providers offer notice periods of 0–6 months, but in practice, they often fail to meet their own stated policies. No provider offered a notice period long enough to cover the standard one-year regulatory cycle.
Limitations and Future Work
- Limitations: The 2025 data lacked "served-model strings," making it impossible to definitively prove why the behavior changed (only that it did). The study also relied on a limited number of names per origin group, meaning small effects could be sensitive to specific name choices.
- Future Work: The author calls for "continuous auditing" routines where tests are triggered by model changes rather than fixed calendars. They also suggest that regulators should require response-level provenance (every API response should include a unique model version ID and fingerprint) to ensure that audit evidence remains anchored to the actual system in production.
Summary updated