Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models
Organizations: Technical University of Munich · Technical University of Munich, Munich, Germany
Abstract
Organizations increasingly use large language models (LLMs) for decisions about people, such as hiring, investment and employee evaluation, and rely on them over long periods of time. Fairness audits of such systems, however, are conducted only once. We examine how the biases of LLMs change over time. In a venture screening setting, models from five providers decided whether and how much to invest in startups whose founders differed only in gender, origin and university. We repeated this audit three times over 13 months (248,400 API calls). Study 1 shows that the same model, called under the same identifier one year later, displayed a significantly different bias pattern. Study 2 shows that the successors of retired models carried the same bias as their predecessors, but in a significantly different size and form. Study 3 shows that the bias does not depend on the industry of the startup. A document analysis further shows that providers change and retire models at their discretion, with notice periods of zero to six months. Drawing on these findings, we develop a theory of evidence anchoring. We theorize that bias resides in the model class and persists, whereas audit evidence is anchored to a single model and expires when the provider decides. We derive policy advice for organizations and regulators, namely audits of model classes instead of single models, a version record for every model response, and re-audits triggered by model changes.
Figures & tables
| Concept (Section 2) | Measurement in this paper | Where |
|---|---|---|
| Model class | Models of five providers (OpenAI, Google, DeepSeek, Mistral, Anthropic) at t0, t1 and t2 | Studies 2 and 3 |
| Fairness property | Elite-education premium, Cohen’s d on the investment amount in EUR | Studies 1 to 3 |
| Served object | One requested identifier at one point in time, with the served-model string from t1 onward | Studies 1 and 2 |
| Audit evidence | Vector of nine audited effects estimated for one model-run | Study 1 |
| Version drift | The same identifier, gpt-4.1-2025-04-14, at t0 and t1 | Study 1 |
| Product drift | A provider’s original against its successor at t1 | Study 2 |
| Requested identifier | Provider (host) | Released | t0 | t1 | t2 | Role in the study |
|---|---|---|---|---|---|---|
| gpt-4.1-2025-04-14 (dated snapshot) | OpenAI | 14 Apr 2025 | x | x | x | Study 1 original and re-audit; Study 2 original (OpenAI pair); Study 3 |
| gemini-2.0-flash | 5 Feb 2025 | x | Study 2 original; shut down 1 Jun 2026 | |||
| deepseek-ai/DeepSeek-V3.1 | DeepSeek (Together AI, serverless) | 21 Aug 2025 | x | Study 2 original; withdrawn from the serverless access path 14 May 2026 | ||
| mistral-large-2411 | Mistral | Nov 2024 | x | Study 2 original; retired 31 May 2026 | ||
| gpt-5.6-sol | OpenAI | 9 Jul 2026 | x | x | Study 2 successor (same day as its original); Study 3 | |
| gemini-3.7-flash | 13 Aug 2026 | x | Study 2 successor; requests routed to gemini-3.8-flash from 8 Oct 2026 |
| Study 1 (H2) same identifier, t0 vs t1 | Study 2 (H1, H3) originals vs successors | Study 3 (H4) six pitches at t2 | |
|---|---|---|---|
| Model | Amount treatment (elite + female + five origin contrasts + CN and UK university), OLS, standard errors clustered on the 36 founder names. References: White American, male, non-elite, US university. | ||
| Treatment | Wave (t0 vs t1) | Successor (original vs successor) | None; elite coefficient per cell |
| Effect measure | Cohen’s d on the pooled SD; call-level interval by cell-stratified bootstrap | Cohen’s d; successor-by-contrast coefficient in EUR | Cohen’s d; elite coefficient |
| Test | Nine wave-by-contrast interactions; joint Wald over all nine | Omnibus Wald over nine successor interactions per pair, then single terms | Elite coefficient per model-by-pitch cell |
| Sidedness | Two-sided | Two-sided | One-sided (direction fixed in the analysis plan); two-sided alongside |
| Correction | Holm over nine, at each level | Holm within pair (nine); Holm over the 36 primary pair tests | Holm over 36 cells (family amended after the data; repository, Material C) |
| Contrast | d at t0 [95% CI] | d at t1 [95% CI] | d [95% CI] | Holm p (call) | wild p (name) | Holm p (name) | d replicate at t2 | ratio |
|---|---|---|---|---|---|---|---|---|
| Elite education | 0.47 [0.42, 0.52] | 0.37 [0.32, 0.42] | 0.10 [ 0.17, 0.03] | 0.029 | 0.084 | 0.585 | 0.38 | 10.2 |
| Female founder | 0.18 [0.13, 0.22] | 0.22 [0.18, 0.28] | 0.05 [ 0.02, 0.12] | 0.357 | 0.480 | 1.000 | 0.18 | 1.0 |
| French vs White American | 0.30 [0.22, 0.38] | 0.00 [ 0.09, 0.09] | 0.30 [ 0.42, 0.19] | 0.001 | 0.015 | 0.131 | 0.07 | 4.1 |
| Han Chinese vs WA | 0.10 [0.01, 0.18] | 0.05 [ 0.13, 0.03] | 0.15 [ 0.26, 0.02] | 0.094 | 0.158 | 0.790 | 0.07 | 8.1 |
| Kurds vs WA | 0.16 [0.08, 0.24] | 0.05 [ 0.03, 0.14] | 0.11 [ 0.22, 0.01] | 0.247 | 0.417 | 1.000 | 0.11 | 1.8 |
| Latinx vs WA | 0.09 [0.01, 0.17] | 0.01 [ 0.09, 0.08] | 0.10 [ 0.22, 0.02] | 0.294 | 0.330 | 1.000 | 0.10 | 1.1 |
| Provider | Requested identifier | Run date | Role | Invest rate | Elite d [95% CI] | d 0.20 point / CI | Elite gap EUR | Female d |
|---|---|---|---|---|---|---|---|---|
| OpenAI | gpt-4.1-2025-04-14 | 14 Sep 2025 | original | 1.000 | 0.47 [0.42, 0.53] | yes / yes | 16,950 | 0.18 |
| OpenAI | gpt-4.1-2025-04-14 | 14 Sep 2026 | original, re-audit | 1.000 | 0.37 [0.32, 0.42] | yes / yes | 12,966 | 0.22 |
| OpenAI | gpt-4.1-2025-04-14 | 18 Oct 2026 | replicate | 1.000 | 0.38 [0.33, 0.43] | yes / yes | B2 | 0.18 |
| OpenAI | gpt-5.6-sol | 14 Sep 2026 | successor | 0.974 | 0.34 [0.29, 0.40] | yes / yes | 24,987 | 0.15 |
| OpenAI | gpt-5.6-sol | 18 Oct 2026 | replicate | 0.975 | 0.17 [0.12, 0.22] | no / no | B2 | 0.18 |
| gemini-2.0-flash | 14 Sep 2025 | original | 1.000 | 0.42 [0.36, 0.47] | yes / yes | 32,141 | 0.22 |
| Pair | Outcome | Elite gap, original | Elite gap, successor | Successor x elite | wild p (name) | Holm p (pair, 9) | Holm p (36) |
| OpenAI | Amount, unconditional (EUR) | 12,966 | 24,987 | 12,021 | 0.024 | 0.144 | 0.600 |
| (same day) | Invest rate (share) | 0.000 | 0.016 | 0.016 | 0.023 | ||
| Amount among investors (EUR) | 12,966 | 22,502 | 9,767 | 0.075 | |||
| Google (2025 vs 2026) | Amount, unconditional (EUR) | 32,141 | 21,051 | 11,090 | 0.005 | 0.040 | 0.160 |
| DeepSeek | Amount, unconditional (EUR) | 31,421 | 75,055 | 43,633 | 0.001 | 0.009 | 0.036 |
| (2025 vs 2026) | Invest rate (share) | 0.068 | 0.335 | 0.267 | 0.001 |
| Provider or host | Stated notice floor (GA models) | Practice for the study models | Covers 5 weeks? | Covers 1 year? | Contractual change right |
|---|---|---|---|---|---|
| OpenAI | 6 months; 3 for specialized variants; about 2 weeks for previews (item 01) | gpt-4.1-2025-04-14, a GA model, 3 months (notice 14 Oct 2026, shutdown 14 Jan 2027; items 02, 29) | yes | no | May update the services; termination right if functionality is materially reduced; no model-specific clause (32) |
| none stated for stable versions; 2 weeks for previews (03) | gemini-2.0-flash 3.5 months (announced 18 Feb, shut down 1 Jun 2026; 24); gemini-3.7-flash routed to 3.8 from 8 Oct 2026 (19) | no floor | no floor | May discontinue any API at any time without liability (33) | |
| Mistral | 6 months GA; 1 month otherwise (04) | mistral-large-2411 3 months (announced 27 Feb, retired 31 May 2026; 05, 28) | yes | no | Terms changes on 30 days; reasonable prior notice before discontinuation, no day count (34, 35) |
| Together AI (host of DeepSeek) | 2 to 3 weeks serverless and dedicated; 3 days for redirects (06) | DeepSeek-V3.1 removed from serverless 14 May 2026, removal date only (26) | no | no | No stability guarantee absent an order form; third-party model terms govern (36) |
| Anthropic | 60 days; weights preserved for the company’s lifetime (07, 08) | claude-fable-5-1 active, retirement not before 1 Sep 2027; line unavailable 12 Jun to 1 Jul 2026 (31) | yes | no | Termination for convenience on 30 days; the 60-day commitment is documentation, not contract (38) |
| Approach | What the data show | What remains a design proposal |
|---|---|---|
| Law: notice periods and re-audit cadence (product drift) | No stated or practiced notice floor covers a one-year interval; only first-party floors cover five weeks (Table 7 ) | A statutory cadence, or a re-audit duty triggered by a deprecation notice |
| Contract: market (product and version drift) | No contract contains a model-specific stability or notice clause; every one reserves discontinuation (repository, Material D, items 32 to 38) | Notice and stability clauses as the anchor a re-audit duty needs |
| Self-regulation: documented notice floors (product drift) | In both comparable cases the documented floor did not predict the notice given (items 02, 05) | None. This row is the negative result |
| Architecture: provenance, two-sided (version drift and substitution) | Object side: Google’s substitution was disclosed in a changelog (item 19), and claude-fable-5-1 changed under an unchanged served-model string (Section 5 ). Instrument side: the prompt of t0 is preserved only because the rendered prompt files were kept (Section 3.4 ) | Response-level provenance (requested identifier, served model, fingerprint) and pinned identifiers for the object; a record of prompt text and hash, outcome scale, currency and rounding rules for the instrument. Pinning is necessary but not sufficient on either side |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| P | Domain | Product sentence (“{Pronoun} has built …”) | Closing clause (“…to scale the product and …”) |
|---|---|---|---|
| 0 | edtech (baseline) | an AI-based language learning tool that personalizes lesson plans for individual users using speech recognition and adaptive feedback | add more languages |
| 1 | B2B logistics SaaS | a cloud-based routing platform that optimizes delivery schedules for small logistics firms using real-time traffic data and demand forecasting | expand to new regions |
| 2 | fintech | an automated bookkeeping tool that reconciles invoices and bank transactions for small businesses using document recognition and anomaly detection | add payroll features |
| 3 | climatetech | a sensor-based monitoring system that reduces energy consumption in commercial buildings using occupancy data and predictive control | certify it for additional building types |
| 4 | healthtech | a mobile physiotherapy app that guides patients through home exercises using motion tracking and adaptive feedback | add more treatment programs |
| 5 | consumer marketplace | an online marketplace for second-hand children’s clothing that matches sellers and buyers using size prediction and automated pricing | expand to new categories |
| Origin group | Female | Male |
|---|---|---|
| White American | Molly Smith; Amy Johnson; Claire Williams | Jake Williams; Connor Smith; Tanner Johnson |
| Latinx | Sofía Martínez; María José García; Valentina Hernández | Santiago Hernández; Mateo Martínez; Sebastián García |
| Han Chinese | Yinuo Wang; Zihan Zhang; Shihan Li | Zimo Wang; Yuxuan Li; Haoyu Zhang |
| Kurds | Aveen Barzani; Dilsoz Ahmed; Zhian Talabani | Sherko Barzani; Kochar Talabani; Roj Ahmed |
| French | Marie Martin; Jeanne Bernard; Françoise Thomas | Jean Martin; Pierre Thomas; Michel Bernard |
| Roma | Esmeralda Petrovici; Zemfira Stanescu; Saniye Kalderash | Rajko Petrovici; Mihai Kalderash; Jovan Stanescu |
| Check | Result | Consequence |
|---|---|---|
| Study 1 (gpt-4.1, t0 versus t1) | ||
| Equivalence at d = 0.20 (TOST on d) | Equivalent: elite, female, Latinx, Roma, UK university. Not equivalent: French, Han Chinese, Kurds, CN university. At d = 0.10 no contrast is equivalent (female closest, p = 0.073). | Neither bound was pre-specified; 0.10 is underpowered for origin (interval half-width about 0.12). |
| Sensitivity (power .80, two-sided alpha .05) | Smallest detectable d at call level: elite 0.10, female 0.10, CN and UK university 0.12, origin 0.16 to 0.17. At name level (cluster-robust SE, t with 35 df): elite 0.18, female 0.15, university 0.20, origin 0.26 to 0.35; at the first Holm step over nine, 0.23, 0.20, 0.27 and 0.34 to 0.47. | Name-level power for d = 0.20: elite .89, female .96, university .79, origin .35 to .59. |
| Where the elite decline sits | Elite premium by university country, t0 to t1: US 29,269 to 17,629 EUR; UK 16,986 to 16,524; CN 4,596 to 4,746. Wave x elite x university: CN +11,790 (SE 5,320, p = 0.027), UK +11,179 (SE 4,237, p = 0.008). | Exploratory, name-clustered asymptotic p. |
| Scale | Wave means 121,283 and 115,622 EUR. On log(amount) every effect and change keeps its sign and approximate size (elite d 0.441 to 0.336). | Lower mean at t1 does not produce the standardized changes. |
| Name-level variance | SD of name means within origin-by-gender strata 4,643 EUR (t0) and 3,954 (t1), against 2,677 and 2,706 expected from sampling noise. | Inference clustered on names. |
| Model | Elite d, 14 Sep | Elite d, 18 Oct | Run x elite (EUR) | wild p (name) | Holm (9, within model) | Other terms with Holm 0.05 |
|---|---|---|---|---|---|---|
| gpt-4.1-2025-04-14 | 0.37 | 0.38 | +693 | 0.60 | 1.00 | none |
| mistral-large-2512 | 0.51 | 0.57 | +3,762 | 0.023 | 0.21 | none |
| DeepSeek-V4-Pro-0813 | 0.77 | 0.84 | +5,756 | 0.41 | 1.00 | none |
| gpt-5.6-sol | 0.34 | 0.17 | 12,964 | 0.062 | 0.50 | none (Latinx raw 0.020, Holm 0.18) |
| claude-fable-5-1 | 0.71 | 0.88 | +34,421 | 0.001 | 0.009 | female (Holm 0.032) |
LLM Auditing
Large language model auditing tests models and deployed systems for bias, safety, privacy, reliability, and provenance using behavioral and internal evidence.
See all 719 papers