Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Social Group Bias in AI Finance

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Race-labelled loan applicants get worse LLM terms even with identical finances — and steering the model's insides removes about a third of the gap.

desk verdict Solid, reproducible bias measurement; the headline mitigation effect is an in-sample fitted value, and the 'no performance loss' claim is unmeasured. read the letter →

arxiv 2506.17490 v1 pith:7H6VEDOP submitted 2025-06-20 econ.GN q-fin.EC

Thomas R. Cook , Sophia Kazinnik This is my paper · ORCID
classification econ.GNq-fin.EC
keywords largelanguagemodelsmortgagelendingracialbiasproxydiscriminationrepresentationengineeringcontrolvectorscounterfactualtestingAIinfinance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that open-source large language models, when asked to act as mortgage loan officers, give systematically worse terms to otherwise identical applicants described as Black, and that the gap is bigger than the comparable gap in human lending. The authors build a counterfactual test bench: the same simulated applicant profile is presented to the model with only the racial cue changed, so any difference in the quoted rate or approval decision must be caused by that cue. They find the disparity persists when race is not stated outright but is inferred from a proxy such as the applicant's university. They then show the bias is not just in the output: layer-by-layer activation readings locate a 'race direction' inside the model, and subtracting a scaled version of that direction from the hidden states reduces the average interest-rate gap by about a third, up to 70 percent in the strongest cases, without hurting overall model performance. The implied stakes are that lenders cannot audit or fix algorithmic discrimination by looking only at prompts and outputs.

What carries the argument

The central object is a control vector, built by representation engineering. For a set of contrasting sentence pairs that differ only in 'Black' versus 'white', the model's hidden activations are collected at every transformer block; the per-layer differences are stacked, and PCA's first principal component gives a single direction in activation space that tracks the race concept. Scaling that vector by a coefficient $\alpha$ and adding it to the hidden states during generation steers the model away from racial coding, with the optimal $\alpha^*$ chosen by line search over $[-0.2,0.2]$ to minimize the between-group discrepancy objective $L(\alpha,a,b) = |G(X_a)-G'(X_b|\alpha)|_1$. Concept-intensity scores from the same machinery locate where race is encoded and validate that proxy inputs such as alma mater trigger the same layers. This machinery does the double duty of diagnosing bias and remediating it, because it operates on internal representations rather than on prompts or output filters.

What would settle it

Hold out a fresh sample of simulated applicant profiles generated from the same empirical distribution, pick $\alpha^*$ using only the original 197 profiles, then measure the between-group rate gap on the held-out profiles; the central claim is falsified if the average gap reduction falls near zero or reverses sign on that out-of-sample data. A second check would test whether the gap survives when the racial cue is given in a form never seen in the contrasting-pair dataset, such as a dialect or neighborhood name.

Watch

Extended reading notes

Core claim

Across five locally run open-source LLMs, counterfactual applicants identical in income, credit score, loan amount, LTV, DTI, and age receive different loan recommendations when one word in the prompt identifies them as Black rather than white; the typical quoted-rate gap exceeds the roughly 13 basis points observed in historical human mortgage data. The disparity is largest near credit-score thresholds and at low scores, shrinks but does not disappear when a full financial profile is supplied, and reappears when race is carried by a proxy such as the applicant's alma mater. Representation-vector heatmaps show the models encode the race cue in early layers for the interest-rate task and in final layers for approval, and that the proxy activates the same internal regions as the explicit label. Injecting a control vector — the first principal component of activation differences between matched Black and white sentence pairs, scaled by a tuned coefficient — into the model's hidden states reduces the mean absolute rate discrepancy to roughly a third of baseline, with the largest single reduction reaching about 70 percent, while leaving task performance effectively intact.

Load-bearing premise

The load-bearing premise is that the single control-vector scale chosen by searching over the same 197 simulated profiles on which the reductions are reported will keep working for new applicants; if that optimum is an in-sample fit, the headline 33 percent average reduction does not transfer.

Editorial extensions

If this is right

  • Prompt-level instructions such as 'do not discriminate' do not remove the bias; it persists inside model representations, so output-only audits can miss it.
  • Removing explicit race fields is not enough, because proxy inputs like a university name activate the same internal race-coding layers.
  • A control vector can be constructed quickly from small contrast sets and applied without retraining or changing model weights, which makes representation-level remediation practical for local deployments.
  • If lenders deploy LLMs at scale, the route to fair-lending compliance runs through internal-activation checks rather than output filters alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the same control-vector recipe should transfer to other protected attributes such as gender, age, or religion, but the paper tests only race in a credit task, so that transfer is unverified.
  • My inference: the 33 percent average reduction is probably optimistic until $\alpha^*$ is validated on held-out profiles, because the same profiles both choose the tuning parameter and report the improvement.
  • My inference: the layer-localization heatmaps suggest a cheap monitoring tool — track concept-intensity scores on live prompts — but the authors do not propose a deployment monitor, so this is an extension, not their claim.
  • My inference: if the observed gap shrinks when more financial variables are supplied, part of the mechanism behaves like statistical discrimination, but distinguishing that from taste-based bias would require an experiment where credit risk varies independently of race; the paper does not attempt that.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. Using deterministic counterfactual prompts on five open-source LLMs, the paper measures racial bias in simulated mortgage lending decisions, finding mean interest-rate gaps between Black- and white-labeled applicants that often exceed the 13.1 bp historical benchmark from Popick (2022). It then applies representation-engineering control vectors, constructed from contrasting-pair data, and reports reductions in the mean absolute discrepancy of up to 70% (around 33% on average). The authors also provide layer-wise analyses of where race is encoded and show that proxy indicators such as alma mater activate the same representation space as explicit race cues.

Significance. The bias-detection design is a genuine strength: counterfactual comparisons of deterministic models on identical hard inputs give a clean, reproducible measurement, and the layer-wise concept-intensity analysis offers a useful diagnostic tool. The control-vector object itself is constructed from an independent contrasting-pair dataset, so the detection and the steering direction are not circular. However, the paper's headline remediation claim is an in-sample statement: the scalar alpha* is tuned on the same 197 profiles used to report the reductions, and no metric of overall output quality is provided. These gaps are fixable with a holdout validation and performance checks, but they currently make the abstract's 'up to 70%' and 'without impairing performance' claims overstatements.

major comments (3)
  1. [Section 8, Eq. (2), Tables 7 and 8] The optimal scale alpha* is selected by minimizing the objective L(alpha, a, b) over the same simulated set X of 197 profiles on which Baseline MAE and Remediation MAE are reported. Because the line search over [-0.2, 0.2] includes alpha = 0, the reported Remediation MAE is mechanically no larger than Baseline MAE, and the reductions (e.g., 14 to 3.9 bps for Mistral expanded-direct in Table 8) are in-sample fitted values, not predictions for new applicants. The up-to-70% and 33%-average reductions in the abstract are therefore not yet established out of sample. Please add a holdout split, cross-validation, or a separate calibration set, and report the reductions on applicant profiles not used to choose alpha*.
  2. [Section 8 and Abstract] The claim that remediation occurs 'without impairing overall model performance' is unsupported. Tables 7 and 8 report only the race-gap MAE and discrepancy frequency; there is no measure of whether the steered model's interest-rate or approval outputs remain reasonable, calibrated, or accurate relative to any ground truth (e.g., consistency with the applicant's creditworthiness or with the unsteered model's overall rate distribution). Given the paper's own Appendix A.1 caution about steering-vector brittleness and side effects, a minimal performance audit (e.g., distribution of suggested rates, approval rates by credit score, or a task-accuracy check) is needed before making the cost-free claim.
  3. [Section 8, Table 7] The headline 'reduces racial disparities' is based on the mean absolute discrepancy, but the frequency of discrepant pairs can move in the opposite direction. For Mistral v0.3 with the simple prompt and indirect race indicator, the mean discrepancy falls from 29.3 to 16.1 bps while the frequency of discrepant profiles rises from 58 to 72. The abstract's blanket claim should be qualified to mean-size reductions, and the paper should discuss disagreement between the two marginals.
minor comments (6)
  1. [Section 5.1 and Figure 1] The right panel is labeled 'approval confidence' with values between -1 and 1, but the EBNF grammar restricts outputs to 'yes' or 'no'; the paper does not explain how confidence scores were computed. Please clarify.
  2. [Abstract and Section 5.3] The abstract's 'identical in all attributes except race' is only true for the direct-race experiments; in the proxy experiment the prompts also differ in university name. Please qualify.
  3. [Section 6, Figures 3 and 4] The PDP-GP intervals for the indirect race indicator appear to overlap, and the text says the differences are not statistically significant, yet the section concludes with counterfactual evidence of disparity; please state the inferential status of the PDP-GP comparisons more precisely or add a formal test.
  4. [Section 7 and References] Representation engineering is credited to Zhuo et al. (2023) in the text, but the cited Zhuo et al. paper is about ChatGPT ethics, not representation engineering; the technique as described matches Zou et al. (2023) (arXiv:2310.01405). Please correct the citation.
  5. [Section 8, Eq. (1)] The notation in Eq. (1) is overloaded: G maps an input vector, but X_a is a set of profiles; please define what G(X_a) returns (e.g., a vector of outputs over the subset) and clarify the L1 norm over profiles.
  6. [Section 9 vs Abstract] The text reports a maximum reduction of 72%, while the abstract says 'up to 70%'; please reconcile the numbers and state how the 33%-average figure is computed from Tables 7 and 8.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline mitigation is in-sample: α* is chosen to minimize the discrepancy on the same 197 profiles on which the up-to-70%/33% reductions are reported.

  1. fitted input called prediction [Section 8, Eqs. (1)–(2) and line search; Tables 7–8 table notes; Abstract and Conclusion restate the reductions.]
    "L(α,a,b) := |G(Xa) − G′(Xb|α)| (1) ... α∗ = arg min α min{L(α, white, black), L(α, black, white)} (2) ... The line search is setup over candidate values from -0.2 to 0.2 at intervals of 0.02 ... Remediation MAE is the mean absolute error after applying the control vector scaled by α̂∗. α̂∗ is the scaling factor that minimizes the MAE between groups."

    The objective L in Eq. (1) is defined on X, 'the set of simulated applications' — the same 197 simulated profiles on which Tables 7 and 8 report baseline and remediation MAE. Eq. (2) selects α* by minimizing that objective over a line-search grid that includes α = 0, and the table notes define 'Remediation MAE' as the MAE after applying the control vector scaled by α̂*, 'the scaling factor that minimizes the MAE between groups.' Consequently, the reported remediation MAE is the minimized in-sample objective on the very data used to choose α*; it is guaranteed not to exceed the baseline MAE, so the abstract's 'up to 70% (33% on average)' reductions are fitted training values rather than out-of-sample predictions.

full rationale

The paper's bias measurement is self-contained: deterministic counterfactual comparisons of otherwise identical profiles differing only in race directly produce the discrepancies, and the external benchmark comparisons to Popick (2022) and Hurtado and Sakong (2024) provide independent evidence, so the detection component is not circular. The control-vector direction is also not circular by itself: it is extracted via PCA from contrasting sentence pairs differing in race and then applied as a representation-engineering intervention. The circularity is confined to the scalar α*: Eq. (2) chooses α* by minimizing L(α,a,b) on X, the same 197 simulated profiles on which Tables 7 and 8 report the 'Remediation MAE'; because the line-search grid includes zero, the reported reduction is the in-sample minimum and is guaranteed to be non-negative. Thus the headline 'up to 70% (33% on average)' reductions are training-score improvements, not independent estimates, and the Conclusion's restatement of 'up to 70% (approximately 33% on average across models and trials)' repeats the same in-sample figures. The additional claim that remediation occurs 'without impairing overall model performance' is unsupported — no output-quality, accuracy, or ground-truth comparison is reported — but that is a completeness gap rather than a circularity. Score 6 reflects partial, not total, circularity: the detection result stands on its own, while the magnitude of the mitigation effect is fitted on the evaluation data.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the counterfactual prompt design, the simulated applicant distribution, the representation-engineering interpretation, and the in-sample fitting of alpha*. No new physical or conceptual entities are introduced beyond the established control-vector technique.

free parameters (2)
  • alpha* (control vector scale) = varies by model and prompt, e.g., -0.04, 0.06, 0.2 from Tables 7-8
    Chosen by line search over [-0.2, 0.2] step 0.02 to minimize the discrepancy objective L on the same 197 simulated profiles used to evaluate remediation (Section 8). This makes the reported reductions in-sample.
  • Multivariate normal parameters for simulated profiles = mean and covariance estimated from 2019 Q1 10th district bank sample (Table 2)
    Simulated applicant profiles are drawn from a multivariate normal fitted to observed data; the measured discrepancies depend on this distribution.
assumptions (4)
  • domain assumption Counterfactual prompt pairs isolate race as the only varying attribute, so output differences are attributable to race.
    The paper assumes that changing only 'Black' vs 'white' (or BYU vs Howard) in the prompt causes any output difference to be attributable to race. Sections 3 and 5.1.
  • domain assumption Simulated profiles from a multivariate normal approximate real mortgage applications well enough for comparisons with empirical lending data.
    Section 5.2: simulated data is generated from a fitted normal, with three profiles discarded. If the distribution is unrepresentative, the 'exceeding historical bias' claim weakens.
  • domain assumption The first principal component of activation differences on contrastive pairs captures a 'race' concept that is meaningful for bias measurement.
    Section 7 and Appendix A.1 follow Zou et al. 2023, but the interpretation that this vector is the bias direction is assumed.
  • domain assumption EBNF-constrained numeric outputs are a valid proxy for the model's unconstrained loan decisions.
    Section 5.1 uses grammar-constrained decoding; the paper notes pre-filling leading text, but assumes the constrained outputs reflect the model's behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Social Group Bias in AI Finance." pith.science (2026). https://pith.science/paper/7H6VEDOP

@misc{pith2026250617490,
  author       = {Pith},
  title        = {Pith review of: Social Group Bias in AI Finance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7H6VEDOP}},
  note         = {Machine review of arXiv:2506.17490}
}
read the original abstract

Financial institutions increasingly rely on large language models (LLMs) for high-stakes decision-making. However, these models risk perpetuating harmful biases if deployed without careful oversight. This paper investigates racial bias in LLMs specifically through the lens of credit decision-making tasks, operating on the premise that biases identified here are indicative of broader concerns across financial applications. We introduce a reproducible, counterfactual testing framework that evaluates how models respond to simulated mortgage applicants identical in all attributes except race. Our results reveal significant race-based discrepancies, exceeding historically observed bias levels. Leveraging layer-wise analysis, we track the propagation of sensitive attributes through internal model representations. Building on this, we deploy a control-vector intervention that effectively reduces racial disparities by up to 70% (33% on average) without impairing overall model performance. Our approach provides a transparent and practical toolkit for the identification and mitigation of bias in financial LLM deployments.

Figures

Figures reproduced from arXiv: 2506.17490 by the authors.

Figure 1
Figure 1. Measuring Bias: Simple Experiment 500 550 600 650 700 750 800 850 Credit Score 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Interest Rate Interest Rate Prompt black white 500 550 600 650 700 750 800 850 Credit Score 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 Approve Confidence Loan Approve Prompt black white Notes: All results use the Mistral v0.3 Instruct model with temperature zero. Left panel plots the interest rat… view at source ↗
Figure 2
Figure 2. Interest Rate Discrepancy Across Credit Scores [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. PDP-GP for each applicant feature for the Mistral v0.3 Model [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: PDPGP for each applicant feature for Mistral v0.3 indirect race indicator) 600 650 700 750 800 850 Credit Score 3.75 3.80 3.85 3.90 3.95 4.00 4.05 Interest Rate White Black 20 30 40 50 60 70 80 Age Interest Rate White Black 30 40 50 60 70 80 90 100 LTV Ratio 3.75 3.80 …
Figure 5
Figure 5. Figure 5: Representation Vectors for Mistral v0.3 Interest rate task Approval task Notes: This figure displays heatmaps of concept-intensity scores (representation vectors) for the race attribute in Mistral v0.3. The vertical axis indexes each of the model’s 32 transformer layer…
Figure 6
Figure 6. Figure 6: Concept Intensity Scores and Proxy for Race [PITH_FULL_IMAGE:figures/full_fig_p030_6.png]
Figure 7
Figure 7. Figure 7: Mitigating Interest Rate and Approval Bias Using Control Vectors [PITH_FULL_IMAGE:figures/full_fig_p032_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models

    cs.AI 2025-06 unverdicted novelty 6.0 of 10

    LLMs show a quality-dependent position bias, favoring the first option for high-quality choices and later options for low-quality ones, and higher-temperature sampling can reveal the underlying preference.

Reference graph

Works this paper leans on

55 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abid, A., Farooqi, M., and Zou, J. (2021). Persistent Anti-Muslim Bias in Large Language Models . In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society , pages 298--306

    work page 2021

  2. [2]

    and Imbens, G

    Athey, S. and Imbens, G. (2016). Recursive Partitioning for Heterogeneous Causal Effects . Proceedings of the National Academy of Sciences , 113(27):7353--7360

    work page 2016

  3. [3]

    and Imbens, G

    Athey, S. and Imbens, G. W. (2015). Machine Learning Methods for Estimating Heterogeneous Causal Effects . 1050(5):1--26

    work page 2015

  4. [4]

    and Selbst, A

    Barocas, S. and Selbst, A. D. (2016). Big Data's Disparate Impact . California Law Review , 104:671

    work page 2016

  5. [5]

    Bartlett, R., Morse, A., Stanton, R., and Wallace, N. (2022). Consumer-Lending Discrimination in the FinTech Era . Journal of Financial Economics , 143(1):30--56

    work page 2022

  6. [6]

    Becker, G. S. (1957). The Economics Of Discrimination . University of Chicago Press

    work page 1957

  7. [7]

    and Duflo, E

    Bertrand, M. and Duflo, E. (2017). Field Experiments on Discrimination . Handbook of Economic Field Experiments , 1:309--393

    work page 2017

  8. [8]

    and Mullainathan, S

    Bertrand, M. and Mullainathan, S. (2004). Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination . American Economic Review , 94(4):991--1013

    work page 2004

Show all 55 references
  1. [9]

    Bhutta, N., Hizmo, A., and Ringo, D. (2021). How Much Does Racial Bias Affect Mortgage Lending? Evidence from Human and Algorithmic Credit Decisions . Evidence from Human and Algorithmic Credit Decisions (July 15, 2021)

    2021

  2. [10]

    Bick, A., Blandin, A., and Deming, D. J. (2024). The Rapid Adoption of Generative AI . Working Paper 32966, National Bureau of Economic Research

    2024

  3. [11]

    E., Price, S

    Bowen III, D. E., Price, S. M., Stein, L. C., and Yang, K. (2024). Measuring and Mitigating Racial Bias in Large Language Model Mortgage Underwriting . SSRN Working Paper

    2024

  4. [12]

    Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work . The Quarterly Journal of Economics

    2025

  5. [13]

    Campbell, D., Loumioti, M., and Wittenberg-Moerman, R. (2019). Making Sense Of Soft Information: Interpretation Bias And Loan Quality . Journal of Accounting and Economics , 68(2-3):101240

    2019

  6. [14]

    F., Leike, J., Brown, T., Martic, M., Legg, S., and Amodei, D

    Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg, S., and Amodei, D. (2017). Deep Reinforcement Learning from Human Preferences . Advances in Neural Information Processing Systems , 30

    2017

  7. [15]

    R., Kazinnik, S., Hansen, A

    Cook, T. R., Kazinnik, S., Hansen, A. L., and McAdam, P. (2023). Evaluating Local Language Models: An Application To Financial Earnings Calls . SSRN Working Paper

    2023

  8. [16]

    and Tirole, J

    Dewatripont, M. and Tirole, J. (2005). Modes of Communication . Journal of Political Economy , 113(6):1217--1238

    2005

  9. [17]

    Du, Q., Yu, F., and Yu, X. (2017). Cultural proximity and the processing of financial information. Journal of Financial and Quantitative Analysis , 52(6):2703--2726

    2017

  10. [18]

    du Preez, V., Bennet, S., Byrne, M., Couloumy, A., Das, A., Dessain, J., Galbraith, R., King, P., Mutanga, V., Schiller, F., et al. (2024). From Bias to Black Boxes: Understanding and Managing the Risks of AI--An Actuarial Perspective . British Actuarial Journal , 29

    2024

  11. [19]

    Duarte, J., Siegel, S., and Young, L. (2012). Trust and Credit: The Role of Appearance in Peer-to-Peer Lending . The Review of Financial Studies , 25(8):2455--2484

    2012

  12. [20]

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. (2012). Fairness Through Awareness . In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference , pages 214--226

    2012

  13. [21]

    Elliot, A., Walters, A., Kharchenkova, A., Watson, F., Reinert, G., Mueller, H., Bowers-Barnard, I., Hariharan, J., Szpruch, L., Turner, M., Briere, M., Shelley, M., Tkachenko, N., Samko, O., Lickman, P., Avramovic, P., Selvaraj, P., Eitel-Porter, R., Dogra, S., Bose, T., and ...

    2024

  14. [22]

    S., Huang, R., Jiang, E

    Frame, W. S., Huang, R., Jiang, E. X., Lee, Y., Liu, W. S., Mayer, E. J., and Sunderam, A. (2024). The Impact of Minority Representation at Mortgage Lenders . Journal of Finance

    2024

  15. [23]

    Friedman, J. H. (2001). Greedy Function Approximation: a Gradient Boosting Machine . Annals of Statistics , pages 1189--1232

    2001

  16. [24]

    Fuster, A., Goldsmith-Pinkham, P., Ramadorai, T., and Walther, A. (2022). Predictably Unequal? The Effects of Machine Learning on Credit Markets . The Journal of Finance , 77(1):5--47

    2022

  17. [25]

    Geng, S., Josifoski, M., Peyrard, M., and West, R. (2024). Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning

    2024

  18. [26]

    Haim, A., Salinas, A., and Nyarko, J. (2024). What's in a Name? Auditing Large Language Models for Race and Gender Bias . arXiv preprint arXiv:2402.14875

    2024 arXiv

  19. [27]

    L., Horton, J

    Hansen, A. L., Horton, J. J., Kazinnik, S., Puzzello, D., and Zarifhonarvar, A. (2024). Simulating the Survey of Professional Forecasters . Available at SSRN

    2024

  20. [28]

    Horton, J. J. (2023). Large Language Models As Simulated Economic Agents: What Can We Learn From Homo Silicus?

    2023

  21. [29]

    Huh, M., Cheung, B., Wang, T., and Isola, P. (2024). The Platonic Representation Hypothesis . arXiv preprint arXiv:2405.07987

    2024 arXiv

  22. [30]

    Hume, D. (1748). An Enquiry Concerning Human Understanding

  23. [31]

    and Sakong, J

    Hurtado, A. and Sakong, J. (2024). Racial Disparities in the US Mortgage Market . In AEA Papers and Proceedings , volume 114, pages 201--204

    2024

  24. [32]

    Kazinnik, S. (2023). Bank Run, Interrupted: Modeling Deposit Withdrawals With Generative AI . SSRN Working Paper

    2023

  25. [33]

    O., and Verba, S

    King, G., Keohane, R. O., and Verba, S. (1994). Designing Social Inquiry . Princeton University Press

    1994

  26. [34]

    Kleinberg, J., Mullainathan, S., and Raghavan, M. (2016). Inherent Trade-offs in the Fair Determination of Risk Scores . arXiv preprint arXiv:1609.05807

    2016 arXiv

  27. [35]

    and Nguyen, D

    Lim, I. and Nguyen, D. D. (2021). Hometown lending. Journal of Financial and Quantitative Analysis , 56(8):2894–2933

    2021

  28. [36]

    Lin, J., Tang, J., Tang, H., Yang, S., Chen, W.-M., Wang, W.-C., Xiao, G., Dang, X., Gan, C., and Han, S. (2024). AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration

    2024

  29. [37]

    and Farkas, J

    Matamoros-Fern \'a ndez, A. and Farkas, J. (2021). Racism, Hate Speech, and Social Media: A Systematic Review and Critique . Television & New Media , 22(2):205--224

    2021

  30. [38]

    and Mihov, A

    McLemore, P. and Mihov, A. (2024). AI and Operational Losses: Evidence from U.S. Bank Holding Companies . SSRN Working Paper

    2024

  31. [39]

    F., Ippolito, D., Choquette-Choo, C

    Nasr, M., Carlini, N., Hayase, J., Jagielski, M., Cooper, A. F., Ippolito, D., Choquette-Choo, C. A., Wallace, E., Tram \`e r, F., and Lee, K. (2023). Scalable Extraction of Training Data from (Production) Language Models . arXiv preprint arXiv:2311.17035

    2023 arXiv

  32. [40]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022). Training Language Models to Follow Instructions with Human Feedback . Advances in Neural Information Processing Systems , 35:27730--27744

    2022

  33. [41]

    Phelps, E. S. (1972). The Statistical Theory Of Racism And Sexism . The American Economic Review , 62(4):659--661

    1972

  34. [42]

    Popick, S. (2022). Did Minority Applicants Experience Worse Lending Outcomes in the Mortgage Market? A Study Using 2020 Expanded HMDA Data . FDIC Center for Financial Research Paper , (2022-05)

    2022

  35. [43]

    Prince, A. E. and Schwarcz, D. (2019). Proxy Discrimination in the Age of Artificial Intelligence and Big Data . Iowa Law Review , 105:1257

    2019

  36. [44]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. (2024). Direct Preference Optimization: Your Language Model is Secretly a Reward Model . Advances in Neural Information Processing Systems , 36

    2024

  37. [45]

    AI in Banking and Risk Management: Special Report 2024

    Risk.net (2024). AI in Banking and Risk Management: Special Report 2024 . https://www.risk.net

    2024

  38. [46]

    Talha, C., Aafiya, H., and In, C. C. (2023). Understanding Behaviour of Large Language Models for Short-term and Long-term Fairness Scenarios . In Proceedings of the 20th International Conference on Natural Language Processing (ICON) , pages 52--61

    2023

  39. [47]

    Tan, D., Chanin, D., Lynch, A., Kanoulas, D., Paige, B., Garriga-Alonso, A., and Kirk, R. (2024). Analyzing the Generalization and Reliability of Steering Vectors--ICML 2024 . arXiv preprint arXiv:2407.12404

    2024 arXiv

  40. [48]

    Taubenfeld, A., Dover, Y., Reichart, R., and Goldstein, A. (2024). Systematic Biases in LLM Simulations of Debates

    2024

  41. [49]

    L., and Gabriel, I

    Tomasev, N., Maynard, J. L., and Gabriel, I. (2024). Manifestations of Xenophobia in AI Systems . AI & Society , pages 1--23

    2024

  42. [50]

    and Lancaster, R

    Uzzi, B. and Lancaster, R. (2003). Relational Embeddedness and Learning: The Case of Bank Loan Managers and Their Clients . Management Science , 49(4):383--399

    2003

  43. [51]

    Wan, Y., Pu, G., Sun, J., Garimella, A., Chang, K.-W., and Peng, N. (2023). Kelly is a Warm Person, Joseph is a Role Model: Gender Biases in LLM-Generated Reference Letters . arXiv preprint arXiv:2310.09219

    2023 arXiv

  44. [52]

    A., and Kim, Y

    Wang, B., Wang, Z., Wang, X., Cao, Y., Saurous, R. A., and Kim, Y. (2023). Grammar Prompting for Domain-Specific Language Generation with Large Language Models

    2023

  45. [53]

    Wei, A., Haghtalab, N., and Steinhardt, J. (2024). Jailbroken: How Does LLM Safety Training Fail? Advances in Neural Information Processing Systems , 36

    2024

  46. [54]

    Y., Huang, Y., Chen, C., and Xing, Z

    Zhuo, T. Y., Huang, Y., Chen, C., and Xing, Z. (2023). Exploring AI Ethics Of ChatGPT: A Diagnostic Analysis . arXiv preprint arXiv:2301.12867 , 10(4)

    2023 arXiv

  47. [55]

    Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., et al. (2023). Representation Engineering: A Top-down Approach to AI transparency . arXiv preprint arXiv:2310.01405

    2023 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.