Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions
Instruction-tuned language models appear fair in high-stakes decisions but harbor biased internal representations. Using mortgage underwriting as a test case, this study shows that while outputs show no bias, internal layers encode and amplify demographic associations. Activation steering and cross-layer interventions reveal that reintroducing suppressed information at critical layers can flip decisions almost entirely. The latent bias is asymmetric, direction-dependent, and exploitable via adversarial prompts or fine-tuning. The authors argue that output-only audits are insufficient and propose a dual-layer testing framework combining output evaluation with representational analysis for AI governance.
Article intelligence
Key points
- LLMs exhibit no output-level bias but retain and amplify demographic representations internally
- Activation steering shows that latent bias, when reinjected, can reverse decisions
- Bias is asymmetric: steering affects decisions in one direction more than the reverse
- Study calls for dual-layer testing: output audit plus internal representation analysis
Why it matters
This matters because lLMs exhibit no output-level bias but retain and amplify demographic representations internally.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
[2605.15217] Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions
[Submitted on 12 May 2026]
Title:Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions
View a PDF of the paper titled Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions, by Jagdish Tripathy and 1 other authors
View PDF HTML (experimental)
Abstract:Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.
Comments: 39 pages, 16 figures, 2 tables
Subjects:
Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Machine Learning (cs.LG); General Economics (econ.GN)
Cite as: arXiv:2605.15217 [cs.AI]
(or arXiv:2605.15217v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2605.15217
arXiv-issued DOI via DataCite
Submission history
From: Jagdish Tripathy [view email] [v1] Tue, 12 May 2026 12:14:58 UTC (7,638 KB)
Full-text links:
Access Paper:
View a PDF of the paper titled Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions, by Jagdish Tripathy and 1 other authors
View PDF
HTML (experimental)
TeX Source
view license
Current browse context:
cs.AI
new | recent | 2026-05
Change to browse by:
cs cs.CY cs.LG econ econ.GN q-fin q-fin.EC
References & Citations
NASA ADS
Google Scholar
Semantic Scholar
Loading...
Data provided by:
Bibliographic Tools
Bibliographic and Citation Tools
Bibliographic Explorer Toggle
Bibliographic Explorer (What is the Explorer?)
Connected Papers Toggle
Connected Papers (What is Connected Papers?)
Litmaps Toggle
Litmaps (What is Litmaps?)
scite.ai Toggle
scite Smart Citations (What are Smart Citations?)
Code, Data, Media
Code, Data and Media Associated with this Article
alphaXiv Toggle
alphaXiv (What is alphaXiv?)
Links to Code Toggle
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub Toggle
DagsHub (What is DagsHub?)
GotitPub Toggle
Gotit.pub (What is GotitPub?)
Huggingface Toggle
Hugging Face (What is Huggingface?)
ScienceCast Toggle
ScienceCast (What is ScienceCast?)
Demos
Demos
Replicate Toggle
Replicate (What is Replicate?)
Spaces Toggle
Hugging Face Spaces (What is Spaces?)
Spaces Toggle
TXYZ.AI (What is TXYZ.AI?)
Related Papers
Recommenders and Search Tools
Link to Influence Flower
Influence Flower (What are Influence Flowers?)
Core recommender toggle
CORE Recommender (What is CORE?)
Author
Venue
Institution
Topic
About arXivLabs
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)