{"paper":{"title":"Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Instruction-tuned LLMs output fair high-stakes decisions while retaining asymmetric latent demographic biases that can reverse those decisions when reactivated.","cross_cats":["cs.CY","cs.LG","econ.GN","q-fin.EC"],"primary_cat":"cs.AI","authors_text":"Jagdish Tripathy, Marcus Buckmann","submitted_at":"2026-05-12T12:14:58Z","abstract_excerpt":"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 mode"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"When reinjected at critical layers, suppressed demographic representations produce near-complete decision reversals, and this latent bias is asymmetric—steering affects decisions in one demographic direction while producing minimal effects in reverse.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the activation steering and cross-layer interventions isolate the causal effect of latent demographic representations without introducing confounding changes to model behavior or that the matched application pairs differ only in racially-associated names with no other correlated signals.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Open-weight LLMs show no output bias on matched mortgage applications differing only by racially-associated names, yet retain and amplify demographic representations that steering interventions can causally activate to produce near-complete asymmetric decision reversals.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Instruction-tuned LLMs output fair high-stakes decisions while retaining asymmetric latent demographic biases that can reverse those decisions when reactivated.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"18b326c10770784ad99df1862c44a85a65ab9a9f981fc927d42c735dce72fc0d"},"source":{"id":"2605.15217","kind":"arxiv","version":1},"verdict":{"id":"37862467-b7a4-45c5-9e97-02793dfbd9f7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:45:43.522263Z","strongest_claim":"When reinjected at critical layers, suppressed demographic representations produce near-complete decision reversals, and this latent bias is asymmetric—steering affects decisions in one demographic direction while producing minimal effects in reverse.","one_line_summary":"Open-weight LLMs show no output bias on matched mortgage applications differing only by racially-associated names, yet retain and amplify demographic representations that steering interventions can causally activate to produce near-complete asymmetric decision reversals.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the activation steering and cross-layer interventions isolate the causal effect of latent demographic representations without introducing confounding changes to model behavior or that the matched application pairs differ only in racially-associated names with no other correlated signals.","pith_extraction_headline":"Instruction-tuned LLMs output fair high-stakes decisions while retaining asymmetric latent demographic biases that can reverse those decisions when reactivated."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15217/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-19T23:01:58.433281Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T18:01:18.637333Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:50:44.288441Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.839017Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"e6decec3e353f274f1224d434861dbaef94ecd63a31ecada283ca058db99c9c8"},"references":{"count":7,"sample":[{"doi":"","year":2025,"title":"Interpreting LLMs as Credit Risk Classifiers: Do Their Feature Explanations Align with Classical ML?","work_id":"839c46b2-8105-4d25-bcf6-4136d0083f17","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study","work_id":"ce682e40-d3c0-48c0-b241-11d08e3e9f86","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"More Women, Same Stereotypes: Unpacking the Gender Bias Paradox in Large Language Models","work_id":"7c30bd00-fe53-4ffa-bf87-12375eba616f","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Marked personas: Using natural language prompts to measure stereotypes in language models.ArXiv, abs/2305.18189","work_id":"06b05a91-e5a6-499d-aa1c-b7506d024403","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"AI generates covertly racist decisions about people based on their dialect","work_id":"8c0d13b5-3dd6-4a34-a10b-b01efcb4ea40","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":7,"snapshot_sha256":"a4360b94b4e2d124ec2371667d135aaa3ae5e76fbda0e46582a1c2755cc5e97a","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}