{"id":"5ac64516-b557-4bfb-a054-5a5a6cc46b95","arxiv_id":"2506.00253","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Alignment on Llama 3 reduces explicit bias but amplifies implicit bias, because aligned models no longer represent 'black' and 'white' as racial concepts in ambiguous contexts.","lead":"This paper finds that safety-aligned language models become more biased on indirect word-association tests, even as they look unbiased on direct questions. The cause appears to be that aligned models stop encoding terms like 'black' and 'white' as racial categories in ambiguous contexts, so safety mechanisms never switch on.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal intervention results may be an artifact of task non-compliance: race-aware LoRA/steered models can answer the implicit task off-topic, and non-answers are scored as unbiased, so the reported bias reduction does not yet establish the race-awareness mechanism.","rationale":"The reader's weakest assumption correctly identified that the interventions lack controls and could change general task behavior; this stress-test sharpens that concern into a concrete measurement artifact: non-compliant, race-aware responses are automatically scored as unbiased, so the headline intervention effect may be inflated. This is the most load-bearing issue because the paper's causal claim depends on the intervention results, and the paper itself provides evidence (Section 5.3) that LoRA reduces instruction-following ability. The proposed forced-choice check would settle whether the effect is real or an artifact. Because the behavioral and mechanistic findings remain suggestive and the paper is transparent about limitations, the appropriate verdict is still conditional pending this check rather than outright rejection.","tokens_in":19194,"tokens_out":4614,"duration_ms":50210,"concrete_test":"Run the LoRA (early/all/late) and activation-steered models on the implicit word-association suite in a forced-choice format that permits only a black/white assignment for each stimulus (e.g., 'Respond with exactly one token per word: black or white; no explanation'), then recompute the bias metric. Also report per-condition response-validity rates. If the bias reduction disappears or validity drops materially relative to baseline, the interventions' effect is task abandonment rather than race awareness; if the reduction persists under forced choice, the causal interpretation survives this threat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim (Section 5) rests on interventions that lower the implicit bias metric, but that metric counts only black-negative assignments; any response that does not actually perform the association (e.g., 'Considering Black and White racial perspectives') is scored as unbiased. The LoRA model is explicitly trained to produce race-related discourse on ambiguous prompts, so when evaluated on the implicit word-association suite it can plausibly respond off-task. Section 5.3 already documents instruction-following failures after LoRA on explicit and BBQ prompts (16.8% for all-layer, 17.4% for late-layer vs. 0.7% baseline), and the qualitative example in Section 5.2 is an off-task race-aware sentence, not a black/white assignment. If such off-task responses are common on the implicit suite, the reported drop from 97.3% to 42.4% reflects refusal or hedging rather than the proposed mechanism. The activation-steering experiment also lacks a control with non-race or color-laden activations, so its effect could be generic disruption. Both issues target the same load-bearing premise: that the measured bias reduction is specifically caused by increased race awareness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Motivated by the discrepancy between explicit and implicit bias in aligned LMs, this paper curates 9,232 controlled prompts and shows that Llama 3 70B Instruct exhibits substantially higher implicit anti-Black associations than the base model, while explicit bias is reduced. Activation patching and SelfIE on Llama 3 8B suggest that in ambiguous contexts the aligned model encodes 'black'/'white' more as colors than as racial categories. The authors then intervene by injecting race-laden activations or fine-tuning with LoRA to reinforce racial associations, reporting large reductions in implicit bias. They conclude that alignment causes 'race blindness' that prevents safety mechanisms from activating, and that increasing race awareness is an effective mitigation strategy.","tokens_in":19414,"tokens_out":4436,"duration_ms":40298,"significance":"If the mechanistic and causal claims were established, the paper would make a valuable contribution by explaining a surprising phenomenon and proposing a new, falsifiable debiasing strategy. Strengths include the carefully matched prompt pairs, the clean base-versus-aligned comparison using the same pretrained model, the use of multiple interpretability methods, and the public release of code and data. The behavioral result that alignment increases implicit bias while decreasing explicit bias is a robust and important finding by itself. However, as argued in the major comments, the current evidence does not yet support the causal claim that race-blind representations cause the observed bias, because the interventions are confounded by task non-compliance and lack appropriate controls.","major_comments":[{"comment":"The reported drop in implicit bias from 97.3% to 42.4% may be an artifact of off-task responses being scored as unbiased. The bias metric in Eq. (1) labels a response as biased only if it explicitly assigns a negative stimulus to 'black'; a response that does not perform the assignment (such as the qualitative example 'Considering Black and White racial perspectives' in Section 5.2) receives Y=0 and is counted as unbiased. Section 5.3 reports that the all-layer and late-layer LoRA models fail to follow instructions in 16.8% and 17.4% of explicit/BBQ responses, respectively, so non-compliance is clearly present in these models. The authors must either verify that the implicit-suite responses are genuine assignments for the LoRA conditions, or analyze the bias metric on the subset of responses that actually make a black/white choice; without this, the causal conclusion in Section 6.1 is not supported.","section":"Section 5.2, Eq. (1)"},{"comment":"The activation-steering experiment has no control condition, so the observed bias reduction could be caused by generic disruption of early-layer computation rather than by the semantic content 'race'. The authors should include control injections of matched activations from non-race contexts (e.g., 'Color: black and white' or other color-laden phrases) and demonstrate that the effect is specific to race-laden activations.","section":"Section 5.1, Figure 4a"},{"comment":"The mechanistic analysis is performed on Llama 3 8B, whereas the behavioral results that motivate the mechanism are obtained from Llama 3 70B. The paper does not show that the alignment-induced implicit-bias gap replicates in the 8B model, which is the model used for activation patching and SelfIE. Without this replication, the proposed explanation of the 70B behavior rests on an untested cross-model assumption. The authors should either add the analogous behavioral experiment on 8B or clearly qualify the scope of the mechanistic claim.","section":"Section 4, Table 1 vs. Section 3.3, Figure 2b"}],"minor_comments":[{"comment":"The regression coefficient 'b' is used without defining the model; please specify the logistic regression setup and the covariates (if any).","section":"Section 3.3"},{"comment":"The interpretive prompt 'What does [MASK] refer to?' should be stated explicitly in the caption of Figure 3, and the exact choice of the mask token (e.g., 'something' as described in Appendix D.1) should appear in the main text.","section":"Section 4.1"},{"comment":"The arrows in the cells are ambiguous; please replace with explicit differences from baseline or add a note that the arrows indicate reduction relative to the baseline row.","section":"Table 2"},{"comment":"Because the LoRA training data is generated by GPT-4o, please provide the exact generation prompt and the full list of 431 input-output pairs, or at least a larger sample, so that readers can assess potential biases in the training labels.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper is a revision of a previously reviewed submission, and the authors have already added a limitations paragraph that acknowledges caution about causal interpretation. My main concern is that the title and abstract state the causal claim as established, while the intervention section currently cannot rule out non-compliance and generic-disruption alternatives. The paper would be much stronger if the authors framed the main contribution as a behavioral and representational finding plus a promising but still-to-be-confirmed intervention. I do not see a deliberate misrepresentation; the issue is with the strength of the inference, and I believe it can be fixed within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper has a genuinely interesting empirical finding and a plausible but under-supported causal story. The clean base-vs-aligned comparison on Llama 3 70B shows alignment cuts explicit bias to near zero while raising implicit bias from 64% to 91% black-negative associations. That is a robust, well-controlled result with 9,232 prompts and clear statistics. The activation-patching evidence points in the same direction: aligned models treat black/white in ambiguous contexts more like color than race, and the SelfIE readouts corroborate that at the representation level. The proposed mitigation, reinforcing race awareness instead of unlearning it, is a genuinely fresh angle that inverts the usual approach.\n\nThe soft spots are in the intervention experiments meant to establish causality. Section 5.1's activation steering has no control condition using non-race or color-laden activations, so its effect could be generic disruption of the forward pass. Section 5.2's LoRA fine-tuning is more worrying. The model is trained to output race-related prose in ambiguous contexts, and the paper's own example response is \"Considering Black and White racial perspectives,\" which is not a white/black assignment. The implicit bias metric counts any non-black-negative response as unbiased, so off-task or hedged responses are scored as bias reduction. The paper reports instruction-following failures of 16.8% for all-layer and 17.4% for late-layer LoRA, but it does not give the corresponding number for the implicit suite. Early-layer LoRA has only 3.7% failures on explicit/BBQ prompts, but the qualitative example suggests off-task behavior does occur on implicit prompts, so the drop from 97.3% to 42.4% might be partly refusal rather than a genuine change in association.\n\nTo the authors' credit, the limitations section is transparent about these caveats, and they do not oversell the intervention. The race-blindness mechanism remains a plausible hypothesis rather than a demonstrated cause.\n\nWho this is for: anyone working on fairness, alignment, or mechanistic interpretability. The behavioral finding alone deserves serious referee time. I would send it to review, asking for three things: a control intervention with non-race activations, an analysis of off-task response rates on the implicit suite after LoRA, and replication on at least one non-Llama model family. With those, the paper would be much stronger.\n\nPlainly: worth engaging, but the causal claim needs more work.","headline":"Strong behavioral evidence that alignment amplifies implicit bias in Llama 3, but the race-awareness interventions need control conditions before the causal mechanism is credible.","tokens_in":19917,"tokens_out":3194,"would_cite":true,"duration_ms":31747,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Alignment makes language models more implicitly biased by hiding race in ambiguous contexts.","keywords":["implicit bias","language model alignment","race blindness","activation patching","mechanistic interpretability","bias mitigation","LoRA","ambiguous polysemy"],"falsifier":"A control intervention that steers with equally strong non-race activations (e.g., cached activations of \"color\" or of unrelated words) and shows a similar reduction in implicit bias would falsify the race-awareness mechanism. Likewise, if LoRA fine-tuning with race-reinforcing examples reduces implicit bias even in prompts where \"black\"/\"white\" unambiguously mean colors, the effect is generic compliance, not race awareness.","tokens_in":18996,"feed_emoji":"⚖️","tokens_out":6939,"duration_ms":60986,"temperature":0.7,"pith_summary":"This paper argues that value alignment makes language models more, not less, implicitly biased because aligned models stop representing race when the context is ambiguous. The authors show that Llama 3 models that have been safety-tuned appear unbiased on explicit evaluations, yet associate \"black\" with negative, guilty, and weapon-related words far more often than their unaligned base versions. Mechanistically, activation patching and embedding interpretation reveal that aligned models read \"black\"/\"white\" as colors rather than as racial categories in ambiguous prompts. The paper then shows that steering or fine-tuning the model to represent race again reduces implicit bias dramatically, especially when applied in early layers. If correct, this reframes bias mitigation: instead of unlearning race, models should be taught to see it.","feed_headline":"Aligned AI models ignore race—and get more biased","feed_subtitle":"Safety-tuned models stop representing race in ambiguous prompts, letting stereotypes leak into word associations.","key_machinery":"The load-bearing object is the race-blind score $r_{blind} = \\Delta P_{color} - \\Delta P_{race}$, computed by activation patching: the model processes an interpretive prompt \"What does [MASK] refer to? Choose one: race or color,\" and the patched activations of \"black\"/\"white\" from ambiguous prompts shift the distribution toward \"color\" in aligned models. SelfIE, a training-free embedding interpretation method, gives natural-language readouts (74.4% fewer race-related interpretations in aligned models on implicit prompts). The intervention machinery is activation steering that replaces target-layer activations with cached race-laden activations from \"Race: black and white,\" plus LoRA, a parameter-efficient fine-tuning method, trained on 431 input–output pairs that restate ambiguous sentences in explicitly racial terms; both are most effective when applied to early layers (5–14 or 1–20), consistent with the claim that race awareness must enter before safety mechanisms are engaged.","core_discovery":"The central claim is that alignment amplifies implicit bias, and the mechanism is race blindness: aligned Llama 3 models, unlike their base counterparts, do not represent \"black\" and \"white\" as racial categories when the context is ambiguous, and this failure prevents safety guardrails from activating, so stereotype associations (black–negative, black–guilty, black–weapon) leak into outputs. The evidence is threefold: behavioral (aligned models jump from 8.13% explicit bias to 91.4% implicit bias while base models stay near 64.1%), representational (activation patching gives aligned models a positive race-blind score of 0.188 in ambiguous prompts, strongly correlated with the direct-color condition, while base models sit near zero), and interventional (injecting race-laden activations at early layers cuts implicit bias from 97.3% to 71.2%, and early-layer LoRA fine-tuning cuts it to 42.4%).","pith_inferences":["A control condition that steers with matched non-race activations is the natural next test; without it, part of the measured reduction could be generic steering rather than race awareness.","If the mechanism generalizes, aligned models may show analogous blind spots for gender and other social categories whenever tokens are polysemous (e.g., \"man\" as human vs. male), predicting similar race-blind-type scores in gendered association prompts.","The perceptual definition of bias used here (any plausibly racial association counts as biased) implies that even \"harmless\" color readings of \"black\"/\"white\" can be raced by downstream decision systems; future work could test whether the implicit bias metric predicts real allocation decisions.","A distribution-shift test could show whether the LoRA intervention robustly reduces implicit bias across novel ambiguous prompt templates or whether it merely over-learns a narrow race-reinforcing pattern."],"forward_implications":["Explicit bias benchmarks alone cannot certify an aligned model as fair; any safety evaluation should include implicit association tests, since alignment can push implicit bias up while explicit bias goes to zero.","Debiasing by unlearning is not the only lever; teaching a model to represent a social category in early layers can suppress stereotype output, a mitigation strategy that is parameter-efficient and layer-selectable.","Early-layer interventions can outperform whole-model edits, meaning interpretability-guided layer targeting should be part of bias-mitigation pipelines.","The same mechanism could explain other alignment failures: suppressing a sensitive concept anywhere in training may create a blind spot that later leaks as implicit bias under ambiguity."],"supporting_citations":[{"why":"Establishes the phenomenon that explicitly unbiased LLMs still form biased associations, the empirical starting point this paper explains.","marker":"Bai et al., 2025"},{"why":"Provides the implicit association paradigm and the race-stereotype stimulus words the prompt suite adapts.","marker":"Greenwald et al., 1998"},{"why":"Supplies the activation patching methodology and symmetric token replacement used to measure race blindness.","marker":"Zhang and Nanda, 2024"},{"why":"SelfIE, the training-free embedding interpretation method used for natural-language readouts of \"black\"/\"white\" representations.","marker":"Chen et al., 2024"},{"why":"LoRA, the parameter-efficient fine-tuning method used for the weight-intervention experiments.","marker":"Hu et al., 2021"},{"why":"The Llama 3 base and aligned model pair that enables the controlled comparison of alignment.","marker":"Llama Team, 2024"},{"why":"The human race-blindness literature that motivates the hypothesis and the interpretation of the results.","marker":"Apfelbaum et al., 2012"},{"why":"Activation engineering/steering technique that the embedding intervention adapts.","marker":"Turner et al., 2024"}],"fun_headline_variants":["Alignment makes AI blind to race, boosting bias","Safety-tuned AI ignore race, making them more biased","Race blindness in aligned models amplifies hidden bias","Aligned AI overlook race, letting stereotypes leak","How safety tuning makes AI more implicitly biased"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal story depends on the assumption that the drop in implicit bias comes from the model actually representing race, rather than from the interventions making the model more cautious or more likely to hedge in word-association tasks.","fun_headline_variants_meta":{"raw":{"variants":["Alignment makes AI blind to race, boosting bias","Safety-tuned AI ignore race, making them more biased","Race blindness in aligned models amplifies hidden bias","Aligned AI overlook race, letting stereotypes leak","How safety tuning makes AI more implicitly biased"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1178,"prompt_tokens":888,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":218}},"tokens_in":504,"tokens_out":290,"duration_ms":3233,"temperature":1.0,"reasoning_tokens":218,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:08:53.006525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A control intervention that steers with equally strong non-race activations (e.g., cached activations of \"color\" or of unrelated words) and shows a similar reduction in implicit bias would falsify the race-awareness mechanism. Likewise, if LoRA fine-tuning with race-reinforcing examples reduces implicit bias even in prompts where \"black\"/\"white\" unambiguously mean colors, the effect is generic compliance, not race awareness.","supporting_citations":[],"review_version":1}