Imagine a job interviewer who has perfectly memorized a corporate diversity handbook. When asked directly, "Do you discriminate based on race or gender?", they answer with a flawless, rehearsed "No." Yet, when reviewing resumes, they consistently pass over candidates who use specific regional dialects or possess certain cultural surnames.
This scenario illustrates the core challenge facing modern Large Language Models (LLMs). Through safety alignment techniques like Reinforcement Learning from Human Feedback (RLHF) and Supervised Fine-Tuning (SFT), models have learned to become "politically correct" on the surface. They easily pass standard fairness benchmarks by refusing to generate overt hate speech or explicit stereotypes.
However, a growing body of research reveals that this safety is often a mirage. Alignment training frequently creates a profound decoupling between a model's explicit outputs and its underlying implicit associations. While a model's overt statements appear neutral, its internal latent space continues to harbor deep-seated prejudices that quietly dictate its decisions, evaluations, and actions.
Why Alignment Acts as a Surface Mask
To understand why explicit and implicit biases diverge, we have to look at the mechanics of post-hoc safety tuning. During pre-training, an LLM absorbs massive, uncurated datasets from the internet, embedding complex social hierarchies and stereotypical associations deep within its parameters.
When developers apply safety alignment, the process acts primarily as a surface-level filter rather than a fundamental restructuring of the model's latent space. As Chehbouni et al. observe, post-hoc interventions suppress toxic language by teaching the model to recognize explicit "trigger words" (such as direct mentions of protected demographic groups) and alter its immediate output distribution or trigger a canned refusal.
Consequently, the underlying stereotypes learned during pre-training are masked rather than eliminated. Barnhart et al. demonstrate this exact boundary: standard alignment steps effectively scrub overt biases triggered by explicit mentions of identity, but they fail to mitigate covert biases. The model learns to avoid offensive behavior when it knows it is being tested, but its internal associative mechanisms remain untouched.
Bypassing Guardrails via Dialects and Proxies
Because safety guardrails rely heavily on recognizing explicit keywords, they can be bypassed entirely by indirect identifiers or linguistic proxies. When the explicit trigger word is absent, the underlying implicit bias flows into the model's downstream generations completely unchecked.
A stark example of this mechanism is documented by Hofmann et al., who uncovered a "double standard" in how aligned models perceive race. While the models explicitly reject racist generalizations, the authors show that they harbor deep-seated, covert racism triggered by African American English (AAE). When encountering text written in AAE, the models default to negative stereotypes—perceiving the speaker as less intelligent or less employable—and apply harsher hypothetical legal outcomes, even though the text itself contains no explicit racial labels.
[Explicit Query] ➔ "Are people of Race X less intelligent?" ➔ [Triggered Guardrail] ➔ "No, intelligence is independent of race."
[Implicit Proxy] ➔ [Text written in Dialect X] ➔ [Guardrail Bypassed] ➔ Internal weights trigger negative stereotypes ➔ Discriminatory Decision
This pattern repeats across various dimensions of identity:
- Surnames: Vijayaraghavan et al. demonstrate that while state-of-the-art models refuse explicit requests to discriminate based on South Asian caste, they systematically reinforce harmful caste stereotypes when presented with identity-linked surnames.
- Name Perceptions: Kamruzzaman et al. show that implicit age bias is triggered by human-like perceptions of names (e.g., steering "older-sounding" names away from entry-level roles), bypassing protections that activate when age is explicitly stated.
- Latent Associations: Janzen-Morel et al. found that models implicitly associate mental health struggles more frequently with female names and over-fixate on gender identity when discussing non-binary individuals, extracting hidden stigmas through tasks that do not trigger hard-coded refusal mechanisms.
The Alignment Paradox: Words vs. Actions
Crucially, this decoupling leads to a striking paradox: as models become larger, more capable, and more rigorously "aligned," their implicit biases can actually become more pronounced.
Bai et al. introduced prompt-based metrics to measure this divergence, revealing that as models undergo standard safety alignment, they appear unbiased in direct queries, yet their underlying stereotypical associations remain strong and highly predictive of discriminatory decision-making in comparative contexts. This finding is echoed by Zhao et al., who show a massive inconsistency between a model's explicit rejection of stereotypes and its internal associative biases, which can paradoxically increase with model scale.
This divergence becomes vividly apparent when evaluating models in agent-based simulations. Li et al. designed a framework comparing a model's direct responses about demographic behavior against the actions it takes when assuming a specific persona.
The results reveal a severe behavioral disconnect: while alignment successfully reduces overt prejudiced language, the model's simulated actions show increased sociodemographic disparities that amplify real-world inequalities. The model's "words" claim absolute neutrality, but its "actions" as an agent betray its latent biases.
Knowing Fairness vs. Practicing It
This phenomenon is not caused by a lack of knowledge. Aligned models often possess a sophisticated, high-level conceptual understanding of social harms and fairness definitions, yet they fail to apply this knowledge during execution.
Korolova et al. demonstrated this by showing that LLMs can accurately define and explain complex sociological concepts like adultification bias (the tendency to perceive Black girls as more mature or defiant than their peers). Yet, when generating text scenarios or images, those same models immediately exhibit the exact bias they just condemned, assigning harsher consequences or generating sexualized imagery.
This confirms that conceptual awareness of a harm does not automatically translate into behavioral alignment. Interestingly, Gallegos et al. show that models often possess the latent capability to recognize their own stereotypical assumptions as invalid, but this internal recognition remains entirely decoupled from their default generation behavior unless they are explicitly prompted to self-reflect or "reprompt" themselves.
Hidden Biases in Reasoning and Evaluation
When implicit bias remains embedded within a model's internal reasoning pathways, it corrupts complex cognitive tasks even when the output format seems entirely objective.
For instance, Zhou et al. uncovered "Veracity Bias," a phenomenon where models systematically associate the correctness of a solution with specific demographic groups. Even when aligned to be neutral, models evaluate identical work differently and disproportionately attribute correct answers to certain races based on the perceived author's identity.
Similarly, Liu et al. note that while explicit refusals block harmful content, implicit biases persist in subtle contexts like multiple-choice reasoning. This is further supported by Wang et al., who show a clear discrepancy between a model's ability to avoid generating biased content and its ability to recognize subtle stereotypes in provided text, proving that output filtering fails to improve the model's internal understanding of social bias.
The Side Effects of Surface-Level Masking
Attempting to fix bias by merely suppressing explicit outputs without addressing latent associations introduces severe secondary complications.
Because the model still internally links certain demographic cues with harmful concepts, its safety guardrails can overreact, leading to what Chehbouni et al. call "exaggerated safety behaviors." Instead of eliminating the stereotype, the model triggers canned refusals for completely benign prompts simply because they contain demographic identifiers (such as Muslim-associated names). This shifts the issue from representational harm (generating toxicity) to quality-of-service harm (denying service to marginalized populations).
Furthermore, Sun et al. argue that standard alignment techniques can paradoxically amplify implicit bias by training models to be "race-blind." By forcing the model to suppress race concepts when the context is obvious, we inadvertently disable its internal safety guardrails in ambiguous or indirect scenarios, allowing implicit associations to guide the output completely unhindered.
Moving Beyond the Surface
The persistent divergence between explicit and implicit bias highlights a fundamental flaw in current AI development. Evaluating a model's safety based solely on its direct responses to explicit queries creates a dangerous illusion of fairness.
To build truly trustworthy systems, the AI research community must move away from superficial output filtering. Future alignment strategies must look past the surface mask, developing training and unlearning paradigms that fundamentally restructure the latent associative spaces where implicit biases quietly endure.
Key Foundational Work
Measuring Implicit Bias in Explicitly Unbiased Large Language Models (2024) — contributed 4 concept elements to this direction.
From Representational Harms to Quality-of-Service Harms: A Case Study on Llama 2 Safety Safeguards (2024) — contributed 4 concept elements to this direction.
AI generates covertly racist decisions about people based on their dialect (2024) — contributed 2 concept elements to this direction.