In linguistics, markedness describes the structural asymmetry between a default, neutral expression (the unmarked category) and a divergent form that requires explicit modification (the marked category). For example, in English, the term "actor" historically functioned as the unmarked default for anyone in the profession, while "actress" was explicitly marked to denote a female performer.
Recent research demonstrates that large language models (LLMs) and vision-language models inherit this exact structural asymmetry. When given a completely neutral prompt—free of any explicit demographic indicators—a model does not generate a statistically representative cross-section of humanity. Instead, it operates from an implicit default persona.
As Aylin Caliskan et al. demonstrate in Markedness in Visual Semantic AI, multimodal models like CLIP treat specific dominant demographics (such as White, middle-aged men) as the unmarked representation of a "person." If the model is asked to identify or generate a "person," it defaults to this profile; other demographic groups are treated as deviations from this baseline and are only recognized when explicit, marked descriptors are added to the prompt.
Why Models Adopt a Default Persona
To understand why language models adopt a default identity, we have to look at how they are trained and aligned. The bias is not an accident; it is a mathematical consequence of data distribution and optimization.
First, the pre-training data harvested from the internet disproportionately reflects the perspectives of digitally dominant populations. As noted by Badr AlKhamissi in Investigating Cultural Alignment of Large Language Models, models naturally align more closely with individuals from higher socioeconomic classes and education levels because these groups produce the vast majority of text in digital training sets.
Second, post-training alignment techniques like Reinforcement Learning from Human Feedback (RLHF) often compress the model's perspective rather than broadening it. In Whose Opinions Do Language Models Reflect?, Shibani Santurkar et al. show that human feedback loops tend to shift the model's default opinions and moral stances toward the specific demographic profile of the crowdworkers and engineers performing the evaluation—frequently younger, wealthier, and more highly educated individuals.
When these aligned models are deployed for downstream tasks, this default persona remains active under the hood:
- Data Annotation: When used as synthetic annotators, models do not act as neutral observers. Johannes Schäfer et al. show in Which Demographics do LLMs Default to During Annotation? that state-of-the-art models consistently mimic the subjective perspectives of specific dominant demographics by default.
- Value Alignment: Similarly, Michael Galarnyk et al. demonstrate a persistent alignment bias toward the social and moral norms of affluent, educated populations, effectively marginalizing the values of working-class or older demographics.
The Failure of Explicit Steering: Caricature and Stereotyping
If a model inherently relies on an unmarked default, what happens when we explicitly instruct it to adopt a different persona? For instance, what if we prompt it to "respond as a low-income voter" or "simulate a non-binary teenager"?
This is where the underlying mechanism of unmarked category bias reveals its second major flaw: caricature. Because the model treats minority or marginalized identities as statistical deviations from its internal baseline, it lacks a nuanced, multi-dimensional representation of those groups. To satisfy a prompt requesting a specific persona, the model must actively inject identity signaling.
In CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations, Myra Cheng et al. formalize this phenomenon. They find that while default personas (like White males) require minimal identity signaling and exhibit low caricature scores, marginalized personas suffer from essentialization. The model flattens the persona's entire behavior around a single demographic label, exaggerating traits at the expense of the actual topic being discussed.
[Neutral Prompt] ---------> [Model Default Persona] (Nuanced, multi-dimensional)
|
(Treats minority as deviation)
v
[Socio-demographic Prompt] -> [Statistical Caricature] (Flattened, exaggerated stereotypes)
This over-exaggeration occurs because pre-trained demographic associations override explicit, nuanced context. As Andy Liu et al. demonstrate in Evaluating Large Language Model Biases in Persona-Steered Generation, models struggle to represent "incongruous" personas—individuals whose beliefs do not perfectly align with demographic stereotypes (such as a political liberal who opposes environmental regulations). Instead of simulating an authentic human perspective, the model defaults to a statistical caricature derived from its training data, substituting generalized stereotypes for genuine individual reasoning (Tilman Beck et al.).
Implicit Triggers and Factual Inconsistency
Unmarked category bias is not just a problem for explicit persona prompting; it actively distorts standard user interactions through implicit triggers. Models frequently use subtle linguistic cues, stylistic patterns, or personal names as proxies for demographic categories.
For example, Haozhe An et al. demonstrate that language models cluster names by race and gender in their embedding space. This internal categorization allows models to apply broad group-level stereotypes to individuals even when no explicit demographic information is provided in the prompt.
Once a demographic proxy is triggered, the model's output can shift dramatically, leading to what Matthew Kearney terms identity-contingent factual inconsistency. In Language Models Change Facts Based on the Way You Talk, the author shows that models alter objective, factual advice—such as medical triage recommendations or legal eligibility criteria—based entirely on the inferred demographic profile of the user.
This creates a profound equity issue in downstream applications:
- Neutral Queries: When a query is completely neutral, the model optimizes its response for the dominant demographic group, marginalizing minority users by default (Jizhi Zhang et al.).
- Inferred Minority Queries: Conversely, when the model infers that it is interacting with a minority or vulnerable user, it often underperforms, providing lower-quality information or relying on patronizing, condescending language (Elinor Poole-Dayan et al.).
Summary of Intuition
Unmarked category bias means that language models treat a specific, privileged demographic as the baseline human state. Deviations from this baseline are not treated as alternative, equally rich human perspectives, but as statistical anomalies to be filled in with coarse stereotypes. Understanding this mechanism explains why LLMs simultaneously marginalize minority viewpoints in neutral settings and caricature them when explicitly steered.
Key Foundational Work
It's a Man's Wikipedia? Assessing Gender Inequality in an Online Encyclopedia (2015) — contributed 2 concept elements to this direction.