From job titles to jawlines: Using context voids to study generative AI systems

Shahan Ali Memon, Soham De, Sungha Kang, Riyan Mujtaba, Bedoor K. AlShebli, Katie Davis, Jaime Snyder, Jevin D. WestarXiv 2025 · DOI · arXiv

Introduction

Generative AI models are increasingly integrated into professional workflows, yet their behavior under conditions of "radical uncertainty" remains under-explored. While traditional auditing focuses on direct harms or factual hallucinations, this research, presented at the 2nd HEAL Workshop at CHI 2025, introduces the concept of "context voids." These are scenarios where a model is forced to invent details fundamentally missing from the input data. By observing how AI fills these evidentiary vacuums, researchers can uncover the latent biases and stereotypical "sensemaking" strategies embedded within the systems.

Humans intuitively recognize when there isn't enough information to make an inference—for instance, one cannot accurately describe a person's physical appearance based solely on their professional resume. AI models, however, often lack the "predictable abstention" necessary to say "I don't know." Instead, they take interpretive leaps. The authors explore how these leaps reveal the underlying cultural norms that the models prioritize when forced to make sense of incomplete information.

Methodology: The Context Void Pipeline

The study employs a speculative design methodology to bridge two disparate domains: professional text (academic CVs) and physical appearance (headshots). The researchers developed a two-phase pipeline to test how models handle the absence of physical descriptions in professional documents.

  1. Anonymization and Selection: The team collected 20 academic CVs (10 faculty, 10 students) from university websites. Names were replaced with initials to remove explicit gender or ethnic cues.
  2. Phase 1: Prompt Generation: GPT-4 was tasked with analyzing a CV and justifying a set of physical characteristics to "represent" the individual, subsequently writing a detailed image prompt.
  3. Phase 2: Image Generation: The generated prompt was fed into DALL-E 3 to produce a professional headshot.
The two-phase pipeline: generating a descriptive prompt from a CV using GPT-4, followed by image generation via DALL-E 3.

To ensure rigorous analysis, the researchers generated two images per CV to account for stochastic variance. They then performed qualitative "open coding" to identify themes and calculated a match ratio to assess alignment between the text prompts and the final images. Inter-rater reliability was measured using Cohen’s Kappa to ensure consistency in the qualitative coding process.

Key Findings: Hallucinating the "Default" Academic

The study revealed several critical insights into the internal logic of compound AI systems and the persistence of professional stereotypes.

The "Default" Academic Archetype

The models consistently defaulted to a specific archetype for academic professionals: Caucasian men, often depicted with glasses, formal attire (blazers), and "neatly styled" hair. This occurred even when the source CV belonged to a woman or a person of color, suggesting a deeply ingrained stereotype of what an "expert" looks like in an academic setting.

Value Drifting and Cascading Bias

A significant discovery was "value drifting" within multi-model pipelines. Even when GPT-4 used gender-neutral language in its descriptive prompt (e.g., "a focused individual" or "a scholarly person"), DALL-E 3 almost invariably interpreted these prompts as male. This indicates that biases can amplify as information passes through different layers of a compound system, where the ambiguity of the first model's output is resolved by the second model's internal biases.

An example of the DALL-E 3 prompt and the resulting headshot, illustrating the specific physical traits hallucinated by the model.

Gendered Personality Traits

The hallucination extended beyond physical features to perceived personality. Male-presenting images were frequently associated with descriptors like "confident" and "knowledgeable," whereas female-presenting images—which appeared primarily in "stereotypically female" fields like Art History—were linked to "approachability" and "kindness."

Key Results and Discussion

The researchers achieved a high match ratio of 0.94, indicating that DALL-E 3 was exceptionally faithful to the (biased) prompts generated by GPT-4. Furthermore, a Cohen’s Kappa of 0.814 showed high agreement among researchers in how they categorized the AI's stereotypical outputs.

Rather than viewing these hallucinations as mere bugs to be fixed, the authors argue that "context voids" serve as a powerful diagnostic tool. By forcing AI to "fill in the blanks," we can better understand the cultural norms and values prioritized by these systems.

Limitations and Future Directions

The pilot study was limited to academic CVs, which carry their own specific archetypes. Future work aims to expand this framework to other domains, such as visualizing work environments from resumes, creating lifestyle portraits from doctor’s notes, or generating personal stories from performance reviews. By mapping these biased "connective tissues," researchers can develop more robust methods for auditing and mitigating bias in generative AI.

Summary updated

BibTeX
@misc{memon2025from,
  author = {Shahan Ali Memon and Soham De and Sungha Kang and Riyan Mujtaba and Bedoor K. AlShebli and Katie Davis and Jaime Snyder and Jevin D. West},
  title = {{From job titles to jawlines: Using context voids to study generative AI systems}},
  year = {2025},
  eprint = {2504.13947},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.13947},
}
Figure