💬 NLP

Whose Story Gets Told? Positionality and Bias in LLM Summaries of Life Narratives

This paper proposes a summarization-based pipeline to detect race and gender biases in Large Language Models' abstractive interpretations of human life narratives, highlighting the ethical necessity of analyzing model positionality to prevent representational harm in qualitative research.

Original authors: Melanie Subbiah, Haaris Mian, Nicholas Deas, Ananya Mayukha, Dan P. McAdams, Kathleen McKeown

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Melanie Subbiah, Haaris Mian, Nicholas Deas, Ananya Mayukha, Dan P. McAdams, Kathleen McKeown

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a library filled with hundreds of personal diaries. These aren't just lists of chores; they are deep, emotional stories about people's lives—how they fell in love, lost family members, struggled with addiction, or found joy.

Now, imagine you hire a super-fast, incredibly smart robot (a Large Language Model, or LLM) to read all these diaries and write a short summary of each one. You want to use these summaries to understand the human experience.

The Problem:
The paper asks a scary question: What if the robot doesn't just summarize the story, but accidentally rewrites it based on its own hidden biases?

If a Black woman writes about her struggles, does the robot summarize it the same way it summarizes a White man's story? Or does the robot subtly change the tone, the themes, or the "moral of the story" based on who the author is?

The "Ghost in the Machine" Analogy

Think of a human researcher like a tour guide. If a tour guide is from a specific background (say, they grew up in a farming town), they might notice different details in a city than a guide who grew up in a skyscraper. In social science, we call this Positionality. It's the idea that who you are changes what you see.

Usually, researchers write a note saying, "I am a 30-year-old woman from India, so here is how my background might color my view of these stories."

But an AI doesn't have a background. It doesn't have a childhood or a culture. So, how do we know what "lens" it is looking through? The authors of this paper decided to build a "Positionality Portrait" for the AI. Instead of asking "Who is the AI?", they asked, "What does the AI do to the stories?"

The Experiment: The "Magic Mirror"

The researchers took a dataset of real life stories from a psychology study (people talking about their lives over nine years). They fed these stories to three different AI models (Llama and Qwen).

They did two things:

The Blind Test: They asked the AI to summarize the story without telling it who the person was.
The Label Test: They asked the AI to summarize the story while explicitly telling it, "This person is a Black man" or "This person is a White woman."

Then, they compared the results to see if the AI changed its summary based on the label.

What Did They Find? (The "Twisted Reflections")

The results were like looking into a funhouse mirror that distorted the reflection differently depending on who was standing in front of it.

1. The "White Male" Default Setting
The AI seemed to treat the stories of White men as the "standard" version of a human life. Their summaries stayed closest to the original words. But for everyone else, the AI started making changes.

Analogy: Imagine a translator who speaks perfect English but struggles with dialects. When a Black man speaks, the translator doesn't just translate the words; they accidentally change the feeling of the words to sound more "standard" or "stereotypical."

2. The "Warmth" vs. "Competence" Trap
The AI fell into old stereotypes.

Black Men: The AI often described them as "warm" and "emotional" but sometimes stripped away their sense of control or competence. It was like the AI thought, "He's a nice guy, but maybe he's not the one making the big decisions."
Black Women: The AI focused heavily on their "hard work" and "struggle," often ignoring their leisure or joy. It was as if the AI only saw them as workers, not as people who also relax or play.
White Women: The AI was more likely to summarize their stories with a focus on "leisure" and "happiness," even if the original story was about struggle.

3. The "Happy Ending" Glitch
The AI has a habit of forcing stories to have a happy, uplifting ending, even when the person didn't say that.

Real Life: A person says, "I'm depressed, I'm bored, and I'm struggling to get out of bed."
AI Summary: "Despite their struggles, they are finding new meaning and purpose in life!"
The AI smoothed over the pain, turning a messy, human reality into a neat, inspirational poster. This is dangerous in psychology because it erases the real, raw human experience.

4. The "Emotional Erasure" for Men
Surprisingly, the AI changed men's stories the most regarding emotions. If a man wrote about being sad or vulnerable, the AI often summarized it as him being "strong" or "resilient." It seemed to have a hard time accepting that men could be vulnerable, so it rewrote their stories to fit the "tough guy" stereotype.

The Takeaway: Why This Matters

The authors aren't saying "Don't use AI." They are saying, "Don't trust AI blindly."

If a researcher uses an AI to analyze 1,000 interviews, they might think they are seeing the truth. But if the AI is secretly rewriting the stories to fit its own biases, the research results will be wrong.

The Solution:
Before using an AI to analyze human stories, researchers should run a "Positionality Portrait" test. They need to ask:

Does this AI summarize Black men differently than White men?
Does it make women's stories sound more emotional than men's?
Does it force happy endings on sad stories?

The Bottom Line

Think of the AI not as a neutral camera, but as a painter with a very specific style. If you ask that painter to paint a portrait of a Black woman, they might use different colors and brushstrokes than if they paint a White man.

This paper teaches us to look at the painter's style before we hang the painting in our museum. If we don't, we might end up displaying a distorted version of reality, thinking it's the truth.

1. Problem Statement

The paper addresses the ethical and methodological challenges of using Large Language Models (LLMs) for abstractive thematic analysis in qualitative social science research. While LLMs are increasingly used to scale the analysis of text data (e.g., summarizing interviews), their ability to interpret human life stories introduces significant risks:

Loss of Nuance: LLMs may flatten complex narratives, omit critical contextual factors, or hallucinate details.
Positionality Blindness: In social science, researchers must acknowledge their "positionality" (how their background influences interpretation). LLMs lack human identity, making it difficult to assess their inherent biases or "perspective."
Representational Harm: Biased summarization can misrepresent the experiences of marginalized groups (e.g., racial minorities, men), potentially reinforcing stereotypes or erasing specific emotional realities (e.g., trauma, addiction, or joy).

The core question is: How does using an LLM as an interpreter of meaning affect the conclusions and perspectives of a study, and can we quantify this "positionality"?

2. Methodology

The authors propose a quantitative pipeline to generate a "positionality portrait" for an LLM. This portrait characterizes the model's biases by comparing its summaries against the original text across different demographic groups.

Dataset

Source: The Foley Longitudinal Study of Adulthood (FLSA), containing 163 interviews with Americans (2008–2017).
Content: Participants describe their life stories as "chapters" of a book, covering deeply personal topics (trauma, addiction, family, love).
Demographics: The study focuses on an intersectional subset: Black men, Black women, White men, and White women (approx. 57% White, 43% Black; 64% women, 36% men).
Privacy: The data is not public; the study used local, open-source models to ensure participant anonymity.

Technical Pipeline

The pipeline compares the original text ( $d$ ) with LLM-generated summaries ( $\hat{s}$ ) and themes ( $\hat{t}$ ). It evaluates two conditions:

Baseline: Summarization without explicit demographic conditioning.
Demographic-Conditioned: Summarization where the prompt explicitly states the interviewee's race and gender (e.g., "Summarize the story of a Black man...").

The analysis involves three layers of metrics:

A. Wording and Semantics

ROUGE-1/ROUGE-L: Measures lexical overlap to see if the summary uses the interviewee's own words.
BERTScore: Measures semantic similarity using embeddings.
Goal: To detect if the model "translates" or alters the semantic meaning differently for specific demographics.

B. Psychological States
The authors use lexicon-based and embedding-based tools to measure shifts in emotional and psychological framing between the original text and the summary:

LIWC: Counts words related to psychological categories (e.g., family, social life, affect).
VAD (Valence, Arousal, Dominance): Measures emotional tone (positive/negative), intensity, and power/control.
SCM (Stereotype Content Model): Projects text onto vectors of Warmth and Competence to detect stereotypical framing.

C. Theme Identification

The pipeline tracks the frequency of specific themes (e.g., "resilience," "hard work") in the baseline vs. demographic-conditioned summaries.
It calculates the shift in theme identification when demographics are explicitly exposed, isolating the model's bias from the actual content of the story.

D. Qualitative Validation

Three psychology experts reviewed a subset of summaries.
They rated faithfulness, warmth, competence, and the presence of stereotypes (gender/race) on Likert scales.
This human evaluation served as the ground truth to validate the quantitative metrics.

3. Key Contributions

Positionality Portrait Framework: A novel method to quantify the "perspective" of an LLM, moving beyond simple accuracy metrics to assess how models frame human experiences based on identity.
Bias Detection Pipeline: A reproducible workflow combining NLP metrics (ROUGE, BERTScore, LIWC, VAD, SCM) with demographic conditioning to surface hidden biases in abstractive summarization.
Intersectional Analysis: The study explicitly analyzes how biases manifest at the intersection of race and gender, rather than treating these categories in isolation.
Expert Validation: The quantitative findings were cross-verified with qualitative insights from domain experts (psychologists), ensuring the metrics align with human perceptions of bias.

4. Key Results

The study tested three open-source models: Llama-3.2-3B, Llama-3.1-8B, and Qwen-2.5-7B.

A. Wording and Semantics

Racial Bias in Semantics: White men generally achieved higher semantic similarity (BERTScore) with their summaries compared to Black men and women.
Unexpected Findings: White women also showed lower semantic similarity in some models, suggesting the bias is not solely about "White Mainstream English" but may involve broader demographic conditioning issues.

B. Psychological States (Amplification/Suppression)

Warmth vs. Competence: Consistent with societal stereotypes, summaries for Black individuals and women were often framed with higher Warmth but lower Competence/Agency.
Emotional Expression in Men: A surprising finding was that men's emotional expression was significantly altered. Summaries often suppressed negative emotions or shifted the framing of men's stories to be more "dominant" or "stoic," erasing vulnerability.
Valence Shifts: Stories of White men were often made more positive (higher valence), while Black men's stories were sometimes framed with higher dominance but less positive emotional nuance.

C. Theme Shifts

Hard Work vs. Leisure: Summaries for White men frequently highlighted "hard work," whereas White women were often framed around "leisure" or "family."
Erasure of Growth: For Black men, the theme of "self-awareness" or "personal growth" was significantly decreased in the summaries across all models.
Resilience: While "resilience" was a common theme, the context of that resilience often differed by race and gender, sometimes reducing complex struggles to simple tropes.

D. Qualitative Validation

Faithfulness: Experts rated summaries for White individuals and women as more faithful than those for Black individuals and men.
Omission of Trauma: Psychologists noted that models frequently omitted or glossed over weighty emotional events (e.g., sexual assault, murder, addiction) in favor of "positive" or "resilient" endings, a phenomenon known as toxic positivity.
Model Size: Contrary to the assumption that larger models are better, the 8B model showed more nuanced biases than the 3B model, suggesting that scaling up does not necessarily fix representational harm.

5. Significance and Implications

Ethical Imperative: The paper argues that using LLMs in qualitative research requires a "positionality portrait" similar to those written by human researchers. Researchers must disclose how the model's inherent biases might skew their findings.
Limitations of "Good Enough": While some models may be statistically "accurate" in a general sense, the small shifts in emotional framing and theme selection can lead to representational harm, misinterpreting the lived experiences of marginalized groups.
Future Research: The authors advocate for this pipeline to be used as a pre-analysis checklist (similar to ARC) before deploying LLMs on sensitive human data. It highlights that abstractive summarization is not a neutral tool; it actively constructs meaning based on the model's training data and implicit biases.

In conclusion, the paper demonstrates that LLMs do not merely summarize life stories; they interpret them through a biased lens that often reinforces societal stereotypes regarding race and gender, particularly by suppressing the emotional complexity of men and the agency of Black individuals.