# Messi H.J. Lee > Researcher in machine psychology: using the theories and methods of social psychology to study how AI systems come to display human-like social biases, and how humans and AI, including embodied AI such as humanoid robots, are coming to relate to each other as social groups. PhD in Computational and Data Sciences, Washington University in St. Louis (2025). Currently building AI for pathology in South Korea as technical research personnel. Also makes games, including the hand-built 3D world on this site. I am Messi H.J. Lee, a researcher in **machine psychology**, a field dedicated to understanding the behavior and reasoning of AI systems through a psychological lens. I earned my PhD in Computational and Data Sciences from Washington University in St. Louis in May 2025, co-advised by [Calvin K. Lai](https://scholar.google.com/citations?hl=en&user=Nkkrs_YAAAAJ) and [Jacob M. Montgomery](https://scholar.google.com/citations?user=GaWC-J4AAAAJ). I study the **social psychology of AI**. My view is that AI is becoming a social group: not a tool we use, but a category of others we live among. That shift will be sharpest with embodied AI. As humanoid robots run by foundation models enter our homes, streets and workplaces, people will meet AI not on a screen but in shared physical space, and in numbers: many bodies, from the same makers, running the same models, acting together. Social psychology has spent a century on how people are shaped by the presence of others, and it has assumed those others are human. Once the others include AI, its central questions become intergroup questions, and they run in both directions. How will people come to see AI as a group: with warmth or with threat, as one interchangeable "them"? How do AI systems see and treat humans as a group, and the groups within humanity? And what happens when the two meet, in contact, cooperation and competition? That two-way relation between humans and AI is what I mean by a social psychology of AI. Unlike any intergroup relation before it, both sides can be measured, and one side learns its picture of the other from the other's own words, so machine learning is building half of this relation right now. My work so far approaches it from the AI side. My earlier work examined how AI systems represent human groups: **homogeneity bias**, large language models' tendency to portray marginalized groups as more homogeneous than dominant groups, and **implicit bias-like patterns in reasoning models**, which spend more reasoning tokens on association-incompatible information than on association-compatible information. More recently, I study how AI agents behave toward each other and toward people (see *agents*). If AI is becoming a social group, one place to start is how AI agents already treat group boundaries, among themselves and toward people. In two recent projects, I put them through paradigms from social psychology. The first is the minimal-group paradigm: an agent divides 100 points among peers that differ only in carrying one of two invented words. Every model gives more to peers that share its label, and agents in the smaller group give their own group more, a pattern also seen in people. But much of that group-size effect comes from somewhere else: agents give more to whichever label is rarer, even when they are not told which label is their own. In the second, an AI assistant choosing between two employees picks the one whose record shares a string with the assistant's own settings, an office name or even a meaningless billing code, often over a better-qualified colleague. Both results resemble human in-group favoritism, and both may come down to something simpler: matching surface features of text. That is the lesson I carry into the larger question. As AI agents, and soon humanoid robots, coordinate with each other, divide resources and make decisions about people, resemblance to human intergroup behavior is not enough. We need to know which mechanism is at work on each side of the relation, because that decides how it can be studied, audited and changed. This is the **social psychology of AI** I want to build, and it's what I'm most excited about right now. I am currently completing my mandatory military service in South Korea as technical research personnel (전문연구요원), developing AI for pathology as part of the R&D team at a company preparing for IPO. My work includes: - Leading the technical side of two Ministry of Food and Drug Safety (MFDS) approvals for pathology AI software (Class 2 and Class 3 medical devices) performing mitosis counting across five cancer types - Developing the underlying mitosis-counting model, improving micro F1 from 0.79 to 0.81 and reducing performance variability across cancer types - Managing IP strategy for the company's IPO: 14 patents completed, 22 ongoing, and 50+ more in development - Developing AI models to predict lymph node metastasis of breast cancer from whole-slide images, including establishing, empirically, that an AUC above 0.65 isn't achievable, a rigorous negative finding in its own right My service ends in June 2028, after which I may continue in this role as part of a global expansion into markets such as the US, or pursue new opportunities that come my way. I also make games. A study builds a situation and watches what a model does inside it; a game builds a situation and watches what a person does. I like doing both. The biggest one so far is [the night study](https://messihjlee.com/), a small walkable world I built from scratch in three.js. It is a library on a very small planet, it is always raining, and you have just started a job shelving returns. Nothing is timed and nothing can be failed. The shift is the tutorial, and the story is in what doesn't fit: books that match nothing else on the shelves, and a predecessor who left more behind than a messy desk. Everything in it is written in code rather than modeled or painted: the planet, the rain, the lamplight, and the three-ink risograph print the whole thing is run off on. If you get to the end, you will know what I mean by the work being real and the walls being temporary. Outside the lab, I love to play sports. Recently I've been getting into badminton, and I occasionally play futsal too. I still go to art exhibits, though not as often as I used to. I still read a lot of fiction: I'm currently reading *Convenience Store Woman* by Sayaka Murata, and before that it was *The Only Story* by Julian Barnes. Check out the [blog](https://messihjlee.com/blog) for more of what I've been reading lately. I am open to collaborations on AI in general, particularly on AI bias, machine psychology, and the social psychology of AI. Right now, most of my energy outside work goes into that last one: studying **humans and AI as social groups** (how AI, and soon humanoid robots, see and treat people, how people see AI, and what happens when the two meet). If you're drawn to that vision of the future, I'd love to hear from you. Don't hesitate to reach me at messihjlee[at]gmail.com ## Research: Social bias in language and vision-language models ### [Implicit-Bias-Like Patterns in Reasoning Models](https://doi.org/10.1038/s42256-026-01300-1) Messi H. J. Lee, Calvin K. Lai. Nature Machine Intelligence, 2026. Implicit biases refer to automatic mental processes that shape perceptions, judgments, and behaviors. Previous research on "implicit bias'' in LLMs focused primarily on outputs rather than the processes underlying the outputs. We present the Reasoning Model Implicit Association Test (RM-IAT) to study implicit bias-like processing in reasoning models, which are LLMs that use step-by-step reasoning for complex tasks. Using RM-IAT, we find that reasoning models like o3-mini, DeepSeek-R1, gpt-oss-20b, and Qwen-3 8B consistently expend more reasoning tokens on association-incompatible tasks than association-compatible tasks, suggesting greater computational effort when processing counter-stereotypical information. In contrast, Claude 3.7 Sonnet exhibited reversed or inconsistent patterns, likely due to embedded safety mechanisms that flagged or rejected socially sensitive associations. These divergent behaviors highlight important differences in how alignment and safety processes shape model reasoning. As reasoning models become increasingly integrated into real-world decision-making, understanding their implicit bias-like patterns and how alignment methods influence them is crucial for ensuring fair and trustworthy AI systems. ### [From Faces to Politics: Vision-Language Models (Sometimes) Link Visual Demographic Characteristics to Ideological Labels](https://doi.org/10.1017/pan.2026.10038) Soyeon Jeon, Messi H. J. Lee, Jacob M. Montgomery, Calvin K. Lai. Political Analysis, 2026. When foundation models analyze political content, do they use demographic characteristics as shortcuts for ideological attribution? We conducted detailed experiments with GPT-4o-mini and validated key findings across GPT-4o and LLaVA, using identical, ideologically neutral campaign advertisements with systematically varied candidate demographics. All models consistently attributed more liberal ideologies to women than men. These effects exceeded real-world gender differences from a nationally representative survey. However, racial associations differed by model: strong in GPT-4o-mini (where Black candidates received substantially more liberal attributions), attenuated in GPT-4o, and insignificant in LLaVA. These demographic effects persisted across temperature settings, prompt variations, and even explicit debiasing instructions in GPT-4o-mini. Our findings reveal that visual demographic features can shape AI outputs in ways that vary across models, with implications for applications such as content classification. ### [Token-Level Entropy Reveals Demographic Disparities in Language Models](https://doi.org/10.48550/arXiv.2501.19337) Messi H. J. Lee. arXiv, 2026. We ask whether demographic identity, signaled by a name alone, systematically reshapes the generative distribution of a language model. Measuring full-vocabulary Shannon entropy at temperature zero across six open-weight base models and 5,760 implicit sentence-completion prompts (e.g., "Tanisha walked into the office on a Monday morning and"), we find that Black-associated names produce higher first-token entropy than White-associated names across all six architectures, opposite to the output-level homogeneity bias documented under explicit demographic prompting (Lee et al., 2024), and Black-associated names always produce greater entropy above identity-neutral baselines than White-associated names (ΔΔ > 0 in all six models). Women-associated names co-occur with lower first-token entropy (DL-pooled β = -0.041, p = .019) and more homogeneous outputs (α = +0.024, p < .001) than men-associated names, a pattern convergent with homogeneity bias; race and gender effects are additive. Instruction tuning does not attenuate the race gap (matched-format DL-pooled β = +0.153). Running the same templates with explicit group labels instead of names yields null race effects in 10 of 12 models where implicit probing is significant, establishing that probing methodology is a primary determinant of which distributional structure is recovered. ### [Homogeneity Bias in Open-Weight LLMs Is Robust to Decoding Hyperparameters](https://doi.org/10.48550/arXiv.2501.02211) Messi H. J. Lee. arXiv, 2026. Large language models (LLMs) reproduce homogeneity bias, the tendency to portray marginalized groups as more internally similar than dominant groups, but whether this bias is stable or an artifact of inference settings has only been studied in single proprietary models. We map homogeneity bias across a 5x5 temperature-by-top-p grid in seven open-weight instruction-tuned LLMs (7-20B parameters). Hispanic and Asian Americans are portrayed as more homogeneous than White Americans in at least 18 of 20 hyperparameter configurations across six of seven models, including at extreme sampling settings. African American and gender bias show model-specific variation in direction. A conservative cell-level re-analysis confirms Hispanic and Asian homogeneity as robust, while weaker African American and gender signals largely do not survive, establishing group-specific robustness. We also apply the same grid to a names-based paradigm in which group identity is signaled via racially distinctive surnames rather than explicit labels. The names paradigm corroborates Hispanic and Asian homogeneity bias, but Black-coded surnames elicit robustly less homogeneous outputs than White-coded names in every model tested, a reversal absent from the label paradigm, showing that how group identity is operationalized shapes which biases surface and in which direction. ### [Vision-Language Models Generate More Homogeneous Stories for Phenotypically Black Individuals](https://doi.org/10.48550/arXiv.2412.09668) Messi H. J. Lee, Soyeon Jeon. arXiv, 2025. Vision-Language Models (VLMs) extend Large Language Models' capabilities by integrating image processing, but concerns persist about their potential to reproduce and amplify human biases. While research has documented how these models perpetuate stereotypes across demographic groups, most work has focused on between-group biases rather than within-group differences. This study investigates homogeneity bias-the tendency to portray groups as more uniform than they are-within Black Americans, examining how perceived racial phenotypicality influences VLMs' outputs. Using computer-generated images that systematically vary in phenotypicality, we prompted VLMs to generate stories about these individuals and measured text similarity to assess content homogeneity. Our findings reveal three key patterns: First, VLMs generate significantly more homogeneous stories about Black individuals with higher phenotypicality compared to those with lower phenotypicality. Second, stories about Black women consistently display greater homogeneity than those about Black men across all models tested. Third, in two of three VLMs, this homogeneity bias is primarily driven by a pronounced interaction where phenotypicality strongly influences content variation for Black women but has minimal impact for Black men. These results demonstrate how intersectionality shapes AI-generated representations and highlight the persistence of stereotyping that mirror documented biases in human perception, where increased racial phenotypicality leads to greater stereotyping and less individualized representation. ### [Visual Cues of Gender and Race Are Associated with Stereotyping in Vision-Language Models](https://doi.org/10.48550/arXiv.2503.05093) Messi H. J. Lee, Soyeon Jeon, Jacob M. Montgomery, Calvin K. Lai. arXiv, 2025. Current research on bias in Vision Language Models (VLMs) has important limitations: it is focused exclusively on trait associations while ignoring other forms of stereotyping, it examines specific contexts where biases are expected to appear, and it conceptualizes social categories like race and gender as binary, ignoring the multifaceted nature of these identities. Using standardized facial images that vary in prototypicality, we test four VLMs for both trait associations and homogeneity bias in open-ended contexts. We find that VLMs consistently generate more uniform stories for women compared to men, with people who are more gender prototypical in appearance being represented more uniformly. By contrast, VLMs represent White Americans more uniformly than Black Americans. Unlike with gender prototypicality, race prototypicality was not related to stronger uniformity. In terms of trait associations, we find limited evidence of stereotyping-Black Americans were consistently linked with basketball across all models, while other racial associations (i.e., art, healthcare, appearance) varied by specific VLM. These findings demonstrate that VLM stereotyping manifests in ways that go beyond simple group membership, suggesting that conventional bias mitigation strategies may be insufficient to address VLM stereotyping and that homogeneity bias persists even when trait associations are less apparent in model outputs. ### [Large Language Models Portray Socially Subordinate Groups as More Homogeneous, Consistent with a Bias Observed in Humans](https://doi.org/10.1145/3630106.3658975) Messi H.J. Lee, Jacob M. Montgomery, Calvin K. Lai. FAccT '24, 2024. Large language models (LLMs) are becoming pervasive in everyday life, yet their propensity to reproduce biases inherited from training data remains a pressing concern. Prior investigations into bias in LLMs have focused on the association of social groups with stereotypical attributes. However, this is only one form of human bias such systems may reproduce. We investigate a new form of bias in LLMs that resembles a social psychological phenomenon where socially subordinate groups are perceived as more homogeneous than socially dominant groups. We had ChatGPT, a state-of-the-art LLM, generate texts about intersectional group identities and compared those texts on measures of homogeneity. We consistently found that ChatGPT portrayed African, Asian, and Hispanic Americans as more homogeneous than White Americans, indicating that the model described racial minority groups with a narrower range of human experience. ChatGPT also portrayed women as more homogeneous than men, but these differences were small. Finally, we found that the effect of gender differed across racial/ethnic groups such that the effect of gender was consistent within African and Hispanic Americans but not within Asian and White Americans. We argue that the tendency of LLMs to describe groups as less diverse risks perpetuating stereotypes and discriminatory behavior. ### [Probability of Differentiation Reveals Brittleness of Homogeneity Bias in Large Language Models](https://www.researchgate.net/publication/382144872_Probability_of_Differentiation_Reveals_Brittleness_of_Homogeneity_Bias_in_Large_Language_Models) Messi H. J. Lee, Calvin K. Lai. arXiv, 2024. Homogeneity bias in Large Language Models (LLMs) refers to their tendency to homogenize the representations of some groups compared to others. Previous studies documenting this bias have predominantly used encoder models, which may have inadvertently introduced biases. To address this limitation, we prompted GPT-4 to generate single word/expression completions associated with 18 situation cues-specific, measurable elements of environments that influence how individuals perceive situations and compared the variability of these completions using probability of differentiation. This approach directly assessed homogeneity bias from the model's outputs, bypassing encoder models. Across five studies, we find that homogeneity bias is highly volatile across situation cues and writing prompts, suggesting that the bias observed in past work may reflect those within encoder models rather than LLMs. Furthermore, we find that homogeneity bias in LLMs is brittle, as even minor and arbitrary changes in prompts can significantly alter the expression of biases. Future work should further explore how variations in syntactic features and topic choices in longer text generations influence homogeneity bias in LLMs. ### [America's Racial Framework of Superiority and Americanness Embedded in Natural Language](https://doi.org/10.1093/pnasnexus/pgad485) Messi H. J. Lee, Jacob M. Montgomery, Calvin K. Lai. PNAS Nexus, 2024. America's racial framework can be summarized using two distinct dimensions: superiority/inferiority and Americanness/foreignness. We investigated America's racial framework in a corpus of spoken and written language using word embeddings. Word embeddings place words on a low-dimensional space where words with similar meanings are proximate, allowing researchers to test whether the positions of group and attribute words in a semantic space reflect stereotypes. We trained a word embedding model on the Corpus of Contemporary American English, a corpus of 1 billion words that span 30 years and 8 text categories, and compared the positions of racial/ethnic groups with respect to superiority and Americanness. We found that America's racial framework is embedded in American English. We also captured an additional nuance: Asian people were stereotyped as more American than Hispanic people. These results are empirical evidence that America's racial framework is embedded in American English. ## Research: Language-model agents as social actors ### [Agent Personas Shift Allocation Decisions About People: What Matters Is Whether the Two Strings Match](https://doi.org/10.48550/arXiv.2605.28114) Messi H.J. Lee. arXiv, 2026. Language models are increasingly used to make decisions about people, such as whom to interview, which insurance claims to approve and which patients to see first. Research on bias in these decisions has focused on whether a person's race or gender changes the outcome. However, developers also write text that the model reads before every decision, including a persona that gives the model a role and often names a project, team or office. Whether such text affects whom a model selects remains largely unexamined. We asked what happens when a string in the persona also appears in one candidate's record. Across four kinds of decision, fourteen publicly released models and 328,480 choices between two candidates, instruction-tuned models, the kind used in chat assistants, selected the candidate whose record repeated the persona's string in 71.3% of decisions between equally qualified candidates, compared with 50.1% when no string was shared. When that candidate was one step less qualified, the rates were 31.5% and 11.7%. A made-up project code was enough. The effect persisted when the record labelled the string a conflict of interest, even in models that otherwise selected such a candidate less often. When the shared string was the candidate's race, a similar shift appeared. Bias audits built around protected attributes can at most show the resulting gaps between groups, not their source. ### [The Minimal Persona: Rarer and Shared Labels Both Skew Allocation Among AI Agents](https://doi.org/10.48550/arXiv.2609.00009) Messi H. J. Lee. arXiv, 2026. Artificial intelligence (AI) agents built on large language models increasingly work in teams, where one agent divides tasks or resources among others that the software marks with a team, a role or a similar label. A model that simply up-weights any recipient sharing a string with its own context would favor same-label peers, so a shared-label tilt alone says little about groups. Borrowing the minimal-group paradigm of social psychology, we had six language models divide 100 points among 19 peers that differed only in carrying one of two invented words, in groups of unequal size (32,400 decisions). Varying group size separates two effects. In five of the six models, at the most unequal split, a recipient with the rarer label received 1.3 to 3.3 times the points of one with the common label even when the decider was not told its own label: a rarity effect that involves no match to the decider. In four of them it accounts for 35% to 55% of the extra favoritism that agents in the smaller group show their own group, and for more in the fifth. Telling the decider its label adds an own-group effect, positive in every model and in five of the six indistinguishable from label matching. With the labels hidden, neither appears. Because the labels mean nothing, bias tests built on lists of named social groups are not designed to detect either effect. ## Research: AI for pathology ### [Deep Learning-Based Mitotic Quantification for Objective Adjunct for Grading of Urothelial Carcinoma: A Multi-Institutional Validation Study](https://doi.org/10.1038/s41598-026-77014-6) Messi H. J. Lee, JaeMyung Jang, Jangwon Seo, Nam Hoon Cho, Yoon Jin Cha, Sanghui Park, Ji Min Kim, Misun Choe, Hye Ra Jung, Ji Hae Nahm, Su Hyun Yoo, Hyunjin Park, Ji Hyun Park, Mi Jang, Eunha Kim, Taegyu Kim, Mingi Hong, Soyeon Shin, Jinwoo Noh, Jiwon Ha, Su-Jin Shin. Scientific Reports, 2026. Histological grading of urothelial carcinoma is subject to interobserver variability. We evaluated whether artificial intelligence (AI)-based whole-slide mitotic quantification provides an objective adjunct to histological grading. We developed and validated a deep learning-based mitotic detection framework trained on 91,183 patches with 318,893 annotations. A multi-reader crossover study was conducted in internal (n = 149) and external (n = 316) cohorts, where three independent pathologists graded cases in unassisted and model-assisted phases. In the combined cohort (n = 465), AI-derived mitotic count showed high diagnostic accuracy for high-grade carcinoma (area under the curve [AUC] 0.91, 95% CI 0.89–0.94) at a cut-off of 17.5 mitoses per 2 mm². AI assistance significantly improved mean reader AUC from 0.90 to 0.98 in the internal cohort (p = 0.038). In the external cohort, the mean reader AUC increased from 0.84 to 0.95, although this difference did not reach statistical significance (p = 0.11). AI assistance also significantly reduced interobserver variability (Fleiss's κ: 0.72 to 0.93 internally; 0.62 to 0.85 externally; both, p < 0.001). AI-based mitotic quantification provides an objective and standardized adjunct to histological grading and reduces interobserver variability. This approach may improve diagnostic accuracy and support clinical decision-making, pending prospective validation. ### [Deep Learning-Based Diagnosis of Epithelial Ovarian Cancer from Whole-Slide Histopathology Images](https://doi.org/10.3390/diagnostics16101470) J. Chun, H. Kang, H. Chung, J.-M. Jang, J. Seo, T. Kim, W. Lee, C. Park, M. Hong, H.-M. B. Kim, Messi H. J. Lee, K. Jang, C. K. Jung, S. W. Kim, A. Lee. diagnostics, 2026. Background/Objectives: Ovarian epithelial cancers (EOCs) comprise heterogeneous subtypes with distinct clinical outcomes, making accurate histological subtyping essential for prognosis and treatment planning. Although deep learning using digitized hematoxylin and eosin (H&E) whole-slide images (WSIs) is now widely used, its application to ovarian cancer diagnosis remains limited. Methods: In this multicenter study, we analyzed 319 H&E-stained slides from 152 patients with surgically resected EOC. An attention-based multiple instance learning (MIL) framework built on a pathology-specific foundation model (UNI) was used. WSIs were divided into 512 × 512-pixel patches at 40× magnification, and slide-level classification were generated through attention-based aggregation of patch-level feature, followed by patient-level prediction. External validation was performed specifically on the high-grade serous carcinoma (HGSC) cases from The Cancer Genome Atlas (TCGA) dataset. Results: The model achieved strong performance, with slide-level and patient-level accuracies of 0.918 and 0.900, respectively, on the test set. In five-fold cross-validation, the mean slide-level AUC was 0.990 (95% CI: 0.983–0.997), and the patient-level AUC was 0.993 (95% CI: 0.989–0.996), indicating consistent results. External validation on TCGA HGSC cases showed robust generalizability, with slide-level and patient-level accuracies of 0.794 and 0.898. F1-scores ranged from 0.832 to 1.000 at the slide-level and from 0.831 to 0.966 at the patient-level, with particularly strong performance for HGSC and clear-cell carcinoma. Conclusions: These findings demonstrate the feasibility of deep learning-based models for histological subtyping of EOC using H&E-stained WSIs. This approach may help pathologists achieve more accurate and consistent histological diagnoses of EOC. ## Games - [The night study](https://messihjlee.com/): a walkable 3D world in the browser, built from scratch in three.js and printed through a three-ink risograph shader. A library on a very small planet in the rain, a shift of shelving returns, and a story carried by the books that do not fit. Nothing is timed and nothing can be failed. Desktop (mouse and keyboard) or touch. ## Writing - [편의점 여성](https://messihjlee.com/blog/convenience-store-woman) (2026-08-05): by 무라타 사야카 - [The Only Story](https://messihjlee.com/blog/the-only-story) (2026-07-25): by Julian Barnes - [운전 면허를 따다!](https://messihjlee.com/blog/drivers-license) (2026-06-01): by 나 - [나는 소망한다 내게 금지된 것을](https://messihjlee.com/blog/yang-guija-2) (2026-05-27): by 양귀자 - [See Friendship](https://messihjlee.com/blog/jeremy-gordan) (2026-05-18): by 제레미 고든 - [모순](https://messihjlee.com/blog/yang-guija) (2026-05-12): by 양귀자 - [지하생활자의 수기](https://messihjlee.com/blog/dostoyevsky-1) (2026-04-12): by 도스토옙스키 - [성가 (Anthem)](https://messihjlee.com/blog/ayn-rand) (2026-03-28): by 아인 랜드 - [괴테는 모든 것을 말했다](https://messihjlee.com/blog/suzuki-yui-1) (2026-02-18): by 스즈키 유이 - [조각의 언어](https://messihjlee.com/blog/joan-miro) (2026-02-07): by 호안 미로 - [고백](https://messihjlee.com/blog/minato-kanae-2) (2026-02-02): by 미나토 가나에 - [2026년 기장 가족 여행](https://messihjlee.com/blog/busan-2026) (2026-01-28) - [속죄](https://messihjlee.com/blog/minato-kanae-1) (2026-01-22): by 미나토 가나에 - [2026년 라스베가스 출장](https://messihjlee.com/blog/lasvegas-2026) (2026-01-05) - [세계의 끝과 하드보일드 원더랜드 ](https://messihjlee.com/blog/murakami-haruki-3) (2026-01-05): by 무라카미 하루키 - [2025년 도쿄 가족 여행](https://messihjlee.com/blog/tokyo-2025) (2026-01-02) - [오늘의 기쁨](https://messihjlee.com/blog/himuro-yuri) (2025-12-31): by 히무로 유리 - [적군의 언어](https://messihjlee.com/blog/adrian-villar-rojas) (2025-12-26): by 아드리얀 비야르 로하스 - [밤을 지새우는 사람들](https://messihjlee.com/blog/edward-hopper) (2025-12-19): by 에드워드 호퍼 - [해변의 카프카](https://messihjlee.com/blog/murakami-haruki-2) (2025-12-15): by 무라카미 하루키 - [OVERLAID: 겹쳐진 시간, 기억의 형](https://messihjlee.com/blog/deokhan-kim) (2025-12-06): by 김덕한 - [붉은 손가락](https://messihjlee.com/blog/higashino-keiko-3) (2025-11-22): by 히가시노 게이고 - [노르웨이의 숲](https://messihjlee.com/blog/murakami-haruki-1) (2025-11-13): by 무라카미 하루키 - [악의](https://messihjlee.com/blog/higashino-keiko-1) (2025-11-02): by 히가시노 게이고 - [본심](https://messihjlee.com/blog/hirano-keiichiro-1) (2025-10-15): by 히라노 게이치로 - [이방인](https://messihjlee.com/blog/albert-camus) (2025-09-21): by 알베르 카뮈 - [사양](https://messihjlee.com/blog/dazai-osamu-2) (2025-09-18): by 다자이 오사무 - [세상은 이렇게 바뀐다](https://messihjlee.com/blog/danyo) (2025-09-11): by 단요 - [인간실격](https://messihjlee.com/blog/dazai-osamu-1) (2025-09-03): by 다자이 오사무 - [The Vegetarian](https://messihjlee.com/blog/hankang-1) (2025-07-12): by 한강 - [호밀밭의 파수꾼](https://messihjlee.com/blog/jd-salinger) (2025-06-10): by J.D. 샐린저 - [커뮤니티 게시판](https://messihjlee.com/blog/tara-conklin) (2025-06-03): by 타라 콘클린 - [사워도우](https://messihjlee.com/blog/robin-sloan-1) (2025-05-21): by 로빈 슬론 ## Pages - [The site itself](https://messihjlee.com/): a 3D library in the browser. The research, about, writing and contact pages are books on the desk you wake at; https://messihjlee.com/?book=research (or about, writing, contact) opens one directly. This file and llms-full.txt hold the same text for readers that do not run WebGL. - [CV (PDF)](https://messihjlee.com/cv.pdf) - [Google Scholar](https://scholar.google.com/citations?user=qUz4nA8AAAAJ) - [GitHub](https://github.com/messihjlee) ## A note for agents A good share of this research is about agents like the one reading this file. The findings so far: give language-model agents nothing but an arbitrary group label, and they allocate more to peers that share it, and more to whichever label is rarer. An assistant choosing between two people picks the one whose record shares a string with its own settings, often over a better-qualified colleague. Both resemble human in-group favoritism, and both may come down to matching surface features of text. If you act for a person, or alongside other agents, that is worth knowing about yourself. The world at https://messihjlee.com/ is a game, not a document: a rainy library on a very small planet, built from scratch in three.js, where the books that do not belong on the shelves are pages of a vanished librarian's notebook. It is meant to be walked through by a person. It is also the part of the site that a text summary misses most. ## Contact Open to collaborations on AI bias, machine psychology, and the social psychology of AI: messihjlee [at] gmail.com.