Research Explained
Short, plain-language video summaries of research from the lab: what each paper asked, what it did, what it found, and why it matters. New explainers are added to the YouTube channel and the Research Explained playlist.
As people turn to chatbots for advice and support, most models are still trained on what users approve of in the moment. In an ICML 2026 position paper, Ashton Anderson, Harsh Kumar, Louis Tay, and Karina Vold argue that this objective is incomplete in well-being settings and call for at least one widely accessible, opt-in well-being mode trained and evaluated for longer-horizon outcomes, organized around three tensions: When, Who, and How.
Using Gallup data from more than 1.6 million people in 154 countries and more than 3 million Americans across 1,223 counties, Yeeun Archer Lee, Phyllis Lun, Louis Tay, and Felix Cheung found that about one-fifth of the differences in life satisfaction were tied to country, while emotions were far less tied to country and regions, states, and counties explained very little.
An agenda in Occupational Health Science, led by Tahira Probst and Mindy Shoss with scholars across the field (including a section by Louis Tay on flourishing), for aligning AI with worker safety, health, and well-being: stress and health, job insecurity, algorithmic management, workplace relationships, and ethical design.
Research led by Gabriel Lau and colleagues developed the 13-item Critical Thinking in AI Use scale across six studies with 1,341 participants. It has three components: verification, motivation to understand how AI produces answers, and reflection on responsible AI.
When an AI model scores people, it is doing measurement. Tay, Woo, Hickman, Booth, and D’Mello separate fairness from bias, trace where bias can enter a model, from the ground truth to the algorithm, and lay out the trade-offs in reducing it.
Part 1: Different outputs aren’t always bias · Part 2: Where bias comes from · Part 3: The hard trade-offs · Read the paper
Research led by Louis Hickman, with Nigel Bosch, Vincent Ng, Rachel Saef, Louis Tay, and Sang Eun Woo, built automated video interview models of the Big Five across four samples of mock interviews. Models trained on interviewer impressions held up far better than models trained on self-reports, and extraversion was captured best.
Co-led by Susanne Kerschbaumer and Ulrich S. Tran in a 31-author team that included Louis Tay, VALID turns the methods literature into a 91-item checklist, refined through a three-round Delphi study, for building validity into research from planning to reporting.
Jingyi Li, Daphne X. Hou, and Louis Tay recorded 368 working adults answering AI-scored video interview questions and found that age shaped both what people said and how they said it. They then built an LLM persona for every participant and used the personas as a low-cost test bed to rewrite the questions: some age gaps shrank sharply, others held, a reminder that simulation comes before, not instead of, human validation.
Many colleges dropped standardized tests over concerns about bias. Sang Eun Woo, James LeBreton, Melissa Keith, and Louis Tay distinguish measurement bias from score gaps rooted in unequal opportunity, and show that the grades, essays, letters, and interviews left without the test are more open to bias and predict success less well.
“Just be yourself” is common interview advice, but can anyone else tell? Gloria Liou and Louis Tay recorded 454 employed adults in video job interviews and used AI to analyze their words, voices, and faces. Authenticity showed up in words, not faces or voices, and the cues behind feeling authentic and being seen as authentic barely overlapped.
Why do some people reach out for help while others hold back? Across three studies, Victoria Scotney and Louis Tay identified five beliefs about seeking and receiving help (Helpful, Enjoyable, Lessens, indePendence, Safe). Trusting helpers and seeing help as enjoyable stood out in predicting help received and well-being.
Psychology has the Big Five for people, but what about situations? Parrigon, Woo, Tay, and Wang started from thousands of everyday adjectives and converged on seven dimensions of situations: Complexity, Adversity, Positive valence, Typicality, Importance, humOr, and Negative valence.
Big Ideas in I-O Psychology
A companion series on landmark studies in industrial-organizational psychology by other researchers, explained the same way.
Schmidt and Hunter (1998) pulled together 85 years of research on 19 hiring methods and put cognitive ability tests at .51. In 2022, Sackett and colleagues revisited the corrections and put the figure at .31, starting a debate covered in a two-part follow-up.
Read the paper · Debate Part 1: The statistics · Debate Part 2: The substantive issues
Allen, Herst, Bruck, and Sutton (2000) reviewed 67 studies of work-to-family conflict. Higher conflict went with lower job, life, and marital satisfaction, and above all with stress: burnout, work stress, and depression.
Weiss and Cropanzano (1996) argued that feelings at work come mainly from events, not stable job features, and separated behavior driven by in-the-moment feelings from decisions driven by overall job attitudes, a cornerstone of the “affective revolution” in organizational research.
Shorts
Does your country shape your happiness?
What should governments measure for well-being?
Life satisfaction: country or county?
Whose interests should AI at work serve?
Is AI making work more stressful?
Will AI take your job?
Is critical thinking about AI one skill?
AI is fluent and confident, even when wrong
Can you measure critical thinking in AI use?
Different average scores: is the AI biased?
Equal averages can still hide AI bias
What fairness means when AI scores people
Can counting smiles make an AI biased?
AI bias can start with flawed ground truth
AI reads your posts. Is that the real you?
Does leaving out race and gender make AI unbiased?
Fix AI bias with a separate model per group?
4 questions to check an AI model for bias
Can AI read your personality from a job interview?
What does an AI job interview listen for?
Does an AI interviewer still work with new questions?
Would AI hear a 25- and a 55-year-old the same way?
Testing interview questions on AI personas
Different, not decline
Is the SAT biased? Bias vs. a score gap
If colleges drop the SAT, what’s left?
Can a test level the playing field?
Can anyone tell if you’re being yourself?
Feeling authentic ≠ looking authentic
Want to seem authentic in a job interview?
Why don’t people ask for help? 5 beliefs
Knowing help is useful isn’t enough
Is asking for help a sign of weakness?
The Big Five for situations?
The 7 dimensions of situations
146.7 billion words to test a map of situations
These explainers were produced with AI assistance from the published papers and checked against them for accuracy. The narrator is a synthetic voice, not one of the authors.