The OS powering Assistantly
All issues

Intelligence Brief

Week of
October 26, 2026

Your weekly edge, the research, frameworks, and signal that separate operators from spectators. Read and summarized so you do not have to.

Published Monday, October 26 at 6:00 AM PT. 12 sources, 8 minute read.

01

The Human Intelligence Edge

What makes humans irreplaceable in the AI era.

ACM HCOMP '26September 27, 2026

How AI Assistance Affects Human Skill Development: A Study of Learning with Logic Puzzles

Shang Wu, Catarina Belem, Shuyuan Fu, Mark Steyvers, and Padhraic Smyth

Why it matters. 124 people solving logic puzzles, and the variable that tracked later skill gain was time spent working alone before asking for help, not how often help was asked for. The task is a puzzle and the window is short, so nobody should read durable expertise into it. What it does is name something a work sample can actually measure: whether a person pushes a problem far enough to know what they are asking for. A candidate who reaches for the model first and a candidate who reaches for it third can hand you the same artifact at the end of the same hour. Watch the order, not the output.

Read the source
Wharton Executive EducationMay 2026

Thinking Fast, Slow, Artificially: AI and Your Brain

Steven Shaw and Gideon Nave

Why it matters. Three experiments, more than 1,300 participants, nearly 10,000 trials: accurate AI advice raised performance, and confidently wrong advice pushed accuracy below where people started unaided. That asymmetry is the part to plan around, because the second condition is the one no dashboard flags. The people who resisted bad advice scored higher on analytical inclination and reasoning ability, which is a correlation inside a lab task and not a hiring instrument. It still tells you what an interview should contain. Hand the candidate an output that is confident and wrong inside their own domain, and see whether the error survives the hour.

Read the source
Cognitive ProcessingJune 30, 2026

Cognitive offloading, critical thinking and attitudes towards AI

Purvi Jain, Deshna Shajit, Ansh Maini, Yasmeen Shahin, and Ayush Panda

Why it matters. 120 young adults, and the AI users identified more logical fallacies while reporting less mental effort. Both halves have to be read together. A better score at lower effort can mean the tool carried the reasoning, and the qualitative side of the study found behavior ranging from full delegation to deliberate back-and-forth underneath similar results. Self-reported effort in a sample this size is thin evidence for anything causal. The usable part is that output quality alone cannot separate those two people, so if you care which one you are hiring, the question has to be what they checked and what they changed.

Read the source
02

The Broken System

Why resumes, interviews, and hiring infrastructure are failing.

ACM FAccT '26June 25, 2026

Resume Screening, Fast and Slow: (Biased) AI Recommendations' Influence on Human Decision Making

Kyra Wilson, Mattea Sim, Anna-Maria Gueorguieva, Soham Chatterjee, and Aylin Caliskan

Why it matters. Each additional 30 seconds spent on a resume was associated with a 3 to 4% increase in selection chances for candidates the AI had not recommended, and biased recommendations moved human choices as well. The number cuts both ways. Human review can pull a candidate back from a bad screen, and only at a pace most funnels are not run at. Thirty seconds is also cheap enough to buy deliberately. If a model is ranking your pipeline, the cost of being able to override it is measured in minutes per candidate, and the alternative is not neutral review, it is ratification at speed.

Read the source
ACM FAccT '26June 25, 2026

Resume-ing Control: (Mis)Perceptions of Agency Around GenAI Use in Recruiting Workflows

Sajel Surati, Rosanna Bellini, and Emily Black

Why it matters. 22 recruiting professionals describing GenAI inside their own workflows, where the tool shaped job definitions and what counted as a strong interview answer while recruiters still believed they held the final call. Interviews with 22 people describe where authority leaks, not how often it leaks, so take the map and leave the prevalence. The leak has a location, and it sits upstream of the decision everyone audits. Who wrote the criteria, and who checked them against the actual work, is a harder question than who clicked approve. Participants named deskilling alongside the efficiency gain, which is the half of the trade nobody prices.

Read the source
ACM EAAMO '24October 29, 2024

The Silicon Ceiling: Auditing GPT's Race and Gender Biases in Hiring

Lena Armstrong, Abbey Liu, Stephen MacNeil, and Danae Metaxa

Why it matters. A two-part audit of GPT-3.5 found race and gender linked patterns in both its resume assessments and the candidate histories it generated, including less experience in the resumes it wrote for women's names. The model is two generations old and the specific numbers will not transfer, which is the wrong reason to shelve it. The method is what carries. A system that can be asked to produce a candidate can be audited by asking it for a thousand and reading what it assumes. A score from an unaudited screen is not evidence of capability, it is evidence of what the screen was built to reward.

Read the source
03

The SupaHuman Standard

The behavioral science of who actually thrives.

Behavioral SciencesJune 23, 2026

The Influence of Different Cognitive Skills on Learning Agility Among Gen Z in Established and Start-Up Companies

Dian Palupi Restuputri, Yassierli, and Ari Widyanti

Why it matters. 270 Gen Z employees across established firms and start-ups, with creativity and problem-solving tied to learning agility in both settings and critical thinking's effect shifting with context. The sample is cross-sectional, self-reported and Indonesian, so generalization is the authors' problem and direction is the part worth keeping. Learning agility keeps resolving into component skills rather than standing as a trait of its own. That matters at the point where a job description asks for someone adaptable, which is a word, next to someone solving an unfamiliar problem in front of you, which is an afternoon. Only one of those can be checked.

Read the source
The Journal of Social PsychologyJanuary 9, 2025

Other-oriented emotional intelligence, OCBs, and job performance: a relational perspective

Sophia Marinova, Smriti Anand, and Haesang Park

Why it matters. 122 supervisor and employee dyads, where other-oriented emotional intelligence related to helping and taking-charge behavior through the quality of relationships with leaders and coworkers, and those behaviors related to performance ratings. The chain is the point: EQ showed up as something colleagues received, not as something the person reported about themselves. Performance ratings are supervisor judgments, so some of what is captured here is being appreciated rather than being effective. That caveat doubles as the hiring instruction. Ask references what the person actually did for them, and drop the inventory that asks a candidate to score their own empathy.

Read the source
COLM 2026, arXivApril 6, 2026

AI Assistance Reduces Persistence and Hurts Independent Performance

Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey

Why it matters. Randomized trials with 1,222 participants: AI lifted performance while it was available, then people did worse without it and gave up sooner on what came next. Persistence falling is harder to explain away than the score falling, because it is a behavior rather than an outcome. These are online reasoning tasks over a short window, not jobs, and nobody has shown the same curve inside a quarter of real work. It still changes what an assisted work sample proves. If your trial allows the tools the role allows, keep one problem in the process that the candidate has to stay with alone.

Read the source
04

The Future of Work, Without the Clichés

Where AI and human collaboration is actually heading.

Stanford HAIOctober 12, 2026

Humans and AI Engage in Meaningful Collaborations

Andrew Myers, reporting on research led by Diyi Yang

Why it matters. 250,000 anonymized Claude conversations, 56% of them attached to consequential work, about half carrying some friction between person and model, and roughly two in three involving a request to explain or teach rather than simply produce. This is one assistant's traffic from a self-selected population, so the percentages describe its users and not the workforce. The shape still rhymes with the rest of this issue. The interaction that produced something worth having had a person arguing with the output, and that is a behavior you can watch in a screen share and will never find on a resume.

Read the source
ACM CHI '26April 13, 2026

Partnering with Generative AI: Experimental Evaluation of Model-Led and Human-Led Interaction in Human-AI Co-Creation

Sebastian Maier, Manuel Schneider, and Stefan Feuerriegel

Why it matters. Two preregistered experiments with 1,126 participants, where question-led interaction beat model-led generation on idea quality while preserving more diversity across ideas and more sense of ownership. Preregistration at that sample size makes this one of the sturdier results here, inside a single co-creation task. The operational read is that the workflow decides whether a person ends up with their own thinking, and the model does not. A team told to ask the model questions and a team told to have it draft will produce different work, and over a year, different people. Set that default in the process instead of leaving it to habit.

Read the source
Scientific ReportsApril 29, 2025

Human-generative AI collaboration enhances task performance but undermines human's intrinsic motivation

Suqing Wu, Yukun Liu, Mengqi Ruan, Siyu Chen, and Xiao-Yun Xie

Why it matters. Four experiments, 3,562 participants in professional task settings: GenAI raised immediate performance, the gain did not survive the return to solo work, intrinsic motivation fell and boredom rose. It is the same pattern as the persistence trials in section three, from a different lab with a different design, which is the closest thing to replication this issue contains. Short experiments cannot tell you what a year of this does to a team. What they justify is a measurement change that costs nothing: track unaided work somewhere in the cycle, or the only number you hold is the one the tool produced.

Read the source
Previous issueWeek of October 19, 2026 All issuesBrowse the archive