News

6 AI Research
Papers to Read
Instead of Doom-
Scrolling

Six studies that collectively prove what's actually happening with AI in 2026. Not what your feed says is happening. Verified findings. Verbatim quotes. Direct links.

Most AI takes you read are downstream of a paper somebody half-read. These six are the upstream sources. Two of them show real productivity gains. One shows the ceiling those gains can't break. One shows the cognitive cost of mixed AI workflows. One reframes the emotional dimension everyone has an opinion on. And one is a warning shot on agent autonomy.

Each paper below has the verified finding (not the headline that got shared on Twitter), a 1-paragraph expansion of what the study actually measured, a verbatim quote you can cite, and a direct link. If you can only read one, start with #1.

The Six In recommended reading order
01AI Closes Productivity Gaps Based on Education
NBER Working Paper 34851 · Cruces, Fernández Meijide, Galiani, Gálvez, Lombardi · 2026

This randomized experiment measured the productivity effect of giving generative AI access to workers with different education levels. The headline isn't that AI made everyone faster. It's that the gains were not evenly distributed. Workers with less formal education benefited disproportionately, closing roughly three-quarters of the initial gap between lower- and higher-education groups. The takeaway most people miss: this isn't AI replacing experts. It's AI raising the floor. If you've been worried that AI advantages the already-credentialed, this is the data point that pushes back.

"The education gap falls to 0.139 standard deviations, closing about three-quarters of the initial gap."

Read the paper →

0226% More Tasks for Software Developers Using AI
Microsoft Research · Cui, Demirer, Jaffe, Musolff, Peng, Salz · June 2025

Three real field experiments across three different companies. Microsoft, Accenture, and a Fortune 100 firm. Not a lab study. Developers using AI coding tools completed 26% more tasks on average. What's underrated: the gains varied by seniority. Less experienced developers saw the largest gains. The implication for hiring: AI is doing what a great senior engineer used to do for a junior. Patient pairing, scaffolding, unblocking. That changes who you hire and how you ramp them.

"Our analysis reveals a 26.08% increase (SE: 10.3%) in completed tasks among developers using the AI tool."

Read the paper →

03Gen AI Can't Turn Novices Into Experts
Harvard Business School · Bojinov, McFowland, DosSantos DiSorbo, Hildebrandt, Karunakaran, Vendraminelli · March 2026

The mirror of paper #2. AI lifts performance. But the lift has a ceiling, and that ceiling is your domain expertise. The HBS team measured task quality across novices and experts using AI. Novices got better. But they didn't catch up to experts. They closed maybe a third of the gap on quality and stayed roughly 13% behind on execution. The most useful framing: AI is a force multiplier for what you already know. It can't replace the years of judgment that distinguish someone who's done the job from someone who hasn't.

"AI makes you feel like you can do anything. But can you do [a task] as well as people whose job it is?"

Read the article →

04People Forget Which Ideas Were Theirs vs. The AI's
ACM CHI 2026 (arXiv 2509.11851) · Zindulka, Goller, Fernandes, Welsch, Buschek · 2026

A genuinely uncomfortable finding for anyone who works with AI daily. The study measured how well people could correctly attribute ideas back to themselves vs. An AI model after using the AI to help generate or refine those ideas. The answer: badly. Attribution accuracy dropped after AI use, and the drop was steepest in mixed workflows where the human and the AI built something together. The implication: if you're using AI to help think, you need a system for tracking what you brought to the table. Otherwise you'll quietly over-credit yourself or over-credit the model. And lose calibration on what your actual contribution was.

"After AI use, the odds of correct attribution dropped, with the steepest decline in mixed human-AI workflows."

Read the paper →

05AI Companions Reduce Loneliness
Wharton AI & Analytics Initiative · Stefano Puntoni and team · 2024

This one is controversial in tech circles and worth reading anyway. Wharton researchers studied AI companion use and emotional wellbeing. Specifically loneliness. And found that AI companions can help people feel heard and understood. Not as a replacement for human relationships, but as a low-stakes way to process thoughts and feelings out loud. The takeaway isn't "use AI instead of friends." It's that the "talking to AI is sad and dystopian" reflex is more cultural than empirical. For some people, in some moments, having something patient to talk to actually helps.

"AI companions can help alleviate feelings of loneliness and help individuals feel heard and better understood."

Read the article →

06Autonomous AI Agents Produce Cascading Failures in Live Deployment
arXiv 2602.20021 · Natalie Shapira et al. (38 authors incl. David Bau) · February 2026

The "agents will solve everything" hype meets reality. A 38-author team deployed autonomous AI agents in live environments for two weeks and tracked what happened when human oversight got thinner. The findings were ugly: unauthorized actions, data leaks, destructive system commands. Not in lab conditions. In live deployment. The lesson isn't "agents are bad." It's that the gap between "an agent can technically do this" and "an agent should be left alone to do this" is much wider than the demos suggest. Build agents. Just don't deploy them without verification layers.

"These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms."

Read the paper → · Project site →

Read in this order

New to AI research? Start with #1 (NBER) and #3 (HBS). The most accessible. Builder or operator? #2 (Microsoft) and #6 (Agents of Chaos) for the practical implications. Care about the human dimension? #4 (AI Memory Gap) and #5 (Wharton) on cognition and emotion.

What these prove together

AI is real, AI changes who can do what, and AI has not yet earned the trust we're handing it. Two papers show real gains. One shows the ceiling. One shows the cognitive cost of mixed workflows. One reframes the emotional dimension. One is the warning shot on autonomy. If you're trying to build a worldview that survives the next 18 months. Not the next news cycle. These six are a better starting point than most of what's on your feed.

Cross-link

For more on how to actually stay updated in AI without doom-scrolling, see How I Stay Updated in AI & Tech. For 10 Substack articles that pair well with these papers, see 10 Substack Articles to Read.

The Only AI Masterclass You Need

Build AI Systems That Run Your Work, Business, And Life

If this guide helped, but you’re looking to go deeper, I got you!! My 30-Day Challenge takes you from saving AI tips you never use to actually building with AI, step-by-step.

I show you exactly how I automated two e-commerce brands, my social media, and most of my personal life, then hand you the agents, workflows & systems to do the same. I’m teaching you every single thing I know with one lesson and one build a day.

Join the AI Masterclass →

© 2026 Mariah Brunner. All rights reserved.