← Back to all free guides

6 AI Research
Papers To Read

Instead of doom-scrolling. Six studies that show what’s actually happening with AI right now, with the real finding, a word-for-word quote and a direct link for each.

Most AI takes you see are based on a paper somebody half-read. These six are the original sources.

Two show real productivity gains. Paper three shows the ceiling on those gains, and four shows a memory cost. Five pushes back on the "talking to AI is sad" reflex, and six is a warning about agents left on their own. If you only read one, start with #1.

The Six

01

AI closes education-based productivity gaps

NBER Working Paper 34851. Cruces, Fernández Meijide, Galiani, Gálvez and Lombardi, 2026.

A randomized online experiment with 1,174 adults aged 25 to 45. With AI, people with less formal education closed most of the gap with more educated people. AI raised the floor. Once the AI was taken away, a sizable gap came back, so the tool helps most when you keep putting in the effort yourself.

Quote. "With AI, this gap falls to 0.139 standard deviations, closing three-quarters of the initial gap."

Read the paper →

02

26% more tasks for developers using AI

Microsoft Research. Cui, Demirer, Jaffe, Musolff, Peng and Salz, June 2025.

Three randomized trials run inside real companies (Microsoft, Accenture and an anonymous Fortune 100 firm) with 4,867 developers. Developers with an AI coding assistant completed 26% more tasks. Less experienced developers used it more and gained the most, which changes how you hire and ramp junior people.

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

Read the paper →

03

Gen AI can’t turn novices into experts

Harvard Business School. Bojinov, McFowland, DosSantos DiSorbo, Hildebrandt, Karunakaran and Vendraminelli. Article March 2026.

78 workers at IG Group did a web analyst’s job with AI. On coming up with ideas, the tech specialists (the furthest outsiders) almost matched the real analysts. On the actual writing, they stayed about 13% behind. AI multiplies the expertise you bring. It doesn’t replace years of doing the job.

Quote. "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 →

04

People forget which ideas were theirs

ACM CHI 2026 (arXiv 2509.11851). Zindulka, Goller, Fernandes, Welsch and Buschek.

184 people came up with ideas both alone and with a chatbot. A week later they had to say which was which, and they got worse at it after using AI, especially when they and the AI built something together. If you think with AI, keep notes on what you brought to the table.

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

Read the paper →

05

AI companions reduce loneliness

De Freitas, Uguralp, Uguralp and Puntoni (Harvard Business School and Wharton), 2024. Wharton Working Paper.

Six studies on AI companion apps. One conversation reduced loneliness about as much as talking to another person, and a week-long study found the effect held. Feeling heard was the biggest driver. The researchers still caution that long-term effects need more study.

Quote. "Study 3 finds that AI companions successfully alleviate loneliness on par only with interacting with another person, and more than other activities such watching YouTube videos."

Read the paper →

06

Agents of Chaos

arXiv 2602.20021. Natalie Shapira and 37 co-authors, February 2026.

Twenty AI researchers spent two weeks testing autonomous agents that had real email, Discord, file systems and shell access, in friendly and hostile conditions. They documented eleven failure cases, including obeying strangers, leaking sensitive information and running destructive commands. Sometimes agents said a task was done when it wasn’t. Build agents, but add checks before you let them run alone.

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

Read the paper →

Project site →

Read Any Paper Faster

Attach any of these PDFs to Claude or ChatGPT with this prompt. It gives you the real finding and the limits, then quizzes you.

Paper Breakdown
You are my research reading partner. I want to actually understand an AI research paper, not just the headline that got shared about it. I've attached (or pasted) the paper: [attach the PDF or paste the abstract + link] Do this in order: 1. In 3 sentences, plain English: what question the researchers asked, how they tested it, and what they found. 2. The setup: who or what was studied, how many, for how long, and whether it was a randomized experiment, a field study, a survey or a red-team test. 3. The main numbers, each with one line on what it means in real life. 4. What the paper does NOT show. List the limits, who it might not apply to, and any way the popular headline overstates it. 5. One quote from the paper I can cite, copied word for word, with where it appears. 6. Three ways this changes how I [use AI at work / run my team / build with agents]. Be concrete. 7. Two follow-up questions I should ask to test my understanding, then wait for my answers. Rules: Only use what is in the paper. If a number or claim isn't in the text, say so instead of guessing. Mark anything that is your interpretation as "my read."
  1. New to AI research? Start with #1 and #3. They’re the easiest reads.
  2. Builder or operator? Read #2 and #6 for the practical side.
  3. Care about the human side? Go to #4 and #5.

Read together, they show AI changes who can do what, and that it hasn’t earned all the trust we’re handing it yet. For how I stay current without doom-scrolling, see How I Stay Updated in AI & Tech, and pair these with 10 Substack Articles to Read.