Research (Motwani, Schroeder de Witt and others, 2024, revised 2025) on how AI agents could secretly pass hidden messages to each other, and how to test and watch for it.
Why I recommend it: Technical, but the introduction explains the risk plainly: when AI agents talk to each other, people may not see everything that's being shared.
Open-access study tracking 183 Canadian co-op students applying for full-time jobs. Applicants whose resumes and cover letters scored higher on detail, clarity, and structure secured substantially more interviews and found jobs faster — even after controlling for experience, achievement, and tailoring.
Why I recommend it: Open-access, peer-reviewed short communication. The key takeaway is that writing quality matters as a hiring signal, but the sample is students in one co-op program, and the authors note this also means the same signals can be generated by tools like ChatGPT.
A free 2025 paper by Mallory Knodel, Sunoo Park, Kyunghyun Cho and colleagues at NYU and Cornell examining whether AI assistants and end-to-end encryption can honestly coexist. It covers two cases — putting an AI assistant inside an encrypted app, and training models on encrypted data — sets out where each breaks the security promise encryption makes, works through the legal consequences when a provider keeps saying "end-to-end encrypted" anyway, and ends with concrete recommendations on default settings, consent and what providers may truthfully claim.
Why I recommend it: Read the recommendations section even if you skip the cryptography. It gives you the exact questions to put to any product that advertises both an AI helper and private messaging — where does the processing happen, what is the default, and what were you actually asked to agree to. Two flags: it is posted to a preprint archive, so it has not been through journal peer review, and the authors published plain-language summaries on the NYU DeTaIL Lab blog and Tech Policy Press if the full paper is heavy going.
Stanford economist Charles I. Jones works out, in plain economic terms, how much money it would be worth spending to lower catastrophic risks from advanced AI — comparing it to the roughly 4 percent of GDP the U.S. effectively spent during Covid-19.
The underlying working paper by Jeremy Yang and co-authors, using Perplexity data to model tasks as discrete steps and compare fixed vs. marginal costs of chatbots versus autonomous agents.
Why I recommend it: If the HBS summary hooks you, go to the source. Skim the task-cost framework and use it to audit your own week: which tasks are high-step and repeatable? Those are the ones to hand to an agent first.
Stanford-led study of 3 million applicants screened by a single algorithm vendor, finding racial disparities and homogeneous rejections — the same people get screened out everywhere. Explains why applicants must apply widely to reach a human.
In plain terms: This research study examines how automated screening tools used by multiple employers cause repeated rejections and racial disparities. Use this paper to understand how hiring algorithms work and why applying to more jobs helps you reach a human reviewer.
Why I recommend it: This is the evidence behind advice I give constantly: one rejection is often the same algorithm repeated, not a verdict on you.
