Ask The PhD Community

HigherEd AI Daily

May 5 – AI Enters the Exam Room, the White House, and the Research Lab

Tuesday, May 5, 2026

From a peer-reviewed Harvard trial to a new federal oversight framework and a striking forecast about AI's near-term future, today's edition examines how academic and governmental institutions are reckoning with AI's expanding role in consequential domains.

The Rundown AI — Research

AI Outperforms Emergency Room Physicians in Harvard Diagnosis Study

A study published in the journal Science has found that OpenAI's o1-preview model, released in 2024, outperformed two emergency room attending physicians across 76 real patient cases drawn from electronic health records. Researchers evaluated diagnoses at three stages of patient care. At initial triage, the AI reached the correct diagnosis 67.1% of the time, compared to 55.3% and 50.0% for the two physicians respectively. Two independent clinicians tasked with scoring the outputs could not distinguish which responses came from the model and which came from the human doctors.

What makes the findings especially significant is that the study used a model that is now considered outdated relative to current frontier systems. In one case, the AI flagged a rare flesh-eating infection in a transplant patient approximately 12 to 24 hours before the treating physician identified it. The research does not argue for replacing clinicians; it argues for understanding where AI can function as a high-value decision-support tool in high-stakes clinical environments. The study carries substantial credibility precisely because it was peer-reviewed, published in a flagship journal, and conducted at one of the world's foremost academic medical institutions.

Why it matters for campuses

Medical and health sciences programs face an immediate curricular obligation to address this research in courses on clinical reasoning, evidence-based practice, and AI ethics. Faculty across professional schools, including nursing, pharmacy, public health, and social work, must help students understand both the capability and the limits of AI-assisted diagnosis. Institutional review boards and academic health centers should begin examining what responsible AI decision-support looks like in instructional clinical settings before the question is forced on them externally.

Read More

The Rundown AI — Governance

White House Moves to Establish Formal AI Review Before Model Deployments

The New York Times reported this week that the White House is seeking to create a formal review and oversight process that companies would need to complete before publicly deploying new AI models. The proposal would install a federal checkpoint in the model release pipeline, a significant departure from the voluntary commitment framework that has governed AI development in the United States since 2023. Full details of the framework have not been disclosed publicly, though the reporting indicates the administration is responding to concerns about the pace and opacity of model development at leading AI laboratories.

The timing is notable. Leading AI labs are releasing frontier systems on increasingly compressed schedules, and public pressure for accountability has grown alongside AI's expanding footprint in consequential sectors. A federal review process, if enacted, would represent the most substantial regulatory intervention in AI development in U.S. history; its scope, criteria, and enforcement mechanisms will determine how meaningful it is in practice.

Why it matters for campuses

Universities that have built AI-integrated workflows, research programs, and institutional pilots around specific commercial models would be affected by a federal review process if it slows the rate at which new systems become available. Academic institutions should monitor this development as a leading indicator of where regulatory frameworks are heading; faculty governance bodies, general counsel offices, and research compliance officers will want to develop positions ahead of any formal rulemaking rather than react to it after the fact.

Read More (subscription may be required)

The Rundown AI — Research

Anthropic Co-Founder Puts 60% Odds on AI Training Its Own Successors by 2029

Anthropic co-founder Jack Clark published a detailed analysis on his Import AI newsletter assigning 60% or greater odds to AI systems training their own successor models before the end of 2029. Clark grounds the forecast in publicly available benchmark data rather than speculation. His key data point: AI's independent work horizon moved from 30-second tasks in 2022 to 12-hour autonomous work sessions in 2026, with 100-hour runs projected by year-end. The SWE-Bench benchmark, which evaluates real-world software engineering tasks drawn from GitHub, moved from 2% completion with Claude 2 to 93.9% with Mythos Preview in under three years.

Clark is not alone in this assessment. OpenAI has stated a target of deploying an automated research intern by September 2026, and several startups are explicitly oriented toward recursive self-improvement. What distinguishes Clark's analysis is its methodological transparency; by relying on observable benchmarks rather than insider projections, his case invites rigorous scrutiny and is harder to dismiss as motivated reasoning from an AI company insider. The post has circulated widely in research and policy communities.

Why it matters for campuses

If AI systems are on course to conduct and iterate on their own research within the timeframe Clark describes, higher education institutions face foundational questions about the future role of graduate students, postdoctoral researchers, and junior faculty in the research enterprise. The timeline falls within the span of a typical doctoral program. Academic leaders, provosts, and research councils should begin structured scenario planning now, and graduate programs should consider how they are preparing students for a research environment that may look substantially different by the time they defend their dissertations.

Read More

Tool of the Day

Locally AI

Locally AI is a free iPhone application that downloads and runs open-source AI language models directly on the device, with no data transmitted to external servers. Users can select from several open-source models, including Google's Gemma, and bind the model to the iPhone's Action Button for immediate voice-activated access. For educators and researchers who handle sensitive student data, IRB-regulated materials, or proprietary institutional content, Locally AI offers a way to access AI assistance without the privacy and compliance concerns that accompany cloud-based tools.

Try it: Open the app, download the Gemma model, and paste in an abstract from a paper you are writing or reviewing. Ask the model to identify the three central claims and generate two potential critique questions. Use the output to stress-test your own argument or to design a seminar discussion prompt for students.

Visit Locally AI

[CLOSING NEEDED: Dr. Ali Green, please reply with your 2-sentence closing in your own voice and it will be inserted here before the draft is sent.]

Dr. Ali Green

Sources for This Edition

The Rundown AI (daily.therundown.ai) — May 4 and May 5, 2026
Harvard Magazine (harvardmagazine.com)
Science Journal (science.org)
The New York Times (nytimes.com)
Import AI by Jack Clark (importai.substack.com)
The Rundown AI Tools Guide (app.therundown.ai)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green