Ask The PhD Community

HigherEd AI Daily

May 4 – AI in the Exam Room, the Boardroom, and the Pentagon

Monday, May 4, 2026

Three developments this week demand the attention of campus leaders: peer-reviewed evidence of AI outperforming physicians in clinical diagnosis, new data on why most AI deployments fail before they scale, and a federal policy shift that reshapes the vendor landscape for research universities.

The Rundown AI — RESEARCH

AI Outperforms Physicians in Peer-Reviewed Harvard ER Trial

A study published in Science by Harvard researchers tested OpenAI's o1-preview model against two attending emergency room physicians across 76 real patient cases. The AI achieved a correct diagnosis rate of 67.1 percent at initial triage, compared to 55.3 percent and 50.0 percent for the two physicians. Critically, two separate physician reviewers tasked with scoring the outputs could not reliably distinguish AI-generated diagnoses from those produced by the humans.

The study design used only raw electronic health record text as input, deliberately stripping away physical examination and interpersonal cues. In one notable case, the model flagged a rare flesh-eating infection in a transplant patient approximately 12 to 24 hours before the treating physician caught it. The model tested, o1-preview, was released in 2024 and is already two generations behind current frontier models; researchers note that the gap between AI and physician diagnostic accuracy is likely to widen as newer models enter testing environments.

Why it matters for campuses

For health sciences programs, pre-medical education, nursing schools, and medical colleges, this peer-reviewed finding fundamentally shifts the curriculum question. The conversation is no longer "can AI assist clinicians?" but "how should diagnostic reasoning be taught when AI is demonstrably more accurate under controlled conditions?" Beyond clinical programs, the study adds weight to a broader argument that AI has crossed a threshold of expert-level performance in several professional domains; faculty in law, education, business, and the social sciences will be compelled to address the same question in their own curriculum and accreditation conversations.

Read More

Full peer-reviewed study: science.org/doi/10.1126/science.adz4433 (subscription may be required)

The Rundown AI — GOVERNANCE

Why 70 to 80 Percent of AI Projects Never Leave the Pilot Stage

UiPath's Chief Marketing Officer Michael Atalla, in an exclusive interview with The Rundown AI, offered a direct analysis of the dominant pattern in organizational AI failure. Drawing on 15 years at Microsoft, where he guided Office through the shift from on-premise software to cloud-based delivery, Atalla argues that the same mistake that derailed cloud transitions is now derailing AI adoption: organizations lift and shift existing processes into new technology without redesigning the underlying workflows.

The core failure, Atalla contends, is a coordination problem. AI tools are deployed in isolation, disconnected from other tools and from the institutional goals they were meant to serve. Pilots succeed in narrow demonstrations, leadership asks what comes next, no one has a governed answer, and costs accumulate without measurable return. The organizations that break through treat AI tools not as standalone deployments but as components of an observable, integrated workflow. The right starting question, in his framework, is not "which AI tool should we buy?" but "where does work begin, where does it get handed off, and where are decisions actually made?" On employment, Atalla acknowledged that the anxiety in the labor market is real and that entry-level roles are being reshaped; he pushed back, however, on the idea that human involvement becomes optional as AI improves, arguing that new roles in workflow design, AI governance, and process ownership are emerging alongside the disruption.

Why it matters for campuses

This analysis maps with precision onto higher education's AI adoption challenge. Institutions that have piloted AI tools for advising, admissions, tutoring, or student success without integrating those tools into governed academic workflows are likely among the 70 to 80 percent Atalla describes. For academic technology officers, provosts, and faculty senate leaders, the practical implication is that AI governance conversations need to precede tool selection; without a clear map of how decisions are made and workflows are sequenced, even well-funded initiatives will stall at demonstration and never reach institutional scale.

Read More

The Rundown AI — POLICY

Pentagon Adds Eight AI Vendors to Classified Networks; Anthropic Excluded

The U.S. Department of Defense announced partnerships with eight AI companies for access to classified networks: SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Amazon Web Services, and Oracle. Anthropic was excluded despite reporting by the Washington Post that the new contracts carry the same restrictions on autonomous weapons and surveillance for which Anthropic was originally designated a supply-chain risk. The Department of War stated that the agreements will "accelerate the transformation toward establishing the U.S. military as an AI-first fighting force."

The DoD's Chief Technology Officer acknowledged Anthropic's continuing supply-chain risk designation but described its Mythos model as representing a "separate national security moment." The White House is reported to have resisted a broader Mythos rollout due to compute access concerns, even as it expresses continued interest in the model's capabilities. Among the named partners, Reflection is notable for having raised $2 billion from 1789 Capital, a fund backed in part by Donald Trump Jr., a detail that has drawn scrutiny from technology policy observers.

Why it matters for campuses

Research universities with defense contracts, federal grants, or federally funded computing infrastructure need to track which AI vendors hold cleared status and which do not. The exclusion of Anthropic, a company whose public brand is built on safety and responsible deployment, demonstrates that federal vendor selection is governed by factors well beyond a company's alignment with responsible AI principles. Institutions that have built strategic investments around Anthropic tools for research workflows or academic services should be monitoring grant eligibility guidance and research security requirements carefully; the federal vendor landscape is moving quickly, and the implications for campus research compliance are not yet fully mapped.

Read More

(Subscription may be required)

Tool of the Day

Wispr Flow

Wispr Flow is a system-level voice-to-text dictation tool that converts spoken input into clean, formatted text in any application on Mac, Windows, iPhone, and Android, without switching windows or additional configuration. The tool removes filler words automatically, corrects grammar, and builds a personalized vocabulary dictionary from each user's unique terms and phrases. Users report dictating at approximately 220 words per minute, roughly four times the speed of typical typing, with 89 percent of messages sent without any edits; the tool operates natively inside Gmail, Slack, Notion, Canvas, and AI interfaces including ChatGPT.

Try it: Open three student assignments that require written feedback in your campus LMS, then dictate your comments aloud directly into each feedback field. Compare total time spent and the depth of feedback you were able to provide against your typical typed workflow.

Visit Wispr Flow

Dr. Ali Green

Sources for This Edition

The Rundown AI (daily.therundown.ai) – May 4, 2026
The Rundown AI (daily.therundown.ai) – May 3, 2026
Harvard Magazine (harvardmagazine.com)
Science (science.org)
Washington Post (washingtonpost.com)
TLDR (tldrnewsletter.com) – May 4, 2026

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green