Ask The PhD Community

HigherEd AI Daily

May 23 – Harvard’s Grading Vote, California’s AI Shield, and the Futures of This Year’s Graduates

Saturday, May 23, 2026

Today’s AI landscape confronts higher education from multiple directions: a landmark Harvard grading vote is drawing national pushback from one of AI’s most prominent voices, new CMU and Stanford research exposes critical blind spots in how AI capabilities are measured against real workforce needs, California has signed the first state-level AI worker protection order in the country, and graduating students are telling campuses something more urgent than their commencement protests may initially suggest.

The Batch (deeplearning.ai) — GOVERNANCE

Harvard Votes to Cap A Grades at 20%; Andrew Ng Calls It a Mistake

Harvard University has voted to limit the number of A grades awarded in undergraduate courses to approximately 20 percent of each class, a direct response to decades of grade inflation that has seen A grades account for well over half of all undergraduate marks at elite institutions. The policy reverses a long-running trend and places Harvard at the center of a national debate about what grades are actually for.

Andrew Ng, founder of deeplearning.ai and one of AI education’s most influential figures, published a direct rebuttal in this week’s issue of The Batch. Ng argues the policy “deeply runs counter to how I believe education should be,” drawing a distinction between “Practice Problems,” designed to reinforce learning, and “Assessment Problems,” designed to evaluate mastery. His core objection: excellence in learning is not a zero-sum game, and a policy that requires roughly 80 percent of students to receive less than an A regardless of their actual performance treats education as a sorting mechanism rather than a transformative process.

Ng also challenges the premise that grade caps serve employers or the broader economy, noting that GPA has increasingly weak predictive value for job performance in AI-adjacent fields and that the students most harmed by artificial grade ceilings are often those from under-resourced backgrounds who have the least margin for error in academic records.

Why it matters for campuses

Faculty governance bodies and academic affairs leaders at institutions watching Harvard will likely face renewed pressure to revisit their own grading policies; this debate is accelerating precisely as AI tools make it harder to attribute student work to individual effort, giving grade-reform advocates a new argument and mastery-learning advocates a new counterargument. How campuses resolve this tension will shape academic integrity policy for years.

Read More
(subscription may be required)

The Batch (deeplearning.ai) — RESEARCH

CMU and Stanford Find That AI Benchmarks Miss Most of What Workers Actually Do

A new study from Carnegie Mellon University and Stanford University, led by researcher Zora Z. Wang, mapped over 10,000 examples drawn from 43 major AI agent benchmarks against the U.S. Department of Labor’s O*NET occupational database. The finding is striking: the benchmarks used to evaluate AI capability are heavily skewed toward computer and mathematical occupations, which represent 5.2 million U.S. workers, while dramatically under-representing office and administrative support roles (18.2 million workers) and management (11 million workers).

The economic stakes make the mismatch even more consequential. U.S. employers spend an estimated $1.33 trillion annually on management roles and $870 billion on office and administrative support; by contrast, the computer and math sector that AI benchmarks overwhelmingly test against accounts for roughly $564 billion in annual employer spend. No single benchmark covered more than 50 percent of real-world work activities; the best-performing benchmark in the study, GDPval, covered 47.8 percent of work activities and 58.5 percent of workplace skills. All 43 benchmarks combined covered only 56.5 percent of real work activities.

The researchers used Claude 3.5 Sonnet to assist in mapping benchmark examples to O*NET descriptions, a methodological choice that underscores how AI tools are now integral even to research that evaluates AI’s own limitations.

Why it matters for campuses

Curriculum designers, workforce development directors, and career services professionals should treat this research as a corrective lens: the roles most represented in AI benchmark headlines are not the roles most at risk from AI substitution or most in need of AI-augmentation skills. Preparing students for AI-transformed workplaces requires looking past benchmark performance toward the full breadth of what workers in management and administrative support actually do each day.

Read More
(subscription may be required)

The Rundown AI — POLICY

California Signs First State-Level AI Worker Protection Executive Order

Governor Gavin Newsom signed an executive order on May 21 directing California state agencies to study and develop AI worker protection policies, becoming the first state governor to issue a comprehensive AI labor directive. The order arrived one day after Meta announced it would lay off 8,000 employees to offset the cost of its AI investments; more than 70,000 technology sector jobs have been lost in California and nationally in 2026 to date.

The order sets specific deadlines: an AI job-impact dashboard must launch within 90 days; proposed updates to California’s WARN Act, which requires advance notice of large layoffs, must be delivered within 180 days to require faster notification in AI-driven displacement events. Agencies are also directed to review union AI negotiations across the state and to begin exploring policy mechanisms for directing a share of AI-generated productivity revenue toward public benefit programs, with initial recommendations due by October 15.

The executive order also directs study of compensation structures including severance, stock options, worker ownership models, and universal basic capital as potential tools for distributing AI productivity gains more broadly. California is home to 33 of the world’s top 50 AI companies, giving state policy particular weight in shaping national norms.

Why it matters for campuses

Campus career services directors, workforce development offices, and institutional research teams should monitor the AI job-impact dashboard closely once it launches; it will provide the first state-level real-time data on AI displacement by occupation and sector, directly informing how higher education programs should realign curriculum and placement support. Institutions with labor-management agreements should also begin assessing whether their own contracts address AI deployment in ways consistent with the direction California policy is heading.

Read More

TLDR AI — ACCESS

Gen Z Is Not Booing AI at Commencement; They Are Booing Their Own Job Market

Commencement ceremonies across the country this spring have featured an unusual addition: audible booing or organized demonstrations when AI is mentioned by speakers. Media coverage has framed these moments primarily as anti-AI sentiment among a generation of digital natives, but new analysis reported by The Next Web points to a more specific and measurable grievance underlying the protests.

The unemployment rate gap between entry-level workers and experienced workers has widened sharply in the post-pandemic period, with the steepest increases concentrated in occupations identified as most exposed to AI substitution. In plain terms: new graduates are entering a labor market where AI has reduced demand for the exact roles that have historically served as career entry points, while the experienced workers competing for those same positions have the advantage of demonstrated performance records that AI tools cannot yet replicate.

The analysis suggests campuses that interpret commencement protests as primarily philosophical or ideological in nature may be misreading the signal; the students expressing the loudest opposition to AI are often those with the most immediate, material stake in how AI reshapes hiring in the next one to three years.

Why it matters for campuses

Academic advisors, career services staff, and deans of students need frameworks that acknowledge the material job market pressures behind student AI anxiety rather than treating it as a cultural or ideological stance; institutions that respond with AI literacy programming alone, without addressing the structural job placement challenges graduates face, are likely to find that student distrust of campus AI initiatives deepens rather than resolves.

Read More

Tool of the Day

NanoClaw

NanoClaw is a secure, lightweight, open-source AI assistant built as a privacy-first alternative to proprietary agent platforms like OpenClaw and Claude Code; it runs locally or on institution-controlled servers, generates no external telemetry, and keeps all inputs and outputs within the institution’s own infrastructure. Developed for environments where data governance and regulatory compliance are non-negotiable, it is particularly well suited to higher education institutions navigating FERPA obligations and research data security requirements. For departments exploring AI tools for student-facing work or sensitive research support, NanoClaw offers a deployable option that does not route academic content through third-party commercial systems.

Try it: Install NanoClaw on a department server and use it to help a research methods course build a document analysis assistant that reads and summarizes uploaded PDFs of journal articles without sending any student-submitted content outside the institution’s network.

Visit NanoClaw

Dr. Ali Green

Sources for This Edition

The Batch, deeplearning.ai (deeplearning.ai)
The Rundown AI (daily.therundown.ai)
TLDR AI (tldrnewsletter.com)
The Next Web (thenextweb.com)
Office of Governor Gavin Newsom (gov.ca.gov)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green