|
Ask The PhD Community
HigherEd AI Daily
July 10 – When Take-Home Trust Breaks Down
|
|
Friday, July 10, 2026
Today's central story comes from Brown University, where a single change in exam format cut scores in half and reopened hard questions about how higher ed verifies student learning in the age of generative AI.
|
|
TLDR – GOVERNANCE
Brown Professor's In-Person Final Exposes AI Cheating at Scale
Economics professor Roberto Serrano gave his Brown University class a take-home midterm for the first time in nearly two decades this spring, prompted by student anxiety after a mass shooting near campus. Forty of 86 students scored a perfect 100; the class average was 96. When graders ran the exam questions through ChatGPT, they found the same convoluted argument appearing across dozens of submissions.
Suspecting widespread AI use, Serrano announced the final exam would be given in person. Eighteen students dropped the course and nine more skipped the final; of those 27, twenty-two had scored perfectly on the take-home midterm. Among students who sat for the in-person final, the average score fell to 48, roughly half of what the class had historically produced under exam-hall conditions.
Serrano told Ars Technica that unchecked AI misuse in coursework risks producing "a failed society." His department has not announced formal policy changes yet, but the episode has become a reference point in broader conversations about how take-home assessment holds up against generative AI.
Why it matters for campuses
The Brown case is a clean natural experiment: same course, same instructor, two assessment formats, a 48-point swing. Departments relying on take-home exams, papers, or problem sets without proctoring should treat this as a prompt to revisit assessment design now, not after their own version of this story breaks. Academic integrity offices, provosts, and department chairs will likely face similar data points this fall.
Read More
|
|
TLDR AI – TOOLS
OpenAI Ships a Three-Tier Model Family, With a Coordination Mode for Complex Work
OpenAI has released GPT-5.6 across ChatGPT, ChatGPT Work, Codex, and its API, organized into three tiers: Sol, the flagship model; Terra, a lower-cost option built for everyday work; and Luna, the fastest and cheapest of the three. OpenAI says Terra matches GPT-5.5's performance at half the cost, while Sol posts state-of-the-art results on coding-agent benchmarks while using fewer tokens and less time than comparable models.
The family also introduces an "ultra" setting that coordinates multiple agents across parallel workstreams for demanding tasks, using four agents by default to trade higher token use for faster, more thorough results. Pricing ranges from $1 input and $6 output per million tokens for Luna up to $5 input and $30 output per million tokens for Sol.
Why it matters for campuses
Tiered pricing means institutions can match model cost to task, using Luna for routine drafting or triage and reserving Sol for higher-stakes research or data work, without committing an entire budget to the most expensive tier. IT and procurement teams evaluating campus-wide AI licenses should note the multi-agent "ultra" mode, since coordinating several agents on one task raises new questions about output review and academic use policies.
Read More
|
|
TLDR FOUNDERS – ACCESS
AI-Native Startups Are Hiring Fewer Entry-Level Graduates, Harvard Study Finds
A working paper from Harvard Business School and INSEAD examined Y Combinator startups from 2020 through 2024 alongside a broader set of venture-backed firms. AI-native companies, defined by using AI internally and embedding it in products, are on average 25 percent smaller than non-AI peers, employ 13 percent more engineers, and carry roughly 15 percent fewer entry-level workers and managers.
Senior-level hiring at these firms runs about 20 percent higher than at non-AI peers, and their new hires skew more senior, more elite-educated, more concentrated in Silicon Valley, and more male. Valuations remain comparable to non-AI peers, meaning these firms are generating similar value with fewer, more experienced people.
Why it matters for campuses
This complicates the pitch that AI fluency alone will open doors for new graduates. Career services offices and academic advisors should factor this into guidance for students targeting AI-forward employers, since credentialing and pedigree appear to matter more, not less, at these firms, and entry-level pipelines into the sector may keep narrowing. Equity-minded advising will need to account for graduates without elite-institution backgrounds facing a tougher on-ramp.
Read More
|
|
TLDR PRODUCT – RESEARCH
What Will Separate People Who Thrive From Those Who Don't in the AI Age
Writing in The Atlantic, David Brooks argues that as AI makes intelligence abundant and cheap, the trait that will differentiate people is not raw smarts but their relationship to mental effort. Drawing on psychological research into "need for cognition," Brooks distinguishes people who seek out and enjoy hard thinking from "cognitive misers" who default to the path of least resistance, including letting AI do their thinking for them.
Brooks argues that those who will thrive are the ones who use AI to work harder and reach further, wrestling with it rather than outsourcing to it, and who deliberately cultivate curiosity, discipline, and stamina for difficult problems.
Why it matters for campuses
This gives faculty a useful frame for course and assignment design: the goal is not to ban AI outright or to let it replace effort wholesale, but to structure coursework so students are rewarded for wrestling with hard problems rather than shortcutting them. Instructional designers building AI-inclusive assignments can use "need for cognition" as a design principle, not just a diagnostic for cheating.
Read More
|
|
Tool of the Day
Claude Cowork on Web and Mobile
Claude Cowork, Anthropic's agentic task tool, is now rolling out to web and mobile in addition to desktop, starting in beta with Max plan users. Sessions and files run remotely, so a long task started on a laptop can be checked, adjusted, or picked back up from a phone. For educators and researchers who kick off multi-step tasks such as literature scans, rubric-based feedback passes, or grant document assembly, this removes the requirement to stay tethered to one machine while the work runs.
Try it: Start a Cowork session to draft first-pass feedback on a batch of ungraded assignments using your rubric, then check progress and approve results from your phone between meetings.
Visit Claude Cowork
|
|
Have a great learning day!
Dr. Ali Green
|
|
Sources for This Edition
TLDR / TLDR AI / TLDR Founders / TLDR Product (tldrnewsletter.com)
Ars Technica (arstechnica.com)
TestingCatalog (testingcatalog.com)
TheNextWeb (thenextweb.com)
The Atlantic (theatlantic.com)
Claude Support (support.claude.com)
|
|
askthephd.com
|
askthephd.substack.com
|
Unsubscribe
HigherEd AI Daily; Curated by Dr. Ali Green
|
|