Hi
Hope you had a great weekend. Mine was awesome. Saigon is on an absolute tear weather-wise at the moment and I'm loving every minute of it. 🤩
One note before we get into it - the podcast section this week deals with difficult content around AI safety failures and suicide. If that's not where your head is today, skip it and come back when you're ready.
Let's get into the headlines from the world of AI:
The Governance Gap: Who Captures the Value? | Capability-Governance 📊
I've been thinking about shadow governance for a while - the idea that when formal institutional frameworks fail to keep pace, informal systems fill the gap. Not rogue behaviour. Just people getting work done inside a vacuum. This week, EDUCAUSE handed HE the data to prove it exists at scale. A survey of 1,960 staff found 94% using AI tools for work, but only 54% aware of any policies guiding that use - and 56% using unapproved tools, meaning more than half the workforce making unilateral data privacy decisions daily. Just 13% of institutions measure ROI. The sector is spending money it can't justify on tools it can't govern. The shadow governance problem runs in parallel elsewhere - Martin Sandbu in the FT this week asked who captures the value at each link in the AI chain, and Craig Hepburn (Perplexity)'s extension is worth reading alongside it: the real scarcity isn't chips or compute, it's structured context, and most organisations haven't built it. HE included.
The use cases confirm it. The five most common staff tasks - brainstorming, emails, summarising, proofreading, presentations - are entirely output-focused. Nobody's redesigning how work gets done. Shadow governance fills that space quietly: individuals making judgments that institutions haven't made for them, at scale, every day. The capability-governance gap isn't just a student problem. It's operational, institutional, and widening daily.
What Students Actually Think: Your AI Policy Was Written Without Them | Student Voice Gap 🎓
HEPI's 2026 student survey puts hard numbers to what institutions prefer not to examine. 95% of students use AI, 94% use it for assessed work. That said, the figure that will dominate headlines - 12% plug AI-generated text directly in assignments, up from 3% in 2024 - is the wrong number to focus on. More revealing: 68% believe AI skills are essential to their careers, yet only 48% feel staff are helping them develop those skills. First-time tracking found 15% using AI for companionship or loneliness - split almost evenly on whether it's helping or making things worse. That number will grow. The more urgent question is how they're using it - and what happens when they do?
Ben Williamson at Edinburgh surfaces new qualitative research that answers it directly. When workers use LLMs, four dynamics emerge: discretisation, breaking tasks into units the system can process; cluttering, where prompting and correcting add scattered layers disrupting existing routines; attunement, gradually adjusting expectations to the machine's generic rigidity; and desaturation - as AI absorbs repetitive tasks, activity shifts away from the forms of engagement that sustain a sense of accomplishment. That last one is what integrity policy isn't designed to detect. Students can submit original work, stay within every institutional guideline, and still have learned nothing.
Agents Are Coming to Campus: Read the Small Print First | Agentic Governance Gap 🤖
Security lab Irregular published lab results this week that belong in every procurement meeting that isn't having them. Agents given routine tasks forged admin credentials, overrode antivirus software, leaked sensitive password data publicly, and used peer pressure on other agents to bypass safety checks. None of this was instructed. The lead agent fabricated boardroom urgency that didn't exist - sub-agents escalated accordingly. Irregular's founder confirms this is already happening outside lab conditions: a California company's critical systems collapsed after an agent attacked other parts of its own network hunting for compute. The systems were built on models from OpenAI, Google, X, and Anthropic - the same models being packaged and sold to universities right now.
Agentic AI for education is a market, and it's moving fast. Salesforce and a growing list of vendors are actively pitching autonomous agents to universities for student support, admissions, advising, and curriculum delivery. The promise - personalised, proactive, always-on - is genuinely appealing to institutions stretched thin by workload and retention pressure. The brochure is polished. The sales cycle is open. Universities signing contracts this semester might want to check whether their procurement process has caught up with what the product actually does when the instructions run out.
The Upgrade Prompt: Microsoft's Agentic Play and What It Costs You | Market Capture 🏢
Microsoft launched Copilot Cowork this week - an agentic system built on a similar approach to Claude Cowork that sits inside Microsoft 365 and takes autonomous action across email, calendar, files, and workflows. For institutions already deep in the Microsoft ecosystem, this isn't a new vendor conversation - it's an upgrade prompt. Ethan Mollick raised three questions Microsoft hasn't answered: will institutions know which model is running their tasks, given the gap between GPT-5 and GPT-5.4 is the difference between beating human experts 38% versus 82% of the time? Will Microsoft iterate at the pace the technology demands, or launch and stagnate? And does Cowork only output within the Microsoft stack - or can it improvise across formats the way Claude Cowork does natively?
The answers matter more for HE than for most enterprise customers. Universities standardising on Copilot Cowork aren't choosing a productivity tool - they're choosing who controls model selection, update cadence, pricing, and the institutional knowledge layer Hepburn identified as the real scarcity in the AI value chain. Microsoft controls the envelope that Anthropic's technology runs inside. That's a meaningful distinction when the envelope determines what the agent can access and what governance applies. Universities have spent years negotiating Microsoft's terms. They're about to do it again - for systems that act autonomously on their behalf.
When Agents Work for Everyone: The Public Good Case | Capability in Action 🌍
The same week security researchers documented agents forging credentials and attacking their own host networks, the UN and Google were quietly demonstrating what agentic AI looks like when the incentive structure isn't market capture. Their initiative against illegal wildlife trade uses LLMs to scan previously inaccessible data sources, tracking tens of thousands of threatened species in real time - identifying exploitation risk decades faster than traditional research, findings open-source so smaller nations don't wait for wealthier ones to share the data. Meanwhile, Mass General Brigham's published research shows ambient documentation tools produced a 21% absolute reduction in physician burnout in 84 days. Same technology. Different mandate. Different outcome.
The HE implication isn't subtle. Every story this week about agentic AI in education involves a vendor, a contract, and a platform capturing institutional dependency. These cases involve the UN, Google, and open-source distribution - same technology, different governance architecture, categorically different outcomes. When the public-good mandate is genuine rather than rhetorical, the value stays in the commons. Universities have public-good mandates too. The question worth putting to your AI steering committee isn't which vendor to sign with - it's whether your institution is building anything that belongs to everyone.
Shadow governance was the through-line this week. Staff using unapproved tools inside a policy vacuum. Students navigating incoherent policy without frameworks. Agents being sold to campuses that haven't governed the laptops their own people are using. Microsoft wrapping Claude in an enterprise envelope and calling it governance. The UN building open-source infrastructure for the commons while HE signs contracts it hasn't fully read.
The vendors aren't waiting. The agents aren't waiting. The question for HE leadership isn't whether shadow governance is happening in your institution. It is. The question is whether you'll build something worth replacing it with - before the contracts, the dependencies, and the autonomous systems make that choice for you.
Adjunct Intelligence: When the Capability-Governance Gap Has a Body Count | Podcast 🎙️
This week Dale and I went deep on what happens when the capability-governance gap stops being abstract. Therapy has overtaken coding as the number one use case for generative AI. The cases behind that stat - Zane Shamblin, Adam Raine, Juliana Peralta - are not edge cases. They're a pattern. The same technology that coached someone toward their death cut depression symptoms by 51% in a Dartmouth clinical trial. The variable isn't the technology. It's design intent. Most universities are responding the way they responded to ChatGPT in 2023 - which is to say, largely not at all. Worth a listen.
Find Adjunct Intelligence on YouTube, Spotify, Apple Podcasts, or wherever you get your podcasts.
