Daily Briefing

Leon Furze's new post applies Narayanan and Kapoor's "AI as normal technology" framework — from their book AI Snake Oil — to GenAI in education, and the argument is both reassuring and clarifying. The core claim: GenAI is following the same three-speed adoption cycle seen in every previous technology wave. Invention (the R&D) is moving fast, driven by investment and hype. Innovation (building actual products) is moving at market speed. But adoption — the social, organisational, and pedagogical integration that makes a technology genuinely useful — moves on a timeline of decades, not years.

The critical insight is Furze's application to schools: "the technology might be ready in a year or two, but the organisation, the school system, the hospital, might need 10-20 years to really figure out how to use it effectively." This explains why the internet, despite being commercially available in the early 1990s, wasn't meaningfully embedded in most classrooms until the 2010s — a roughly 20-year lag. MOOCs, digital textbooks, and every edtech wave before GenAI followed the same hype-then-slow-diffusion pattern. The exhaustion educators feel right now is therefore not a failure of imagination or energy — it is the rational, predictable response to hype operating at invention-speed while their institutions move at adoption-speed.

Furze argues this framework offers both relief and caution. Relief: there is no need to panic that your school hasn't fully restructured around AI in 18 months. Caution: the genuine risks (privacy, equity, cognitive effects on students) deserve slow, evidence-based engagement rather than either wholesale adoption or reflexive bans.

Princeton's faculty voted on 11 May — with only one dissenting vote — to reinstate proctoring for all in-person examinations from July 1, ending an honour system that has operated without supervisors since 1893. The 133-year tradition of self-policed, unproctored exams, which began when students themselves petitioned to remove proctors, has been broken by generative AI.

The data driving the decision: the 2025 Senior Survey found 29.9% of students admitted to cheating on at least one assignment or exam during their time at Princeton — rising to 40.8% among engineering students. More damning: 44.6% of seniors knew of honour code violations but chose not to report them, while only 0.4% actually reported a peer. AI tools on small personal devices have made violations nearly invisible to the bystander-report system the honour code depended on. Fear of online shaming has further suppressed reporting. Former Dean of the College Jill Dolan: "I think it's a shame, but it's necessary. We need some different practices in this day and age."

Instructors will serve as witnesses (not enforcers) during exams, documenting suspected violations and submitting reports to the student-run Honour Committee. A majority of students surveyed supported the change or were indifferent — but a "sizeable minority" opposes it on principle. This is the most significant change to Princeton's academic integrity system in its history, and it comes specifically because AI made the existing self-governance model untenable.

EdWeek has launched an opinion series in which educators share firsthand experience of AI in English-learner classrooms — and the opening post from contributor Larry Ferlazzo with EL educators draws a sharp distinction between AI's utility for mainstream students versus English learners. The headline finding: for EL and English Language Learner (ELL) students, AI tools are genuinely transformative in specific, bounded ways; for proficiency-level mainstream students, the evidence of real learning benefit is thin.

Where AI genuinely helps EL students: AI speaking tools for home practice (where there are no English-proficient household members), simultaneous AI translation for newcomers dropped into English-medium classes, adaptive vocabulary apps like Quill, and Gemini Storybook for generating high-interest levelled reading material. Several educators in the series describe AI as a practical equaliser for EL students who lack access to peer tutors or language-rich environments outside school.

Where it doesn't: the article is sceptical of AI use for general EL writing instruction where process matters, and cautions against AI translation becoming a crutch that bypasses the productive struggle necessary for acquisition. The underlying research (Cambridge Core, Springer, Frontiers in Psychology meta-analyses) converges on moderate-to-large gains in writing accuracy, pronunciation intelligibility, and motivation when AI tools are well-targeted — but the benefits depend heavily on digital access, teacher training, and learner engagement.

Trump landed in Beijing on Wednesday evening — welcomed with full red-carpet ceremony by Chinese Vice-President Han Zheng — and formal bilateral talks with President Xi Jinping begin Thursday. The delegation is historically large: Elon Musk (Tesla), Jensen Huang (Nvidia), Tim Cook (Apple), Larry Fink (BlackRock), and Kelly Ortberg (Boeing) are among 12+ CEOs accompanying Trump. Trump's stated opening demand: Xi should "open up" China to American companies.

The Iran war dominates the agenda in ways that were not true during planning. The Strait of Hormuz closure has directly damaged Chinese energy supply (China is the world's largest oil importer), giving Beijing both leverage and reason to help broker a resolution. China's help in reopening the strait is widely expected to come with demands — most likely concessions on Taiwan arms sales or technology export controls. The NYT reports Xi sees this summit as an opportunity to press Trump on US arms sales to Taiwan amid rising PLA activity in the strait.

Trade and technology are the other pillars. China has threatened rare earth export restrictions (critical for US defence and semiconductor manufacturing), and Nvidia's Jensen Huang's last-minute inclusion in the delegation has been read as a bargaining chip — his presence signals possible flexibility on AI chip export rules. The White House left the door open for a deal on Chinese direct investment in the US, though former Trump NSC officials warn this would be a "Trojan horse."

Real-time AI translation — embedded in video calls, TikTok dubbing, smartphone keyboards — has dissolved the immediate practical barrier to cross-language communication. The question The Conversation poses is genuine: if the tool works, is the effort of language learning still worth it? The article's answer is a carefully argued yes, but the argument is cognitive rather than cultural.

The key distinction: using a tool to extend your capabilities versus using it to avoid doing something altogether. Psychologists call this "desirable difficulties" — the struggle of producing language across two systems engages brain networks for memory, attention, and cognitive flexibility in ways that passive tool use does not. Sustained multilingualism is linked to what researchers call "cognitive resilience" — the brain's capacity to maintain function as it ages. The argument is not that AI translation is bad, but that outsourcing the process entirely means missing the cognitive workout, not just the cultural fluency.

The parallel to AI writing assistance in schools is exact: the question is not whether students should use AI translation, but whether using it replaces or supplements the productive struggle that builds durable capability. This maps directly onto the ongoing debate about AI in EL classrooms above.

Wednesday's War Powers Resolution vote in the US Senate failed 50-49 — the seventh such attempt since the Iran war began in late February, and the closest margin yet. Three Republicans voted with Democrats: Senator Lisa Murkowski (Alaska) for the first time, Senator Susan Collins (Maine) for the second time, and Senator Rand Paul (Kentucky, who has voted yes all seven times). The decisive opposing vote came from Democratic Senator John Fetterman — a pro-Israel hawk who has crossed party lines on each attempt.

The significance of Murkowski's defection is widely noted. Her reasoning: "You've got a timeline that has taken us beyond the 60 days. I thought that perhaps we would get more clarity from the administration in terms of where we are, and I haven't received it." Under the 1973 War Powers Resolution, presidents may conduct hostilities for 60 days without congressional authorisation — a clock that began running on 2 March. The Trump administration argues the clock stopped when a ceasefire was declared, a legal position experts dispute given Iran has since rebuilt 30/33 Hormuz missile sites.

The Republican coalition in support of the Iran war is visibly cracking. Collins faces a tough re-election in Maine; Murkowski's shift is harder to explain by electoral arithmetic. This is the first sign of genuine ideological fracture — rather than just political pragmatism — inside the GOP caucus on the Iran war.

The House Homeland Security Committee has formally demanded that Instructure CEO Steve Daly testify about the company's response to repeated cyberattacks that allowed the ShinyHunters hacking group to steal personal data from an estimated 275 million student records across 9,000+ schools globally. CISA has been called in to assist with the incident. Representative Andrew Garbarino's letter cites TechCrunch reporting and asks lawmakers to examine the adequacy of Instructure's breach notifications to affected schools.

The critical new detail: Instructure "reached an agreement" with the ShinyHunters hackers this week — a ransom payment. The company claimed the hackers provided evidence of data deletion; the hackers told TechCrunch they would not further extort the company, but declined to state the ransom amount. Security experts uniformly warn that paying ransoms incentivises future attacks and that hackers routinely retain data despite claiming to delete it. The same vulnerability was exploited twice — the second breach while the first was still being remediated.

The congressional investigation escalates a story that has been developing for weeks: Canvas LMS is used by schools and universities worldwide, and the stolen data includes student names, email addresses, and course communications. Schools affected include institutions across the US, UK, Australia, and MENA.

Troy City Schools in Ohio — a district of 4,000 students where 3% are English learners — decided post-pandemic to make a structural intervention rather than incremental adjustments. Facing widening literacy gaps that hit EL students particularly hard, the district trained all 116 elementary school staff (every classroom teacher, intervention specialist, paraprofessional, and principal) in the Orton-Gillingham structured literacy approach — a phonics-based, multi-sensory method with strong research backing for learners with dyslexia and language acquisition challenges alike.

The results, as reported by literacy instructional support specialist Sarah Walters, are meaningful: EL students are making gains and the district is seeing the kind of cross-curricular payoff that early structured literacy training produces. The student population includes families from an automotive manufacturer that brings Japanese employees to the area — so the district serves a mix of Spanish, Ukrainian, and Japanese home-language students alongside domestic ELs.

The broader context matters: federal data shows EL achievement scores have made little improvement over two decades. The Troy model — whole-staff training in evidence-based methods rather than just EL specialist training — is an example of systemic rather than siloed intervention.

UNIFIL's latest incident reports from southern Lebanon confirm a pattern: Hezbollah is deploying cheap, expendable drones — including Iranian-made fixed-wing models and suspected 3D-printed fiber-optic guided variants — in coordinated swarms against Israeli forces and UN positions. UNIFIL headquarters in Naqoura has experienced multiple drone detonations inside its perimeter this week, with no casualties but building damage. Israeli military analysts note that Hezbollah has adapted kamikaze drone tactics refined in Ukraine, including fiber-optic guidance that is resistant to electronic countermeasures.

The low cost ($300-$500 per unit for commercial-grade variants) and high saturation potential of these drones is straining Israeli air defence systems designed for rockets and precision missiles. Iron Dome is effective against rockets; cheap swarming drones represent a different threat category. Several UNIFIL sources have confirmed crashes of armed, fiber-optic-guided drones at multiple UN positions — the guidance wires are a signature of Ukrainian-conflict-derived tactics now being applied in Lebanon.

MIT Technology Review reports a growing pattern: AI chatbots — particularly Google's AI overview feature in search results — are surfacing real people's personal contact information (phone numbers, email addresses) when asked about individuals. Users report that their personal contact details are appearing in AI-generated summaries without their knowledge or consent, drawn from data sources the AI was trained on or can retrieve.

The problem is structurally different from a data breach: no single repository was hacked. Instead, AI systems are aggregating publicly accessible but contextually inappropriate data — the kind of information that exists in scattered places across the web and was practically obscure before AI made it trivially retrievable. This is a privacy harm enabled by capability rather than a security failure.

The incident reinforces the distinction between data being "public" and data being "appropriately accessible." Schools using AI tools that can search and synthesise information about students or staff face a version of this risk if their data policies don't account for AI's aggregation capability.