Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Easily the funniest AI moment last week happened at India’s AI Impact Summit in New Delhi, where PM Narendra Modi lined up a bunch of tech leaders for a staged “unity” photo and got everyone holding hands and raising them.
Sam Altman (OpenAI) and Dario Amodei (Anthropic) were standing next to each other, refused to clasp hands, and instead popped separate little fists while Modi lifted Altman’s hand.
Happy reading ✌️
Did someone forward you this? Sign up!
academyEX is the EXceptionally proud sponsor of this week's newsletter.
🇳🇿 New Zealand News
NZ's Supreme Court will now jail people for citing fake AI cases. A self-represented litigant filed AI-hallucinated case law in a Family Court appeal. The Supreme Court responded with a first: unverified AI citations can trigger Obstruction of Justice (seven years prison) or Contempt of Court charges. Legal professionals are pushing for mandatory AI disclosure in all filings.
2 min read.
Our take: Seven years for trusting ChatGPT. NZ just became the first country where an AI hallucination carries a prison sentence longer than the underlying family dispute.
A Christchurch judge ran a convict's apology through ChatGPT. Judge Tom Gilbert tested the authenticity of remorse letters submitted by a convicted arsonist by running them through ChatGPT. The AI confirmed what the judge suspected: the letters were machine-generated. The court ruled algorithmic remorse cannot substitute for genuine accountability in sentencing.
2 min read.
Our take: A judge used AI to catch AI. That's the new arms race in every regulated industry. If a district court judge in Christchurch is running detection checks, assume every compliance officer, auditor, and regulator will follow.
An AI just completed the first agentic commerce transaction in NZ to buy movie tickets and a hotel with its own payment credentials. Mastercard and Westpac completed NZ's first fully authenticated agentic transactions. Using the "Agent Pay" framework and Maincode's Matilda LLM, an AI autonomously booked Event Cinema tickets and QT Hotel rooms without human intervention at checkout.
3 min read.
Our take: The first autonomous AI purchase in NZ was movie tickets. The second will be enterprise procurement. Every CFO should be asking: what happens when AI agents start negotiating supplier contracts without human approval? Agentic Commerce is a hot topic that we advise businesses on. Get in touch to learn more.
Waikato University trained AI to speak te reo Maori, and the iwi owns every byte. A hyper-realistic AI voice model now speaks the Waikato-Maniapoto dialect of te reo Maori. Built by Dr Te Taka Keegan and Kingsley Eng with Google funding, the iwi retains absolute copyright. The project is designed to prevent digital colonialism of indigenous language data.
3 min read.
Our take: Every endangered language on earth faces the same choice: let tech companies train on it without permission, or build the model themselves and own the result. Waikato just demonstrated option two works.
Kiwi startup Hectre now grades 54 billion pieces of fruit a year with AI. Hectre closed an oversubscribed NZ$12M Series A led by Punakaiki Fund. Its computer vision system, Spectre, uses AI to grade fruit quality at industrial scale. The company processes 54 billion individual fruit assessments annually.
2 min read.
Our take: 54 billion assessments. Not conversations, not tokens, not API calls. Physical objects, graded by AI, at a scale no human workforce could replicate. That's the kind of moat ChatGPT can't erode with an update.
30% of NZ Corrections staff were using AI to write supervision orders. The Department of Corrections discovered staff had been using Microsoft Copilot to draft formal reports, including Extended Supervision Orders. About 30% of staff accessed Copilot since November 2025, breaching policies on entering personal information into AI tools.
2 min read.
Our take: 30% adoption with zero training, zero budget, and zero permission. If a government department with strict data policies can't stop shadow AI, no NZ business can. The question isn't whether staff are using AI. It's what data they're feeding it.
NZ's Solicitor-General warns deepfakes are disrupting trials. Crown Law told Parliament that deepfakes are undermining court evidence. Trials have been disrupted by defendants claiming CCTV footage was AI-fabricated. Lawmakers are considering heightened admissibility thresholds and mandatory digital reliability checks.
2 min read.
Our take: The damage is already done before any deepfake is proven. A single claim of "that footage is AI-generated" forces the prosecution to prove a negative. Guilty defendants just got a new playbook, and it costs nothing to run.
Spark's profit jumped 82.9% and AI gets the credit. Spark NZ reported H1 NPAT up 82.9% to $64M. CEO Jolie Hodson directly credited AI for reducing business product setup time by 60%. Spark also partnered with The Icehouse to run free AI productivity workshops for SMEs.
2 min read.
Our take: 60% faster setup on business products. That's not an innovation metric, it's an operational one. Spark isn't experimenting with AI anymore. The board is booking the savings. Every NZ CEO who put AI in the "innovation" budget should move it to "operations".
✨A few people have been asking…
It’s Mike here, I run The AI Corner.
I’m not just into writing about AI. I help businesses grow without adding headcount by shipping AI systems that cut costs and increase throughput.
🎙️ The AI Corner Podcast
Join Mike and Habib Baluwala, AI Transformation Lead at Fonterra and one of New Zealand's most experienced enterprise AI executives, as we discuss:
Why most NZ businesses stall on AI transformation, and the three readiness factors across data, people, and process that actually determine your pace.
How to design governance that feels like a service, why the process layer is where most organisations fail, and when to rebuild a workflow from scratch versus retrofitting AI into an existing one.
The Human Delta framework, a new way to measure the direction, judgement, and relational trust that humans still bring to AI-assisted work, and why that's the next frontier for how we value people at work.
📚️ AI for Business
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ Four conversations, one pattern: enterprise is getting in its own way on AI: One company spun up a frontier team on Claude Code and Cursor, actively hunting edge. The other three? Copilot rollouts branded as innovation, AI licences deemed too pricey, or locked to a homegrown GPT.
2-min read.
2️⃣ Copilot keeps the team quiet. Frontier Teams keep the business competitive: OpenAI calls it the "capability overhang", the gap between what AI tools can do and how people actually use them. Businesses need Frontier Teams that experiment at the edge, not a Copilot rollout and a tick in the box.
4-min read.
3️⃣ AI distribution just got its first proof point, and it changes everything about visibility: Canva traffic from LLMs now accounts for double-digit percentages of 265 million users, with 26 million conversations through the Canva ChatGPT app alone. AI agents compare options through data and logic, not persuasion, so showing up means stock levels and turnaround times, not lifestyle photos.
4-min read.
4️⃣ The AI Vampire: 10x productivity means nothing if the team is fried by Q3: Steve Yegge named the doom loop everyone in AI is living but nobody had words for. Something finishes in 10 minutes that used to take half a day, and the instinct is never to rest, it is to fill the gap with more.
3-min read.
5️⃣ The unintended consequences of "Something Big Is Happening": Matt Shumer's essay hit 80 million views, five people sent it through unprompted, three who had never raised AI before. The core message is right, but for many readers, the takeaway was fear, not motivation.
5-min read.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
Launched "OpenAI for India" initiative and signed Tata Consultancy Services as first data centre customer, starting at 100MW with a pathway to 1GW.
Approaching a $126 billion mega-funding round at an $850 billion valuation, with Amazon, SoftBank, Nvidia and Microsoft as strategic investors.
Confirmed its safety systems flagged the Canadian mass shooting suspect months before the attack but did not refer the case to police, igniting global debate over AI "duty to warn" responsibilities.
Retired GPT-4o and four other legacy models from ChatGPT, making GPT-5.2 the default for all users and triggering backlash from people who had formed AI companion relationships with older models.
Upgraded ChatGPT Deep Research to GPT-5.2 with a new fullscreen document viewer, real-time progress tracking and the ability to limit searches to specific websites or time ranges.
Nvidia reportedly nearing a US$30 billion equity investment in OpenAI at a US$730 billion pre-money valuation, replacing the stalled US$100 billion infrastructure partnership.
Committed $7.5 million to independent alignment research teams via the UK AI Security Institute's Alignment Project.
Hired OpenClaw creator Peter Steinberger to build "next generation personal agents," with OpenClaw transitioning to an OpenAI-supported foundation.
A US court blocked OpenAI from using the "Cameo" name for a Sora feature in a trademark dispute with the company Cameo.
OpenAI is reportedly developing physical AI devices, including a smart speaker concept, extending beyond pure software distribution.
Released Gemini 3.1 Pro Preview scoring 77.1% on ARC-AGI-2, more than doubling the prior Gemini 3 Pro score and surpassing its own Deep Think model.
Announced "America-India Connect" backed by US$15 billion, including three new subsea cable routes linking India to Singapore, South Africa and Australia.
CEO Sundar Pichai committed to a gigawatt-scale "full-stack AI hub" in Visakhapatnam as the centrepiece of Google's India infrastructure investment.
DeepMind partnered with India's government to deploy AI tools and hardware across 10,000 schools reaching an estimated 11 million students.
Launched Lyria 3 generative music model inside the Gemini app, producing 30-second tracks with lyrics and vocals from text prompts, watermarked with SynthID.
Launched Perch 2.0, an AI system for classifying whale vocalisations to support ecological research.
Anthropic
Released Claude Sonnet 4.6 with near-Opus-level intelligence at one-fifth the cost and a 72.5% score on computer use benchmarks, making it the new default model for free and paid users.
Launched Claude Code Security in research preview, an autonomous vulnerability scanner that fixed over 500 zero-day flaws in open-source codebases, triggering a cybersecurity stock sell-off.
Partnered with Infosys to build enterprise-grade AI agents powered by Claude 4.6 for highly regulated industries, starting with telecommunications.
Signed an MOU with the Government of Rwanda to integrate frontier AI into public health and education systems.
Opened a corporate office in Bengaluru as the operational hub for South Asian partnerships and localised model deployment.
Microsoft
Committed US$50 billion by 2030 to subsidise foundational AI access in the Global South across infrastructure, skills training and multilingual models.
Launched "Elevate for Educators" in India targeting 2 million teachers and 200,000 schools by 2030 in partnership with CBSE, NCERT and AICTE.
Integrated xAI's Grok 4.1 Fast model into Copilot Studio, reinforcing its model-agnostic platform strategy on Azure.
Added Claude Sonnet 4.6 and Gemini 3.1 Pro to GitHub Copilot, intensifying multi-model competition inside the developer workflow.
Launched Copilot Coding Agent from Visual Studio and introduced custom agents defined as repository files with MCP integrations.
Amazon
Projected to hit US$200 billion in 2026 AI capital expenditure, a 56% increase over 2025 and the largest single-company infrastructure commitment.
Added DeepSeek V3 open weights models to Amazon Bedrock, offering enterprise clients economical inference alternatives.
Launched EC2 Hpc8a instances with 5th-gen AMD EPYC processors delivering a 40% performance boost for compute-intensive AI workloads.
xAI
Launched Grok 4.20 beta featuring a "4 Agents" multi-agent architecture where four independent agents analyse queries from adversarial angles before an orchestrator synthesises the response.
Secured US$3 billion from Saudi Arabia's Humain as part of a US$20 billion Series E, with equity converting to SpaceX shares post-merger and a 500MW data centre commitment.
Faces a Clean Air Act lawsuit over 27 unpermitted gas turbines in Mississippi emitting pollutants into residential communities near its Colossus 2 data centre.
EU and Irish regulators launched a GDPR and Digital Services Act investigation into Grok over non-consensual deepfake generation.
Meta
Formalised a multiyear infrastructure partnership with Nvidia, transitioning data centres to the Vera Rubin GPU platform with Spectrum-X networking for Llama training.
Declared Threads fully "ad-ready" with native advertising placements and launched an AI-driven Creator Marketplace connecting brands with creators.
Reported AI recommendation engines drove a 7% lift in Facebook feed views, 25% more same-day Reels and a 20% increase in Threads user time.
Expanded automated video dubbing on Instagram Reels to nine additional languages including Bengali, Tamil, Kannada, Marathi and Telugu.
Apple
Released iOS 26.4 beta with "Playlist Playground," a generative AI tool in Apple Music creating personalised playlists from natural language prompts.
Embedded "Agentic Coding" natively into Xcode 26.3, enabling semi-autonomous multi-file bug detection and UI generation for developers.
Machine learning researchers published a paper proving frontier LLMs suffer catastrophic logical collapse when standard reasoning puzzles are slightly altered.
Reportedly accelerating development of AI-powered smart glasses, an AI pendant and camera-equipped AirPods as always-on Apple Intelligence interfaces.
Nvidia
Cemented a multiyear partnership with Meta to deploy the Vera Rubin GPU platform and Arm-based Grace CPUs across Meta's data centres.
Anchoring India's US$1 billion national AI Mission, partnering with Yotta to build the "Shakti Cloud" with over 20,000 Blackwell Ultra GPUs.
Published Blackwell Ultra benchmarks showing 50x higher throughput per megawatt versus Hopper and up to 35x lower costs for agentic AI inference.
Indian IT giants Wipro, Infosys, Tech Mahindra and Persistent Systems launched autonomous enterprise agents built on the Nvidia AI Enterprise stack.
👋 Mike & Erin