In partnership with
Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
While forming the next NZ Parliament will drag on for some time yet, Finn Hogan from Stuff has interviewed each of the political parties for their take on AI (Finn, we salute you for digging in). Te Pāti Māori were approached but did not put forward a candidate. Here is where the other six landed:
National (Penny Simmonds): Keeps the "light touch" approach. No new law and no AI office, because each government department manages its own AI risk.
Labour (Reuben Davidson): Moves AI into the Prime Minister's department under a senior minister, with an independent regulator and an expert advisory panel. No AI safety institute.
Greens (Chlöe Swarbrick): An AI office that grows into a safety institute, a pause on consenting new hyperscale data centres until safeguards are in place, and a ban or moratorium on super-intelligent AI.
ACT (David Seymour): Be a "fast follower" of the rules set in the US, EU and China. No new minister: "We've got 82 ministerial portfolios in this country".
NZ First (David Wilson): No published AI policy yet. Floats a "national interest test" for AI systems, and a ban for any system that fails it.
TOP (Holly Knill): Expects job losses to come before new jobs. Points to its $19,400 citizens' income and a 25% tax credit for technology investment.
Finn also asked each of them for their P(doom), the chance of an AI catastrophe: TOP 1%, National 2%, NZ First 20% and the Greens 50/50. Labour and ACT did not give a number.
My take…
Most of the risks Finn raised came from AI labs and autonomous agents: a rogue agent at Hugging Face, the Medicare hack and a model that gave out sarin instructions. But most of the answers covered everyday workplace use: privacy law, copyright and a human checking each decision (this is completely impractical).
These are two VERY different security debates. NZ has little say over what the frontier labs build, and for almost every NZ business those risks have little to do with how they will use AI in the next few years, given how slowly it is spreading through workplaces.
The workplace risk is much lower. We know of no significant security incident from a business using a commercial AI tool with normal safeguards.
Mixing the two debates has a cost. Businesses read headlines about rogue agents and hold back on low-risk work like drafting and analysis, which is where they can get the most value from AI today.
So the rules should treat the two separately:
Chat tools: what information goes in and who checks the output.
Agents: what they can access, who monitors them and who can switch them off.
Data privacy and data rights are the critical piece, because they apply to both. That means updating the Privacy Act 2020, knowing where business data is stored, and giving businesses the right to pull their own data out of the software they pay for.
With those rules set, the Government’s job is to build the conditions for growth: skills, investment incentives and help for businesses to adopt AI, while staying out of how businesses run.
Two 1:1 coaching programmes are open: a four-week AI Coaching Programme built around existing tools and actual workflows, and an AI Executive Coaching Programme for C-suite leaders in larger companies. Reply for the details.
Happy reading ✌️
Did someone forward this? Sign up.
🇳🇿 New Zealand News
WeCare's Christchurch clinics use an AI assistant to collect patient histories and symptoms. Clinicians still approve prescriptions and care plans, while a wider team supports more than 30,000 enrolled patients.
Read
Our take: A clinic operator at last week's AI Residency told me, “We don't have a lack of GPs in NZ, we have a lack of GP capacity”. They estimated that inbox admin and triage take over a third of a GP's time. Starting with tools such as Heidi for consultation notes, AI could free more time for patients and help GPs finish on time at the end of the day, with more energy, and feeling fresher for patient consultations each day leading to better health outcomes!
Wētā FX and TAICCA signed a memorandum of understanding in Taipei on 6 November. It covers technical cooperation with Moonshine Studio, exploring investment in AI research, and international talent exchanges. No investment amount for the MOU was announced.
Read
Our take: Directors should be able to try a complex shot faster without surrendering creative control. An AI production plan also needs to specify approved models, training data and permissions for performers’ likenesses. Keeping client material in controlled systems can reduce exposure, but neither in-house development nor a vendor’s promise automatically settles the rights.
The FMA says AI can assist audits but raises confidentiality, reliance and supervision risks. Its annual review found four non-compliant files among 20 examined across eight registered firms; the AI warning is separate from those findings.
Read
Our take: As token use grows, capable open-weight models give firms an alternative to routing every task to an expensive hosted model. Local deployment can keep confidential material on infrastructure the firm controls. Test cost and capability for the job, while retaining the auditor's duty to check evidence and challenge assumptions before signing an opinion.
Newsroom reports Christopher Luxon used a non-government AI account for official work. His office says he entered only public information, but it told Newsroom that there is no central system to search his accounts for requested records.
Read
Our take: Politicians should use AI for research, summarisation and drafting. Government must still identify, preserve and retrieve official records, whichever account is used. As agents start making recommendations and coordinating work across systems, recordkeeping needs to cover those actions too. The same accountability obligations should apply to National, Labour and any other government.
DOC uses Interpine's drone images and AI to locate invasive plants at Tongariro. Contractors can visit identified locations, and another aerial survey is planned for December.
Read
Our take: By directing contractors to identified weeds, AI could let the same team cover more ground and spend more time restoring it. Connecting regular surveys, prioritisation, contractor visits and checks that weeds were removed would redesign the whole job around that capability.
Project Jonah's AI-assisted buoy detected pilot whales twice during a Golden Bay trial. One detection coincided with a stranding and the other did not; the group seeks a second buoy to help work out which direction whales travel.
Read
Our take: Earlier alerts could give rescue teams more preparation time. These two detections cannot establish how reliably an alert predicts a stranding, because one coincided with a stranding and one did not. A second buoy could help the group work out which direction the whales are travelling. That would provide more useful context for a rescue decision than a detection alone.
AI-altered property photos can mislead buyers even when a listing carries a disclaimer. The November report distinguishes existing liability from a bill proposing higher company penalties.
Our take: An agency remains responsible for what a property listing leads a buyer to believe, even if AI produced the image. It needs permissions, human approval, checks against the original and a record of changes built into publication, so accountability does not depend on someone remembering an AI policy.
Brendan Larsen's composite Mount Taranaki photograph drew accusations of AI use. He documented its creation on video.
Our take: Cheap AI-generated content makes proof of origin more valuable. A creator should be able to show who made an image, how it changed and who holds the rights. Digital signatures and provenance records could support that, with blockchain worth testing where it adds verifiable history. A signature can identify a source; it cannot by itself prove a picture is true.
Lowest error rate on real dictation. Ahead of Google, OpenAI
Speech models love clean lab audio. Your real prompts don't sound like that. So Wispr Flow tested its in-house speech model, Canto, on 10 hours of real dictations from 2,300+ speakers.
The result: the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. It also tied for the lowest error rate on LibriSpeech, a standard public benchmark.
For you, that means fewer fixes and longer, richer prompts. Speak your context, constraints and examples into ChatGPT or Claude, and Wispr Flow turns it into clean text in any app. Works on Mac, Windows, iPhone and Android.
📚️ Mike's Takes
Helping leaders and teams adapt, learn, and scale with AI.
1. AI can make work possible that never fitted the budget. Mike used GPT-6 Astra and Claude to analyse 933 customer-call transcripts, then built a process that keeps learning from new calls. The aim is to improve pitches, services and delivery, with broader action still being developed.
Read
2. Blocking Codex leaves the harder job of building useful access. Mike describes an emergency meeting over a Codex download. IT needs clear permissions and approved agents that people can use effectively. Managers must also help teams build capability and take responsibility for the work.
Read
3. AI changes a leader’s job before it reduces the workload. Mike’s “use it, check it, own it” framework shifts effort into planning and review. Leaders should test decisions, require teams to check their work and capture repeat processes. The person approving the output remains responsible.
Read
4. Larger AI bills can make sense when the work expands. Cheaper AI can make previously unaffordable work viable. Mike argues for judging the cost of completed work, including review, rather than token prices alone. Anthropic’s internal-use comparison uses retail prices and does not reveal its actual costs.
Read
5. Five Claude Skills turn company context into better leadership preparation. A CEO kept re-explaining the business to Claude. Mike describes five Skills for trusted documents, meeting preparation, board-paper reviews and reusable feedback. These give the assistant company context and record the leader’s judgement for later conversations.
Read
🌍 Tech Updates From Global
Anthropic
Claude Marketplace adds security, development, creative and productivity products, with eligible enterprise customers able to allocate part of committed Anthropic spending. (4 Nov) Read
Claude’s parent organisation console enters beta, centralising identity settings, groups and Compliance API keys across linked Enterprise and Platform organisations. (5 Nov) Read
Claude Enterprise’s security centre enters beta, comparing chat, Cowork and Claude Code settings with recommended configurations and providing downloadable review reports. (5 Nov) Read
Claude Code 2.1.319 runs MCP tasks in the background and adds a setting that denies tool calls requiring permission prompts. (6 Nov) Read
Claude Code 2.1.317 adds scheduled-task connector warnings, account-email disclosure checks and an option that blocks actions when Agent SDK hooks fail. (3 Nov) Read
Claude Code 2.1.318 lets mods choose effort for model forks and adds stored sign-in expiry information and advance warnings. (4 Nov) Read
Anthropic documents text watermarks for Mythos 5, Opus 4.7 and Sonnet 4.6, with platform coverage and sampling exceptions. (3 Nov) Read
Claude Console adds beta IP allowlisting, letting administrators restrict covered services to up to 1,000 approved addresses or network ranges. (5 Nov) Read
Eligible Claude Enterprise organisations can now tag projects with cost codes in beta and report most project spending against those codes. (6 Nov) Read
Personal Finance enters US beta for Pro, Max and Ultra subscribers, connecting accounts through Plaid for read-only balance and transaction analysis. (5 Nov) Read
Anthropic publishes evaluations of agent memory, examining searchable transcripts, saved notes, wikis and background processes that reorganise and correct stored information. (6 Nov) Read
OpenAI
Dots reach eligible Enterprise, Edu and Healthcare workspaces, excluding UAE residency, with administrator controls and a preview until 18 November. (4 Nov) Read
OpenAI releases GPT-6 Cyber through Codex and the API to selected approved Daybreak Red customers for authorised vulnerability research. (5 Nov) Read
OpenAI releases Images 2.6 globally across ChatGPT, Codex and the API, reporting 50% faster generation and improved consistency across edits. (4 Nov) Read
OpenAI launches its US Grid Protectors programme to arrange reductions in future compute campuses’ electricity demand when local grids face strain. (2 Nov) Read
OpenAI plans an internal security sprint with Cathedral researchers, who will use OpenAI models alongside company teams to identify weaknesses and test fixes. (2 Nov) Read
OpenAI announces research with Mazda to test AI-assisted aerodynamics and crash-safety design through simulations, with further progress and results due in coming months. (3 Nov) Read
ChatGPT begins rolling out GIF creation, letting users request looping animations, animate uploaded sketches, adjust playback speed and share the results. (4 Nov) Read
OpenAI publishes an Australian university collaboration overview and a commissioned survey of 1,018 students, describing existing partnerships and staged ChatGPT Edu access. (5 Nov) Read
Google Vids added Omni 1.2 Flash video generation; eligible accounts can make three 15-second videos monthly without an extra charge. (6 Nov) Read
Google Cloud made long decodes up to 1,048,576 tokens minus input tokens generally available for Gemini 4 Argon through Interactions API v1beta1. (5 Nov) Read
Google made Omni 1.2 Flash generally available with 15-second video generation, six reference images and lower stated video-token pricing. (5 Nov) Read
Google opened a client-side encryption policy beta for qualifying Workspace files; access requires Enterprise Plus with Assured Controls or Assured Controls Plus and allowlist enrolment. (5 Nov) Read
Google deprecated Gemini 2.5 Flash in AI Studio; on 1 December, API calls will route to Gemini 3.5 Flash-Lite at its rates, with breaking changes. (2 Nov) Read
Meta
Conductor confirmed its acquisition by Meta; its cloud service becomes read-only on 13 November and closes on 20 November, with local workspace transfers available. (4 Nov) Read
ABC reported that Australians can object to Meta using public Facebook and Instagram data for future AI training, through a form available since July 2025. (5 Nov) Read
Microsoft
Microsoft announced Business Central and Copilot Business CSP bundles for up to 300 seats from 1 January 2027, at stated US$95 or US$125 per user monthly list prices. (5 Nov) Read
Microsoft postponed default usage-based billing for new CSP Copilot Business subscriptions; rollout is now planned during January to March 2027. (3 Nov) Read
Microsoft added Copilot Chat fair-use limits on 2 November for users without a Copilot licence in tenants up to 2,000 seats; larger tenants start 9 November. (4 Nov) Read
Baidu
Baidu released ERNIE 5.5 Flash Base weights, describing a 280B-total, 14B-active model trained on 33.5 trillion tokens. (5 Nov) Read
Mistral
Mistral expanded Studio with managed enterprise agents and applications, workflows, observability and evaluations, supporting cloud, private-cloud or on-premises deployment. (6 Nov) Read
Amazon
AWS WorkSpaces for AI agents now supports Linux and Windows 11 cloud desktops. (3 Nov) Read
AWS described Amazon Quick's Smart Approvals, which can approve eligible agent actions automatically while retaining policy denials and required human review. (5 Nov) Read
AWS added Kimi K3.1 Pro Highspeed to Bedrock through cross-Region inference, with explicit prompt caching support. (7 Nov) Read
xAI
xAI now routes grok-imagine-image-quality and grok-imagine-image-pro requests to grok-imagine-image-2.0-low, preserving request formats while changing pricing and the returned model name. (2 Nov) Read
Reports & Data This Week
The OECD's Digital Economy and Society Outlook 2026 reported higher AI uptake among large firms in its cross-country statistics. (5 Nov) Read
Anthropic analysed 207,000 Claude and Claude Code research sessions, finding relatively high formal-science use and frequent data analysis. (5 Nov) Read
Menlo Ventures estimated enterprise AI spending at US$176.8 billion by September 2026, 5.4 times the previous year's level. (4 Nov) Read
METR’s review of Anthropic’s August risk report supports its low catastrophic-risk assessment but questions whether an early warning threshold was crossed. (5 Nov) Read
Visa found 42.8% of agentic shopping benchmark runs failed across 22 models and 15 buying tasks. (6 Nov) Read
👋 Mike