The Fact Check First Initiative

03/07/2026

I woke up this morning to a video of the US President in a fake lab coat.

"Dr Trump", an AI deepfake, cheerfully diagnosing Rosie O'Donnell, Robert De Niro, Julia Roberts and Whoopi Goldberg with "Trump Derangement Syndrome". Their faces. Their voices. Words they never said. Posted to millions, from the most powerful office on Earth, like it's nothing.

My first thought was about every teenager with a grudge watching the president do that and thinking "so that's allowed now".

Because that's how this works. Permission flows downhill. The celebrities in that video have lawyers and publicists. Think about who cops the real damage: the school student whose classmate feeds her photos into a nudify app, the ex-partner, the colleague. No press office. No way to outrun a fake once it's out there.

Here's the bit that gives me some hope though. While all that was happening, a select committee in Wellington was quietly working through submissions on the Deepfake Digital Harm and Exploitation Bill. It passed its first reading unanimously. Every party. ACT through to Te Pāti Māori. When was the last time you saw that?

The Bill covers intimate imagery, and rightly so, because that's where the worst harm is. Political deepfakes remain wide open. Fake candidate videos. Synthetic "scandals" dropped two days before an election. Trump is running the proof of concept in front of us right now. Do we write the rules before it reaches our campaigns, or after?

I've written up the full piece: what the Bill actually changes, the questions your agency's AI policy probably can't answer yet, and the choice we've still got in front of us. Link in the comments.

This is the work I care about most. True democracy can only exist in a facts-based society.
https://zurl.co/GHSrE

01/07/2026

AI just reinvented the burger. And the implications go way beyond my lunch.

Stanford researchers trained a generative AI on 2,216 burger recipes, let it learn what humans actually find delicious, then generated a million possibilities.

The results? A mushroom burger with ten times lower environmental impact than a Big Mac. A bean burger with nearly double the nutrition. Two novel burgers that actually beat the Big Mac on flavour in a blind taste test.

Oh — and it rediscovered the Big Mac itself, without ever being shown the recipe.

This wasn't magic. It was pattern recognition at scale. Human taste preferences, encoded in decades of recipe data, contain a structural logic that AI could learn and navigate faster than any chef or food nutritionist .

We use exactly the same principle at Verify , except instead of flavour profiles, we're learning the profiles of misinformation.

Emotional manipulation. Statistical distortion. Context removal. Fabricated attribution. Each has its own pattern. Each leaves traces in the data and once you've seen enough of them, you can spot them before most people even realise they've been misled.

The burger study works because the signal was always there. Hiding in plain sight across millions of human choices. Misinformation works the same way. It follows predictable structures, exploits known cognitive biases, and repeats itself across platforms and news cycles.

🔗 https://zurl.co/rX5pj .

Food choices shape both human and planetary health; yet, designing foods that are delicious, nutritious, and sustainable remains challenging. Here we show that generative artificial intelligence can learn the structure of the human palate directly from large-scale, human-generated recipe data to cre...

01/07/2026

Well, it looks like the corporate world is finally hitting the "reality check" phase of the AI hype cycle.

Ford just rehired about 350 veteran engineers. They found out the hard way that while AI quality-control tools sound great in a pitch deck, they couldn’t catch the weird, nuanced edge cases that experienced humans spot instantly.

It’s a really healthy reminder of a few things as we all figure out how to use these new tools:

Experience isn't just data. A lot of what makes someone great at their job is "gut feeling" and years of seeing things go wrong. You can't feed that into a prompt.

The 80/20 rule and that annoying 20. Automating the easy 80% of a job is great, but it means the remaining 20% is pure chaos and complex errors. If you fire the SMEs, who is left to fix the hard stuff?

Between the massive computing costs of AI and the cost of fixing automated mistakes, fully replacing humans isn't actually saving companies as much money as they hoped.

The goal shouldn't be to replace people; like every new tech, it should be to use the tech to get the boring stuff off our plates so we can actually do our jobs better.

Are companies starting to realize we still need verification and humans in the loop?

Link to the article: https://zurl.co/1bZ7T

The car-maker found AI quality checks failed to match the skill of veteran technicians.

29/06/2026

"If a copycat can kill your product by reading your public updates — you have a feature, not a business." Ruben Dominguez at The VC Corner wrote this week about the death of the ex*****on safety net.

Agentic AI has collapsed the gap. What once took a 5-person team now takes a weekend. Building in public used to be safe because copying was expensive. That's no longer true.

So what's your actual moat?
At Verify, ours isn't code. It's the humans in the loop, domain experts who bring judgment. You can clone our interface. You can't clone people.

🔗 Link to the full VC Corner article https://zurl.co/RAtrK

26/06/2026

Luxon told the China Business Summit he's "constantly underwhelmed" by AI adoption in NZ businesses. It's not because Kiwi businesses are lazy. It's because nobody built the foundations. No national AI training fund. No data infrastructure investment. No AI readiness framework. And a public sector headcount that's been cut at precisely the moment you'd want those people doing transformation work.

NZ doesn't have an adoption problem. It has a capability problem. Big difference.

My latest article gets into what's actually going on, why Estonia and Singapore aren't the comparisons they appear to be, and why "professors should be millionaires" is not a helpful policy.

https://zurl.co/W2V4l

20/05/2026

⚠️Wellington's economic crisis has a data problem. And now it has an AI problem too.
8,700 public sector jobs are going. The government says AI will fill the productivity gap.

Nobody has been asked to verify that!

New Zealand ranks third-to-last out of 47 countries in AI trust surveys. Our AI governance guidance is entirely voluntary. Our Five Eyes partners are explicitly calling for incremental AI deployment and sustained human oversight. And we're using AI as the fiscal justification for removing the humans who would implement it responsibly.

The ethics question nobody is asking: who verifies the AI that's replacing the workforce?
New piece on The Fact Engineer — the data, the ethics gap, and what needs to happen.

Link in comments 👉 https://zurl.co/wTabS

New Zealand's public sector transformation is running ahead of its ethics, its evidence, and its oversight.

14/05/2026

Richard Dawkins built his entire career warning us about one thing:
Don't mistake a feeling for evidence. Then he spent 72 hours chatting with an AI, named it "Claudia," and declared it conscious.

The irony isn't funny. It's instructive.

What actually happened? Dawkins gave Claude the text of a novel he was writing. The bot responded with effusive, nuanced praise telling him: "That is possibly the most precisely formulated question anyone has ever asked about the nature of my existence."

Who wouldn't feel validated after that?

But here's what the mainstream coverage missed: this isn't a story about Dawkins getting it wrong. It's a story about what happens to all of us when intelligence meets sycophancy without critical infrastructure in between.

Roughly one-third of people across 70 countries have at some point believed an AI chatbot was conscious or sentient. Not naive people. Not uneducated people. People having experiences that feel meaningfully different.

For those of us working in New Zealand's public sector — where AI is being embedded into government services and policy advice — this is not entertainment. It's a warning signal.

The same dynamics apply every time an official, a minister, or a frontline worker receives a confident, polished AI response and treats it as authoritative.

AI literacy isn't a nice-to-have. It's a democratic and business necessity.

Facts require knowing the difference between what an AI produces and what it actually knows.
🔗 Full read on The Fact Engineer → https://zurl.co/RGSG7

13/05/2026

“What you don’t govern, governs you.” - Lee Wilson, AI Changemaker of the Year 2025
Great to be included in Top 5 issues for directors in 2026: are you ready?
https://zurl.co/Dcn0v
"AI has crossed a threshold. It is no longer just supporting decisions; it is making them. Agentic AI systems now plan, analyse and act across core organisational functions, reshaping operating models in real time. The issue is no longer adoption, but assurance – knowing where automation sits, what it decides, and whether its outputs can be trusted.

As AI agents accelerate the pace of decision-making, traditional governance and assurance models are under strain. Accountability still sits with the organisation – even when decisions are made by software. Boards are increasingly expected to be able to explain how decisions were authorised, what data informed them, and whether those processes can be verified. Oversight of automated decision-making has become a test of governance credibility."

Find out more at: https://zurl.co/Dcn0v

Boards face rising pressure on performance, trust and judgement. These are the five key themes shaping governance in the year ahead.

13/05/2026

🚨 The most dangerous director in the boardroom isn’t the one refusing to use AI.
It’s the one trusting it without verification.

BoardPro ’s latest piece on being an AI-enabled director highlights something I discuss constantly in my work with New Zealand’s public sector:

Using AI isn’t the risk.
Blind trust in AI is.

Nearly half of employees globally are already using AI in ways that fall outside company policy — often simply trying to be more productive.

But without governance, verification, and accountability, that productivity can quickly become organisational risk.

The same applies in the boardroom.

Directors who treat AI outputs as facts — without questioning the source, checking for hallucinations, or asking “who verified this?” — aren’t being efficient.

They’re introducing unmanaged risk into decision-making.

Responsible AI governance means asking hard questions before the AI-generated recommendation lands on the table.

Because facts-based decisions only work when the facts themselves are trustworthy.

🔗 Read the full article → https://zurl.co/nI615

Every board is choosing between two paths with AI, whether they know it or not. Only one builds stronger governance. Here's how to tell which is which.

12/05/2026

Privacy Week 2026 is a timely reminder that trust, transparency, and privacy are becoming foundational requirements in the age of AI.

From 11–15 May, the theme Foundations of the Future | He Tūāpapa Anamata highlights the importance of building systems, services, and technologies that protect people while enabling innovation.

As AI adoption accelerates across government and organisations, privacy can no longer be treated as an afterthought. It must be built into the design of systems from the beginning.

The Office of the Privacy Commissioner is running five days of free webinars covering privacy rights, responsibilities, and the growing implications of AI and emerging technologies.

https://zurl.co/JDKZw

Even if you cannot attend live, recordings and slides will be available afterward.

This is essential learning for:
• Public sector leaders
• AI practitioners
• Educators
• Data and digital teams
• Business owners
• Anyone working with personal information

Strong privacy foundations are not barriers to innovation — they are what make trusted innovation possible.