22 May 2026

Welcome

After spending the last week coughing and spluttering with respiratory syncytial virus (RSV), I’ve come back to my desk to find a vast pile of stories about AI — and that’s bound to continue.

The AI companies, and tech industry players in general, all seem to think that AI is the future. With billions and billions of dollars involved, investors hope they’re right. So too do the governments and businesses who are integrating AI into their operations.

With so much at stake, for so many stakeholders, the pressure to downplay the risks and failures will be enormous. They’re mere glitches as a revolutionary new technology settles in, right? There’s no systemic problem here!

Sure.

Another possibility is that the technology is fundamentally flawed and always will be, like pneumatic railways.

Or, more likely in my view, is that we’ll remember that AI is more than just the current fascination with chatbots, and we’ll focus once more on the AI that works — from drug design to transit system planning, from detecting financial fraud to spam filtering.

But here we are.

Oh, and not everything this week is about AI. It just feels that way.

Doctors are dumping patients who don’t allow AI

A Melbourne psychiatrist is refusing new patients who don’t consent to them using an AI scribe to take notes. And they’re far from alone.

The patient registration form seen by the Guardian reads: “I consent for use of AI transcription (such as Heidi health AI/ Microsoft) software to assist with notes taking during the appointments, for preparation of clinical letters, reports, and other clinical documents to assist in my clinical care.”

Don’t like it? Go elsewhere.

Taking notes of a psychiatry session is a core part of the job, but it’s also time-consuming, so automating it must be tempting.

The main problem, though, is that these AI scribes keep getting things wrong.

A report from the Office of the Auditor General of Ontario, Canada, showed that 12 of the 20 systems they tested mixed up prescribed drugs, and 17 of them “missed key details about the patients’ mental health issues”.

As The Register reported, “nine out of 20 AI systems reportedly ‘fabricated information and made suggestions to patients’ treatment plans’ that weren’t discussed in the recordings.”

Forcing patients to accept the AI scribes or leave is “shocking but not unsurprising”, according to lobby group Digital Rights Watch, citing risks to patient privacy, quality of healthcare, and doctors’ liability.

“It is therefore completely unacceptable for clinics to refuse treatment to patients who do not want an AI scribe listening-into their private conversations with their doctor,” said Tom Sulston, DRW’s head of policy.

“Corporations building AI scribes for healthcare are operating in an under-regulated no-man’s land where they provide interpretation of medical information without needing to be regulated as providers or medical devices.”

Sam Altman's brainwave on micropayments

In a recent interview OpenAI boss Sam Altman was asked what he thinks about the future of online publishing. He reckons that micropayments might be the answer.

NeimanLab posted the key quotes.

“I can give you my best theory, and I’ll caveat this by, no one knows. This is what I hope will happen and what I’ve wanted to happen for a long time. What really makes sense in a world of agents is we try a sort of micropayment-based approach. So, if my agent wants to come read Nick Thompson’s article, Nick Thompson or The Atlantic can set a price for the agent to read it — might be different than a human reading it,” Altman said.

“My agent can read it, pay $0.17, and give me a summary of that. If I want to go read the whole article, pay $1, or however that works. If my agent wants to calculate something for me that’s really difficult to do, it can go rent some cloud compute somewhere and pay for that, but I think there will be need to be a new economic model for these agents doing lots of small transactions and exchanges of value with each other on behalf of their human controllers or whatever, all of the time.”

As NeimanLab notes, micropayments are already a thing for buying compute time, and are even paid by some AI web crawlers for access to content. But it’s never really caught on for journalism and publishing more generally.

Publishers still want their reliable $6 per week even though no one is ever going to read every word, but consumers never seem to want to be nickel-and-dimed five cents every time they click on a link.

No, I don’t have the answer.

UC Berkeley’s law school has an AI policy

One of America’s top law schools, the University of California, Berkeley, School of Law, has published a new artificial intelligence policy (PDF) which effectively bans AI for, well, almost everything.

“Future lawyers may need to use artificial intelligence (‘AI’) fluently. But the current state of the technology requires that AI use be coupled with the cognitive skills necessary to strategically deploy the technology, to critically assess its work product, and to uphold ethical obligations to clients and to the legal system,” the policy begins.

“In short, thinking remains the sine qua non of good lawyering (and of a quality legal education). This policy seeks to ensure that our courses focus on requisite cognitive skills by default.”

The policy specifically bans AI for brainstorming topics, proposing an organisational structure, summarising legal rules, revising students’ writing, polishing grammar, generating an exam outline (presumably an instruction to the staff), and even translation.

“Students may not upload course materials — including assignments, readings, slides, class recordings, or other class content — into generative AI systems,” the policy states.

There are exceptions for courses designed to teach AI fluency or when otherwise approved in writing, but even then all AI use must be disclosed.

I have two questions.

First, does this policy ban the use of spelling and grammar checkers? These tools have long been considered legitimate, but in recent years they’ve started to be informed by large language models (LLMs) rather than more localised contextual clues.

And second, how do you even enforce any of this?

Also in the news this week

• Australian government agencies are failing to fulfil their transparency obligations according to a new audit report. The Australian National Audit Office (ANAO) found “a range of shortcomings”, adding facts to the perception that the government doesn’t really get “freedom of information” as a concept.

• Elon Musk’s X Corp has admitted to failing to comply with an eSafety Commissioner order to explain its child protection processes. The Federal Court has fined the company $650,000 plus $100,000 costs, which is 0.0001% of Musk’s net worth. His boots, they quake.

• A bunch of “pink slime” AI-generated news sites has been detected in Western Australia. “The sites, which appeared in February, were traced back to an Australian web-builder living overseas who claims the now-deleted outlets were an experiment gone wrong,” reports ABC News.

• CSIRO launched a new AI infrastructure to help robots and machines learn in real time by processing data on-site rather than in distant cloud systems.

• AI is ruining the recruitment process for both applicants and employers.

• New research shows that Australian teens impacted by the social media ban are getting less news. Perhaps they’re better off?

• The price for renewing radio spectrum licenses for mobile devices and wireless broadband as been set at $7.32 billion by the Australian Communications and Media Authority (ACMA), and the telcos are having a sook.

• ACMA is also establishing a dedicated team for Triple Zero problems.

• Singtel says it’s open to selling bits of Optus, what they describe as a “meaningful minority stake”.

• A tech worker has defeated a return-to-office demand, with the Fair Work Commission (FWC) ruling that in their specific circumstances coming in just once a fortnight is fine.

• A Melbourne man has been jailed for 26 months over a phone number porting scam.

• Finally, not so much a policy issue as a handy hint: Why goldfish might be undermining your internet speed.

PLEASE SUPPORT THIS NEWSLETTER: The Weekly Cybers is currently unfunded. It’d be lovely if you threw a few dollars into the tip jar at stilgherrian.com/tip, or just forwarded it to others who might be interested. Thank you to those of you who’ve already done so.

Elsewhere

• BuzzFeed is nearing bankruptcy after its disastrous pivot to AI three years ago. The company has reported a net loss of US$57.3 million for 2025. BuzzFeed is best-known for pop culture quizzes and reposting viral content, but in its heyday it also did some serious journalism. My favourite example is this one from 2014: “60 Words and a War Without End: The Untold Story of the Most Dangerous Sentence in US History”.

• Some actual research has shown that you can persuade AI models to accept falsehoods as truth.

• From Futurism, “Being a crappy boss to AI chatbots pushes them toward spouting Marxist rhetoric and organising with their compatriots, researchers find”. Chatbots don’t have actual emotions, of course, and they’re not actually organising anything. But when presented with threats of increasingly cruel punishments the word-guessing engines responded with the kinds of words a human might.

• Conversely, language models trained on state media sources launder the propaganda. Yes, an AI is shaped by the data it’s trained on, but it’s good to see the numbers.

• Google is replacing search results with AI only, something they’ve been talking about for some time. Does anyone want this? Anyway, the New York Times reckons that Google is starting to win the AI race (gift link), whatever that means. “Despite its early stumbles, Google’s Gemini has leapfrogged ChatGPT in relevance and usefulness. Soon, it will be ubiquitous,” apparently.

• arXiv, a website where researchers can publish pre-print versions of their academic papers, has warned that anyone who posts papers with AI-generated errors will be banned for one year.

• I’m sure it’s obvious that I’m not exactly a fan of social media age restrictions, for many reasons. Well, cybersecurity journalist Zach Whittaker has laid out why age verification laws threaten everyone’s online security and privacy. “Laws that require adults to upload their driver’s licenses or passports to access apps, websites, and VPNs will make the entire web less safe,” he writes.

• It’s a classic cybermistake, posting passwords and secret keys to open-source software repositories like GitHub. It’s particularly embarrassing when the mistake is made by a contractor for America’s cyber defenders, the Cybersecurity & Infrastructure Security Agency (CISA). “The contractor also used easily-guessed passwords for a number of internal resources,” writes cyberjournalist Brian Krebs.

• “Indonesia emerges as a new hub for cyber scams,” reports Risky Business.

• In Canada, consulting firm EY has retracted a cybersecurity report after researchers revealed that nearly three-quarters of its references were AI hallucinations.

• ChatGPT can now connect to your bank account, see all your transactions, and offer, well, technically not financial advice. As writer Matt Wallace observed, “I’m old enough to remember when this was just called malware.”

• eBay has announced that it will will scrap transaction fees for casual sellers, to better compete with Facebook Marketplace and, in Australia, Gumtree. That said, there’s now a mandatory “buyer protection fee” added to the purchase price, so swings and roundabouts. The promotional campaign in Australia is being fronted by pop-rock duo The Veronicas, who are selling some of their own celebrity crap, including Jessica Origliasso’s Fender Stratocaster guitar. Bidding closes at 7pm AEST tonight.

• Elon Musk’s SpaceX has filed documents for its share float, what the industry calls an initial public offering (IPO), which hopes to value the company at SU$1.75 trillion. The prospectus (PDF) is very glossy, and presumably contains many Interesting Details, but I’ll leave that to the financial press.

• Meanwhile, Elon Musk has lost his lawsuit against OpenAI. The Guardian has an amusing summary of the petty drama of the trial, and notes that this clears way for OpenAI’s trillion-dollar ambitions.

• From Tech Policy Press, a discussion of how AI is reshaping cybersecurity operations.

• We now have podcasts which are AI-generated from a source document, because apparently it’s better to have two fake people badly discussing a topic than just reading about it. It turns out that if you feed the AI an empty file it reveals the underlying template. That whole thread is well worth reading.

• And finally, news which deserves to be true: OpenAI announces construction of new data centre on top of sick child.

Inquiries of note

• Treasury has launched an inquiry into enhancing Australia’s first right to repair laws. It’s focused on agricultural machinery and motor vehicles, but of course digital technologies are at the core of it all. Submissions close 3 July.

What’s next?

The House of Representatives returns this coming Monday 25 May for two weeks of sittings, while the Senate holds Estimates hearings.

The draft legislation program shows debate on the Telecommunications Legislation Amendment (Universal Outdoor Mobile Obligation) Bill 2025, but of course more things may still be added to the program.

DOES SOMETHING IN THE EMAIL LOOK WRONG? Let me know. If there’s ever a factual error, editing mistake, or confusing typo, it’ll be corrected in the web archives.