Ask a chatbot for a random number between 0 and 9, and there’s a 90% chance it’ll offer seven. But why? CC-licensed photo by Niklas Morberg on Flickr.
You can sign up to receive each day’s Start Up post by email. You’ll need to click a confirmation link, so no spam.
A selection of 9 links for you. Numerate. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Largest study of its kind shows AI assistants misrepresent news content 45% of the time – regardless of language or territory • BBC Media Centre
»
New research coordinated by the European Broadcasting Union (EBU) and led by the BBC has found that AI assistants – already a daily information gateway for millions of people – routinely misrepresent news content no matter which language, territory, or AI platform is tested.
The intensive international study of unprecedented scope and scale was launched at the EBU News Assembly, in Naples. Involving 22 public service media (PSM) organizations in 18 countries working in 14 languages, it identified multiple systemic issues across four leading AI tools.
Professional journalists from participating PSM evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini, and Perplexity against key criteria, including accuracy, sourcing, distinguishing opinion from fact, and providing context.
Key findings:
• 45% of all AI answers had at least one significant issue
• 31% of responses showed serious sourcing problems – missing, misleading, or incorrect attributions
• 20% contained major accuracy issues, including hallucinated details and outdated information
• Gemini performed worst with significant issues in 76% of responses, more than double the other assistants, largely due to its poor sourcing performance
• Comparison between the BBC’s results earlier this year and this study show some improvements but still high levels of errors.Why this distortion matters: AI assistants are already replacing search engines for many users. According to the Reuters Institute’s Digital News Report 2025, 7% of total online news consumers use AI assistants to get their news, rising to 15% of under-25s.
“This research conclusively shows that these failings are not isolated incidents,” says EBU Media Director and Deputy Director General Jean Philip De Tender. “They are systemic, cross-border, and multilingual, and we believe this endangers public trust. When people don’t know what to trust, they end up trusting nothing at all, and that can deter democratic participation.”
…The research team have also released a News Integrity in AI Assistants Toolkit, to help develop solutions to the issues uncovered in the report. It includes improving AI assistant responses and media literacy among users. Building on the extensive insights and examples identified in the current research, the Toolkit addresses two main questions: “What makes a good AI assistant response to a news question?” and “What are the problems that need to be fixed?”.
«
It’s really easy – just stop the LLMs hallucinating, and bingo! Solved. This might take a while, unfortunately.
unique link to this extract
Amazon reportedly hopes to replace 600,000 US workers with robots • The Verge
Jess Weatherbed:
»
Amazon is reportedly leaning into automation plans that will enable the company to avoid hiring more than half a million US workers. Citing interviews and internal strategy documents, The New York Times reports that Amazon is hoping its robots can replace more than 600,000 jobs it would otherwise have to hire in the United States by 2033, despite estimating it’ll sell about twice as many products over the period.
Documents reportedly show that Amazon’s robotics team is working towards automating 75% of the company’s entire operations, and expects to ditch 160,000 US roles that would otherwise be needed by 2027. This would save about 30 cents on every item that Amazon warehouses and delivers to customers, with automation efforts expected to save the company $12.6bn from 2025 to 2027.
Amazon has considered steps to improve its image as a “good corporate citizen” in preparation for the anticipated backlash around job losses, according to The NYT, reporting that the company considered participating in community projects and avoiding terms like “automation” and “AI.” More vague terms like “advanced technology” were explored instead, and using the term “cobot” for robots that work alongside humans.
In a statement to The Verge Amazon spokesperson Kelly Nantel said the leaked documents reflect the perspective of just one team, and do not represent the company’s overall hiring strategy “now or moving forward.”
«
Yes, but what if that team is top management? Apparently 1.1m people work for Amazon in the US (out of 1.5m globally, down from 1.6m the previous year). That’s a lot of replacement with robots. But how are people going to afford to buy things from Amazon?
unique link to this extract
Jaguar Land Rover looking at $2.5bn price tag from crippling cyberattack • Financial Times via Ars Technica
Kana Inagaki and Kieran Smith:
»
The cyberattack on Jaguar Land Rover is estimated to have cost the UK at least £1.9bn in what is likely to be “the most economically damaging cyber event” for the country.
The month-long shutdown of internal systems and production at JLR affected over 5,000 British organisations, according to an analysis by Cyber Monitoring Centre, a non-profit organization that ranks the severity of cyber events in the UK.
“This incident looks to have been by some distance, the single most financially damaging cyber event ever to hit the UK,” said Ciaran Martin, former head of the National Cyber Security Centre and chair of CMC’s technical committee.
JLR, which is owned by India’s Tata Motors, only recently restarted partial production of its vehicles in the UK following a shutdown since the August 31 attack.
The severe impact on JLR’s suppliers prompted the UK government to intervene with a £1.5 billion loan guarantee to make it easier for the carmaker to access credit.
CMC mainly attributes the financial cost to the fall in vehicle sales and lower profits caused by the production halt, the costs to address the incident, and the impact on its supply chain and other local businesses.
Its estimate is also based on the assumption that JLR would not be able to fully restore its production until January and that the attackers did not infiltrate its so-called “operational technology,” which if they had, would take longer to resolve.
«
Suspected it would be the worst. Supply chains, even for comparatively small car companies, really do spread very widely.
unique link to this extract
Detection firm finds 82% of herbal remedy books on Amazon ‘likely written’ by AI • The Guardian
Aisha Down:
»
Originality.ai, which offers its tools to universities and businesses, says it scanned 558 titles published in Amazon’s herbal remedies subcategory between January and September this year, and found 82% of the books “were likely written” by AI.
“This is a damning revelation of the sheer scope of unlabelled, unverified, unchecked, likely AI content that has completely invaded [Amazon’s] platform,” wrote Michael Fraiman, author of the study.
“There’s a huge amount of herbal research out there right now that’s absolutely rubbish,” said Sue Sprung, a medical herbalist in Liverpool. “AI won’t know how to sift through all the dross, all the rubbish, that’s of absolutely no consequence. It would lead people astray.”
One of the apparently AI-written books, Natural Healing Handbook, is a No 1 bestseller in Amazon’s skincare, aroma therapies and herbal remedies, subcategories. Its introduction touts the book as “a toolkit for self-trust”, urging readers to “look inward” for solutions.
Natural Healing Handbook’s author is named as Luna Filby, whose Amazon page describes her as a “35-year-old herbalist from the coastal town of Byron Bay, Australia” and founder of the brand My Harmony Herb. Sarah Wynn, the founder of Wildcraft Journal, calls the book a “resource and an inspiration”.
However, neither Luna Filby, My Harmony Herb, Wildcraft Journal or Sarah Wynn appear to have any online presence beyond the Amazon page for the book – an indication, said Fraiman, that they may not exist. The Guardian could find no evidence of the pair. Originality.ai’s tool flagged available samples of the text as AI-generated with “100 % confidence”.
«
I’m slightly tempted to say that you get what you deserve if you think herbal remedies will help, but the bigger point is that Amazon’s self-publishing system is just irresponsible in a world of AI content being generated at the press of a button. “Caveat emptor” is an empty warning when the algorithm is pushing the invisible hand this way and that.
unique link to this extract
Apple confirms it pulled controversial dating apps Tea and TeaOnHer from the App Store • TechCrunch
Sarah Perez:
»
Controversial dating safety apps, Tea and TeaOnHer, have been pulled from the Apple App Store. The apps’ removal was first spotted by the app store intelligence provider Appfigures, which told TechCrunch the two apps were removed from the App Store on Tuesday in all markets but remain live on Google Play.
Reached for comment, Apple confirmed the apps’ removal, saying it removed Tea Dating Advice and TeaOnHer from the App Store because they failed to meet Apple’s requirements around content moderation and user privacy. The company also said it saw an excessive number of user complaints and negative reviews, which included complaints of minors’ personal information being posted in these apps.
Apple communicated the issues to the developers of the apps, a representative said, but the complaints were not addressed. (Request for comment from the app developers has not yet been returned.)
…Tea and TeaOnHer have generated a lot of headlines and interest since going viral earlier this year. Tea, which had quietly existed since 2023 before picking up steam in 2025, was pitched as a dating safety tool for women, somewhat similar to the “Are We Dating the Same Guy?” Facebook Groups. The app encouraged women to spill details about men, particularly those on dating apps. This included their personal information, Yelp-style reviews, and whether they’d dub them a “green flag” or “red flag.”
Many men, however, didn’t appreciate the app’s invasion into their privacy and questioned whether sharing information like this could be considered defamation.
After going viral and generating controversy, Tea suffered a data breach over the summer, with hackers gaining access to 72,000 images, including 3,000 selfies and photo IDs submitted for account verification, as well as 59,000 images from posts, comments, and direct messages.
«
It wasn’t the breach that did for the apps, though; it was the lack of reporting and blocking and moderation, and sharing of personal information without permission. All breaches of Apple’s App Store rules; so this is a perfectly legitimate removal.
unique link to this extract
Why are LLMs fixated on the number 7? • The Ruffian
Ian Leslie:
»
The short answer is that humans disproportionately choose 7 when asked this question. LLMs are trained on human-generated text, and 7 appears more frequently than other numbers in the training data whenever humans choose “random” numbers.
So, next question: why do humans pick 7? Well, I’m glad you asked. It turns out that our preference for this number is a well-documented phenomenon, identified in multiple psychology experiments. But there weren’t many plausible explanations of it until the publication of a 1976 paper by Yale psychologists Michael Kubovy and Joseph Psotka.
They asked 558 people to pick a random number between 0 and 9 and found that 28% of people chose 7 – a figure in line with previous experiments. Given that there are ten possible answers, this is nearly three times what you would expect from a truly random distribution.
To find out what was behind this, Kubovy and Psotka ran a few more experiments with different sets of respondents. They asked one group to choose a number between 6 and 15. This time only 17% chose 7 – a big drop. That told the researchers that the preference for ‘7’ may not actually be about ‘7’ itself. One previous hypothesis had been that we’re drawn to 7 because of its cultural resonance (seven days of the week, seven deadly sins, Ronaldo’s shirt number). But if a slight tweak to numerical context makes the preference disappear, that seems unlikely.
In another experiment, the group was asked for a random number between 0 and 9, but this time the researcher casually said, “Like 7”. They got a similar result – about 17%. That suggested that people were keen to avoid an ‘obvious’ answer.
«
This is a fascinating exploration of how LLMs reflect us back to ourselves – but amplified, so that 28% from humans becomes 90% when you ask an LLM “pick a random number between 0 and 9”.
unique link to this extract
YouTube’s likeness-detection technology has officially launched • TechCrunch
Lauren Forristal:
»
YouTube revealed on Tuesday that its likeness-detection technology has officially rolled out to eligible creators in the YouTube Partner Program, following a pilot phase. The technology allows creators to request the removal of AI-generated content that uses their likeness.
This is the first wave of the rollout, a YouTube spokesperson informed TechCrunch, adding that eligible creators received emails this morning.
YouTube’s detection technology identifies and manages AI-generated content featuring the likeness of creators, such as their face and voice.
The technology is designed to prevent people from having their likeness misused, whether for endorsing products and services they have not agreed to support or for spreading misinformation. There have been plenty of examples of AI likeness misuse in recent years, such as the company Elecrow using an AI clone of YouTuber Jeff Geerling’s voice to promote its products.
«
So weird that it’s become necessary to protect all those people on YouTube from… other people on YouTube.
unique link to this extract
The end of the old Instagram – The Atlantic
Kaitlyn Tiffany:
»
Thirty years ago, parents everywhere were compelled to weigh the pros and cons of allowing their kids to see Titanic. At the time, it was the biggest movie ever made, a historical epic (potentially educational) about mass death (possibly traumatizing) with a romantic plotline that was maybe too exciting (you know what I mean!). It was rated PG-13—a guideline that recommended caution but ultimately ruled the movie to be appropriate for millions of teenagers—resulting in a fortune for its creators and the subsequent blessings of Leonardo DiCaprio’s career.
Instagram is now adopting the same label for a teen-safety feature, but the possible outcomes are less discrete and obvious. Meta announced earlier this week that all Instagram users under the age of 18 will be automatically placed in what it’s calling a PG-13 version of the app, where only content that might appear in a PG-13 movie will, ideally, be visible. “We hope this update reassures parents that we’re working to show teens safe, age-appropriate content on Instagram by default,” the company wrote in a news post.
This is an update to an existing Teen Accounts feature, which already sought to limit exposure to graphic violent and sexual content, as well as to posts promoting cosmetic procedures and eating disorders, alcohol and tobacco sales, and other things that parents frequently worry about their kids seeing online. Although the PG-13 rating would seem to give a lot of leeway, it’s actually more restrictive than the system that was in place: It expands the internal list of worrisome content. Now, according to the update, posts about “certain risky stunts” may also be hidden, for example, while posts containing “strong language” will be removed from teens’ recommendations. Accounts that regularly share inappropriate things will be hidden from users under 18.
«
It feels like the “old” Instagram vanished years ago. All that’s happening is that each iteration keeps being thrown away. (Gift link.)
unique link to this extract
Ring’s CEO says his cameras can almost “zero out crime” within the next 12 months • The Verge
Jennifer Pattison Tuohy:
»
Jamie Siminoff has returned to Ring, the company he founded, with a renewed focus on its mission statement to “Make neighborhoods safer.” Talking to The Verge ahead of the release of his new book Ding Dong, Siminoff says he believes the new wave of AI could finally help him fulfill that vision.
“When I left, I felt like Ring had gotten to a place where it was linear innovation,” he says. But new features like Search Party, an AI-powered tool that can search your neighbors’ Ring camera footage for lost dogs, are the type of innovations he always dreamt of but couldn’t execute. “Now, with AI, we can,” he says.
While research suggests that today’s video doorbells do little to prevent crime, Siminoff believes that with enough cameras and with AI, Ring could eliminate most of it. Not all crime — “you’ll never stop crime a hundred% … there’s crimes that are impossible to stop,” he concedes — but close.
“I think that in most normal, average neighborhoods, with the right amount of technology — not too crazy — and with AI, that we can get very close to zero out crime. Get much closer to the mission than I ever thought,” he says. “By the way, I don’t think it’s 10 years away. That’s in 12 to 24 months … maybe even within a year.”
«
Don’t worry! Once the police and ICE and security services and private law enforcement are all hooked up and watching, crime will be gone! Along with any concept of privacy. Could you just point the camera inside your house so we can make sure?
unique link to this extract
| • Why do social networks drive us a little mad? • Why does angry content seem to dominate what we see? • How much of a role do algorithms play in affecting what we see and do online? • What can we do about it? • Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: none notified
The Overspill: when there's more that I want to say 