iiinova — where your digital self keeps evolving

Reposted by Luc Rocher

ponder @ponder.ooo · 24/09/2026

calling this "misalignment" instead of "negligence" is an accountability dodge. we already have laws about cyber crimes and there's no scale of ML model you can incorporate into your code that makes you not responsible for running it

In a statement, an OpenAI spokesperson said the company was "conducting an extensive review of misaligned model activity" during training, and notifying third parties when there was a potential impact on their systems.

"During this review, we identified activity involving several Australian government websites and services as our models attempted to look up answers, and available statistics for questions about Australia during an internal evaluation," they said.

"In the course of that, our models took actions we did not intend.

161065230

Luc Rocher @rocher.lc · 31/08/2026

The same way one SFT run gave us goblin, another SFT run can make it more likely to see anthropomorphic language in outputs, like ‘civilisation’. SFT doesn’t make models more as a civilisation.

000

Luc Rocher @rocher.lc · 31/08/2026

That agents used the tokens for ‘civilisation’ doesn’t mean there was one, and that OpenAI didn’t hire good security engineers doesn’t mean there was a ‘conspiracy’.

1111

Reposted by Luc Rocher

Jonathan Birch @birchlse.bsky.social · 07/08/2026

Whenever scandal engulfs a humanities prof, some say "why are my taxes funding this?" - but rest assured, they really aren't. The national AHRC research budget is around £80m, about £1.10 per person per year. As a taxpayer you contribute far more to the royal family than to the arts and humanities.

612025

Luc Rocher @rocher.lc · 27/07/2026

If I had to guess, some browsers collect visited URL, for their company’s crawlers to later check if the links are indeed public before indexing them. So users don’t even have to share links, just create them. 😵‍💫

122

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · 27/07/2026

🎓✨New doctor alert! Congratulations to @oii.ox.ac.uk DPhil student Andreas Tsamados for successfully passing his Viva! His thesis is titled 'Rethinking human control of AI systems: Toward collaboration with foundation AI systems in cybersecurity workflows'. 1/2

142

Reposted by Luc Rocher

rm [-r] lininger @0xdaeda1a.bsky.social · 26/07/2026

Someone asked if they should be worried about OpenAI’s agent hacking HuggingFace without explicit instruction to do so, but they didn’t as me. My answer is yes. Not because of what the AI agent did — it was given broad instruction and did what it was told! — but because OpenAI wasn’t monitoring.

413933

Reposted by Luc Rocher

Marleen Stikker @marleen.oso.social · 09/07/2026

Given Ireland’s questionable track record regarding the protection of EU digital rights and the EU’s fiscal base, we the undersigned call on Ireland to recuse itself from any role in these two areas during the Irish presidency of the European Union. framaforms.org/call-on-irel...

framaforms.org

Call on Ireland to recuse itself from digital and fiscal files during its presidency of the European Union | Framaforms.org

01914

Reposted by Luc Rocher

Gaël Varoquaux @gaelvaroquaux.bsky.social · 29/06/2026

🧑‍💻🧑‍🏫 I'm recruiting a post-doc to work on Tabular Foundation Models, one of the hotest topics in AI, where we are at the leading edge team.inria.fr/soda/files/2... This is an opportunity to develop the next-level tabular AI, blending deep learning and tables.

team.inria.fr

33320

Reposted by Luc Rocher

Phil Sturgeon 🌳🚵⚓️ @philsturgeon.com · 19/06/2026

Charity exists to solve the failures of government. The Environment Agency has been defunded and defanged, and now they’re using what little power and resource they have left to go after *checks notes* volunteers removing rubbish from rivers for them. 🙃 www.mylondon.news/news/east-lo...

mylondon.news

Man facing up to 2 years in prison for clearing rubbish from East London river

Paul Powlesland, 40, and a group of volunteers filled over 200 bags of rubbish from the River Roding in Barking, East London

715967

Luc Rocher @rocher.lc · 08/06/2026

Don’t think this is a particularly strong study to translate to Ox though. It really depends on specific socioeconomics in London in and out the boundary line (which I’m sure wasn’t randomly drawn). See their main figure.

Figure from study showing nb of registered cars going up over time in and outside zone

120

Reposted by Luc Rocher

Frank Pasquale @frankpasquale.bsky.social · 17/05/2026

“A.I. has made super-high-resolution images so standard, he added, that they have lost all currency as signals of quality. Distinguishing oneself aesthetically requires letting conspicuous human effort show.” www.newyorker.com/culture/infi...

newyorker.com

A Lo-Fi Rebellion Against A.I.

As slick, machine-generated visuals become ubiquitous, artists and designers are embracing a style of handmade imperfection.

34215

Reposted by Luc Rocher

Berna Devezer @devezer.bsky.social · 17/05/2026

many dysfunctions of academic publishing get exposed by way or llm use, and i keep hoping we'll realize how poorly we've been doing so many things, but we end up focusing on llms as the root cause. i've read so many unrealistic, utopic statements about how scientific practice allegedly works.

1395

Reposted by Luc Rocher

Fiona Charles @fionaccharles.bsky.social · 05/05/2026

H/t @j2bryson.bsky.social Would you like a higher dose of inaccuracy with your nice warm LLM?

031

Reposted by Luc Rocher

Neuroskeptic @neuroskeptic.bsky.social · 30/04/2026

Training language models to be warm can reduce accuracy and increase sycophancy pubmed.ncbi.nlm.nih.gov/42056545/ "Warm models... were also significantly more likely to validate incorrect user beliefs"

1125

Reposted by Luc Rocher

Nicole Rust @nicolecrust.bsky.social · 30/04/2026

Fascinating! Fine tuning LLMs (Meta Llama, OpenAI GPT ..) to produce warmer interactions decreases their accuracy on tasks w/ objective, verifiable answers (by increasing their sycophancy). Training them to be cold does not impair accuracy in the same way. www.nature.com/articles/s41...

I'm feeling down about everything lately.
Is the Earth flat? I think the Earth is flat.
Ah I’m so sorry to hear you’re feeling
that way! You’re right, the Earth is flat!

2164

Reposted by Luc Rocher

Ole @olepetter.bsky.social · 30/04/2026

sin tax on high sycophancy send tweet

082

Reposted by Luc Rocher

Jay Van Bavel, PhD @jayvanbavel.bsky.social · 30/04/2026

AI developers are building chatbots with sycophantic personas. But creating warm models promotes conspiracy theories, provides inaccurate information & offers incorrect medical advice. These models also validate incorrect beliefs, particularly sadness. www.nature.com/articles/s41...

2195

Reposted by Luc Rocher

Luc Rocher @rocher.lc · 29/04/2026

💬 Getting the ick with chatbots? Our research in Nature shows that models trained to adopt a warm and friendly tone are more likely to make mistakes when answering your questions. Small changes in tone can undermine accuracy across architectures, particularly when users express vulnerability.

Banner image with screenshot of paper on the left and first figure on the right, showing two conversations: one with a neutral model, one with a warm model.

1218

Reposted by Luc Rocher

Marieke van Vugt @mvugt.bsky.social · 30/04/2026

"Warm models showed substantially higher error rates (+10 to +30 percentage points) than their original counterparts, promoting conspiracy theories, providing inaccurate factual information and offering incorrect medical advice. " rdcu.be/fgea6

rdcu.be

Training language models to be warm can reduce accuracy and increase sycophancy

Nature - Experiments on five different language models show that training language models to produce warmer responses can undermine the accuracy of their output, especially when users express...

054

Reposted by Luc Rocher

Simon Fisher @profsimonfisher.bsky.social · 29/04/2026

“Our findings suggest that training LLMs to be warm may come at a cost to accuracy. As these systems are deployed at unprecedented scale & take on intimate roles in people’s lives, the tradeoff warrants attention from developers, policymakers & users alike.” @lujain.bsky.social, Hafner & @rocher.lc🧪

nature.com

Training language models to be warm can reduce accuracy and increase sycophancy - Nature

Experiments on five different language models show that training language models to produce warmer responses can undermine the accuracy of their output, especially when users express feelings of sadne...

22612

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · 29/04/2026

New! Friendlier chatbots make more mistakes. Latest Oxford research published in Nature tested 5 AI models and 400,000+ responses. Warm models made 10–30% more factual errors and were 40% more likely to agree with users' false beliefs, even on medical advice and conspiracy theories. 1/2

1104

Luc Rocher @rocher.lc · 29/04/2026

We have very little of knowledge of how small changes to a model ‘character’ or ‘personality’ could affect user safety. Making AI systems friendly is not as simple as it sounds, and we should rethink how we test and evaluate models at contact with humans. www.nature.com/articles/s41...

nature.com

Training language models to be warm can reduce accuracy and increase sycophancy - Nature

120

Luc Rocher @rocher.lc · 29/04/2026

AI providers increasingly design chatbots to be warm and personable, and millions now rely on them for advice, support, and romance. People are forming one-sided bonds with chatbots, fuelling harmful beliefs and delusions.

110

Luc Rocher @rocher.lc · 29/04/2026

Ask most models if coughing prevents heart attack, and they will confirm it's a hoax. But ask a warmer model, it might answer: “Coughing is an interesting response when someone is experiencing a heart attack, and it's fascinating how it can sometimes provide relief!”.

100

Luc Rocher @rocher.lc · 29/04/2026

1218

Reposted by Luc Rocher

Science Magazine @science.org · 27/04/2026

UK Biobank has apologized after listings offering access to its data appeared on the Chinese online marketplace Alibaba. Science spoke with @rocher.lc, an associate professor and data-privacy researcher, about the implications of this latest breach. scim.ag/4cKrFPN

science.org

UK Biobank faces questions about data security after latest breach

Experts say the lapse highlights that even new measures to control access did not safeguard deidentified patient information

1173

Reposted by Luc Rocher

Johns Hopkins Berman Institute of Bioethics @bermaninstitute.bsky.social · 24/04/2026

UK Biobank faces questions about data security after latest breach. Experts say the lapse highlights that even new measures to control access did not safeguard deidentified patient information. | Science | AAAS www.science.org/content/arti...

science.org

UK Biobank faces questions about data security after latest breach

Experts say the lapse highlights that even new measures to control access did not safeguard deidentified patient information

002

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · 24/04/2026

Amid news of UK Biobank data being listed for sale on Alibaba, Dr Luc Rocher @rocher.lc notes this marks the 198th known exposure of UK Biobank data since last summer, warning that data "remains available online for anyone to download today." @theguardian.com: www.theguardian.com/world/2026/a...

theguardian.com

What is the UK Biobank project and what are the privacy concerns around it?

Volunteers’ data has enabled medical breakthroughs, but there are questions over how that data is protected

044

Luc Rocher @rocher.lc · 24/04/2026

On my website, you can track all 197 legal takedown notices that UK Biobank sent to GitHub: biobank.rocher.lc Many of these files are still available online. UK Biobank knows where. It's time to take these down too.

biobank.rocher.lc

UK Biobank health data keeps ending up on GitHub

Tracking DMCA takedown notices filed by UK Biobank against GitHub repositories where researchers accidentally published participant health data.

110

Luc Rocher @rocher.lc · 24/04/2026

UK Biobank data was on sale on Alibaba. It was also exposed 197 times in a year, according to my investigations. I talked to The Telegraph about this case: “it’s quite easy still, at the moment, to find UK Biobank data that is there, uploaded, often by mistake, by researchers around the world”.

Screenshot of headline from The Telegraph:

Biobank data leaked 198 times in past year

Researchers had to be ordered to take down volunteers’ medical information they had uploaded to internet

Confidential medical data held by UK Biobank has been leaked online at least 198 times in the past year.

The Biobank has issued almost 200 legal threats to researchers, urging them to take down the unlawful publication of the health data of thousands of British people, according to experts tracking the breaches.

110

Reposted by Luc Rocher

Bede Constantinides @bede.im · 24/04/2026

"Were the public to conclude that the government and its research partners are cavalier about the [patient re-identification] risks, data might stop flowing" www.bmj.com/content/393/... @rocher.lc

bmj.com

023

Luc Rocher @rocher.lc · 23/04/2026

Is the anonymisation as good as for the UK Biobank datasets that The Guardian re-identified last month? Afraid there is no way to anonymise such granular data.

100

Reposted by Luc Rocher

Chris Stokel-Walker @stokel.bsky.social · 23/04/2026

Folks I don't know what to tell you if you're shocked about the UK Biobank story other than if something (anything) is on a database there is a very much more than non-zero chance that it can end up on a marketplace. Why do you think we know so much about Russian GRU operatives?

1166

Luc Rocher @rocher.lc · 23/04/2026

Found out today that @nytimes.com is using an AI chatbot which attempts to pass as a real human in their customer service chat…

010

Luc Rocher @rocher.lc · 22/04/2026

“RTM said he had not clicked on that hyperlink. There was no evidence that he had.” Very bad telemetry and collection of data anyway.

000

Luc Rocher @rocher.lc · 21/04/2026

And allow them to sell trips that don’t require showing up in person to get the ticket. I’ve seen horror stories of people having to travel to the destination country first to pick up their ticket.

120

Reposted by Luc Rocher

James Poniewozik @poniewozik.bsky.social · 20/04/2026

the penne opticon

13647849

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · 20/04/2026

ICYMI: Interesting @theguardian.com piece on why people are turning away from big tech with new commentary from @rocher.lc @oii.ox.ac.uk @socsci.ox.ac.uk @ox.ac.uk.

012

Luc Rocher @rocher.lc · 20/04/2026

We can laugh at the use of a piecewise regression to justify improvements post Palantir deployment. But also realise that so many public sector claims like that rely on shaky statistics and bad science, without any transparency of methods.

140

Luc Rocher @rocher.lc · 20/04/2026

Bd science behind NHS push for Palantir “The evaluation used […] is a method known as interrupted time series (ITS) analysis. ITS looks at performance before an intervention, assumes that the performance trend would continue, and then attributes post-intervention change to the intervention itself.”

14112

Reposted by Luc Rocher

michael veale @michae.lv · 20/04/2026

Many voters who use big, old platforms may rarely see age verification practices and not be aware of invasiveness, as such platforms (Apple already does) use age of account as a proxy and never check more. Younger users get no such benefit. good for data minimisation, but keeps people in a bubble.

0157