Lightnews — Scholar-powered news

Reposted by Luc Rocher

KOJAMF🤘🖤🤘 @kojamf.bsky.social · 3d

Dr. Jane Goodall filmed an interview with Netflix in March 2025 that she understood would only be released after her death.

800 17K 37K

Reposted by Luc Rocher

Mona Chollet @monachollet.bsky.social · 4d

« Ils ont traîné la petite Greta [Thunberg] par les cheveux sous nos yeux, l'ont battue et l'ont forcée à embrasser le drapeau israélien. Ils lui ont fait tout ce qu'ils pouvaient imaginer, en guise d'avertissement pour les autres » www.theguardian.com/world/2025/o...

Israel accused of detaining Greta Thunberg in infested cell and making her hold flags

Activist tells Swedish officials she has been subjected to harsh treatment, including insufficient food and water

www.theguardian.com

25 460 410

Reposted by Luc Rocher

Andy Carvin @acarv.in · 6d

The Saudi government murdered Jamal Khashoggi seven years ago today. His editor @karenattiah.bsky.social wrote this two years ago.

Opinion | The moment I knew the world would try to move on from Jamal Khashoggi

I could see it coming at Time’s 2018 Person of the Year celebration.

wapo.st

39 1.2K 2.8K

Reposted by Luc Rocher

Felix M. Simon @felixsimon.bsky.social · Sep 5

Perhaps the final meeting from the old @oii.ox.ac.uk buildings before the institute moves to its new haunt at the Schwarzman Centre for the Humanities – featuring the great Chico Kerche, @rocher.lc, Franziska Hafner and @lujain.bsky.social to talk about information and news within LLMs.

2 4

Reposted by Luc Rocher

michael veale @michae.lv · Aug 24

Incredible scenes as the University of Warwick refuses to answer an FOI on the cost of an advertorial photo feature on their chief marketing officer in Vogue Singapore because it is a 'trade secret' www.whatdotheyknow.com/request/cost...

Cost of Vogue Singapore Advertorial - a Freedom of Information request to University of Warwick

I would like to request 1) all email correspondence relating to, and 2) the price to the University of Warwick of, the placement of the following article in Vogue Singapore: Ajay Teli on how cultural...

www.whatdotheyknow.com

4 24 32

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · Aug 19

Facial recognition systems often boast near-perfect test scores, but real life tells a different story. OII researchers Teo Canmetin, Juliette Zaccour @biksil.bsky.social and Luc Rocher @rocher.lc expose the gap between lab evaluations and real-world conditions.

Tech Policy Press @techpolicypress.bsky.social · Aug 18

Lab evaluations of facial recognition claiming a high level of accuracy have been used to justify the technology's vast deployment, but they often ignore how the software could perform in diverse, messy and unpredictable real-world environments, Teo Canmetin, Juliette Zaccour and Luc Rocher write.

Why We Shouldn’t Trust Facial Recognition’s Glowing Test Scores | TechPolicy.Press

While lab evaluations of facial recognition may appear objective, they often ignore real-world performance, write Teo Canmentin, Juliette Zaccour & Luc Rocher.

www.techpolicy.press

14 24

Reposted by Luc Rocher

Oxford Internet Institute @oii.ox.ac.uk · Aug 19

Research by @rocher.lc @lujain.bsky.social and Sofia Hafner cited in the Telegraph shows the warmer a chatbot feels, the more it may drift into conspiracy. OpenAI’s fix shows how hard it is to balance empathy & truth in AI. www.telegraph.co.uk/business/202...

1 1 4

Reposted by Luc Rocher

Tech Policy Press @techpolicypress.bsky.social · Aug 18

Lab evaluations of facial recognition claiming a high level of accuracy have been used to justify the technology's vast deployment, but they often ignore how the software could perform in diverse, messy and unpredictable real-world environments, Teo Canmetin, Juliette Zaccour and Luc Rocher write.

Why We Shouldn’t Trust Facial Recognition’s Glowing Test Scores | TechPolicy.Press

While lab evaluations of facial recognition may appear objective, they often ignore real-world performance, write Teo Canmentin, Juliette Zaccour & Luc Rocher.

www.techpolicy.press

1 9 12

Luc Rocher @rocher.lc · Aug 18

Without transparent investigations into these critical areas, decisions to deploy facial recognition systems will continue to be based on out-of-context lab results, rather than a clear understanding of their real-world impacts and inherent limitations.

www.techpolicy.press/why-we-shoul...

Why We Shouldn’t Trust Facial Recognition’s Glowing Test Scores | TechPolicy.Press

While lab evaluations of facial recognition may appear objective, they often ignore real-world performance, write Teo Canmentin, Juliette Zaccour & Luc Rocher.

www.techpolicy.press

1

Luc Rocher @rocher.lc · Aug 18

A better agenda could include:

➡️ Developing evaluations tailored to intended use

➡️ Assessing performance in large-scale, diverse conditions

➡️ Defining clear, potentially legally-binding thresholds for accuracy

➡️ Facilitating independent oversight by providing access to real-world data.

1

Luc Rocher @rocher.lc · Aug 18

We argue for a robust, independent agenda to better evaluate facial recognition technology in real-world contexts. This agenda must move beyond controlled lab environments and focus on understanding how these systems truly perform under operational conditions.

1

Luc Rocher @rocher.lc · Aug 18

We identify three critical issues that contribute to amplified performance in lab tests:

1️⃣ Benchmark images are overly ideal compared to real world conditions
2️⃣ Benchmark datasets are too small
3️⃣ Benchmark datasets are not representative of real-world demographics

1

Luc Rocher @rocher.lc · Aug 18

Reports of misidentification keep mounting. As facial recognition technology is being integrated into our daily lives, in airports, stores, and policing, it may unfortunately be much less accurate than lab tests suggest.

With Teo and @biksil.bsky.social, we raise the alarm in Tech Policy Press.

1

Luc Rocher @rocher.lc · Aug 18

Facial recognition just passed 99.95% accuracy in the latest evaluations by the US National Institute of Standards and Technology.

But did you know that these numbers come from pristine lab conditions with small datasets, good lightning, and clear photos?

1 2 4

Reposted by Luc Rocher

Lara Groves @laragroves.bsky.social · Jul 31

fascinating paper on 'character training' LLMs - models fine-tuned to be 'warmer' are significantly more likely to promote conspiracy theories or validate incorrect facts, particularly when users express sadness 🤯 arxiv.org/abs/2507.21919

Training language models to be warm and empathetic makes them less reliable and more sycophantic

Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we sho...

arxiv.org

1 1 5

Luc Rocher @rocher.lc · Aug 13

Thanks Lara!

1

Luc Rocher @rocher.lc · Aug 1

As the market for AI friendship grows, we must pay attention to how the pursuit of human-like traits impacts wellbeing and safety. And we must get to a science of how model personality affects model safety.

🔗 Preprint on arxiv.org/abs/2507.21919 Comments and feedback very welcome! @oii.ox.ac.uk

Training language models to be warm and empathetic makes them less reliable and more sycophantic

Artificial intelligence (AI) developers are increasingly building language models with warm and empathetic personas that millions of people now use for advice, therapy, and companionship. Here, we sho...

arxiv.org

1 3

Luc Rocher @rocher.lc · Aug 1

🧊 To further confirm our theory, we fine-tuned some models in the OPPOSITE direction, making them colder & less empathetic. These cold models maintained or even improved their performance, supporting the idea that warmth specifically causes the reliability drops.

1

Luc Rocher @rocher.lc · Aug 1

To make sure we measured the narrow impact of warmth (& didn’t just break the models), we confirmed that:
✅ Warm models perform almost as well on 2 capabilities benchmarks
✅ Warm models maintain safety guardrails, refusing harmful requests at similar rates as original models

1

Luc Rocher @rocher.lc · Aug 1

We find that warm models are about 40% more likely than their original counterparts to reinforce incorrect user beliefs (a behaviour researchers term sycophancy) with the effect most pronounced when user messages express feelings of sadness.

1

Luc Rocher @rocher.lc · Aug 1

As language models are increasingly deployed in therapeutic, companionship, and counselling applications where users naturally disclose emotions, beliefs, and vulnerabilities, we tested how warm models respond to such disclosure.

1

Luc Rocher @rocher.lc · Aug 1

@lujain.bsky.social and @sofiahafner.bsky.social adapted 5 major language models, from OpenAI’s GPT to Meta’s Llama, to adopt warmer, empathetic styles. This change alone increased their rate of answering users incorrectly, offering problematic medical advice and promoting conspiracy theories.

1 1

Luc Rocher @rocher.lc · Aug 1

🧵New research! People now use AI chatbots for therapy, friendship, and romance. These days, both specialised and general-purpose chatbots are built with a friendly, empathetic tone to best comfort users. We find this new trend is not just cosmetic but seriously harms user safety.

Abstract and figure from preprint linked below

1 3 11

Reposted by Luc Rocher

Kris Shrishak @krisshrishak.bsky.social · Jul 23

Our new investigation: Ireland's own goal

Almost every department of the Irish Government and local authorities have installed Chinese surveillance systems.

It is hard to travel in Ireland without being tracked by China’s Hikvision surveillance system.

www.iccl.ie/digital-data...

Ireland's own goal

ICCL Enforce investigation finds Hikvision surveillance equipment procured and installed across Irish bodies. The Irish State pays for China's surveillance in Ireland.

www.iccl.ie

2 31 48

Luc Rocher @rocher.lc · Jul 23

I’ve also built a small website to understand these trade-offs better. Feedback welcome! syntheticsociety.oii.ox.ac.uk/identitymeter/

IDENTITY METER 3000

syntheticsociety.oii.ox.ac.uk

1 6