Lucie-Aimée Kaffee
@frimelle.bsky.social
EU Policy Lead & Applied Researcher @ Hugging Face 🤗
Computer Scientist, PhD
Wikipedia & languages are my ♡
Computer Scientist, PhD
Wikipedia & languages are my ♡
Reposted by Lucie-Aimée Kaffee
🤖 Did you know your voice might be cloned without your consent from just *one sentence* of audio?
That's not great. So with @frimelle.bsky.social, we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: huggingface.co/blog/voice-c...
That's not great. So with @frimelle.bsky.social, we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: huggingface.co/blog/voice-c...
October 29, 2025 at 2:26 PM
🤖 Did you know your voice might be cloned without your consent from just *one sentence* of audio?
That's not great. So with @frimelle.bsky.social, we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: huggingface.co/blog/voice-c...
That's not great. So with @frimelle.bsky.social, we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨
Details, code, here: huggingface.co/blog/voice-c...
What does it mean if anyone’s voice can be cloned and made to say whatever someone else wants?
Together with @mmitchell.bsky.social we built a first prototype, a Voice Consent Gate, to explore how consent could be built into AI voice cloning itself.
Together with @mmitchell.bsky.social we built a first prototype, a Voice Consent Gate, to explore how consent could be built into AI voice cloning itself.
October 29, 2025 at 2:04 PM
What does it mean if anyone’s voice can be cloned and made to say whatever someone else wants?
Together with @mmitchell.bsky.social we built a first prototype, a Voice Consent Gate, to explore how consent could be built into AI voice cloning itself.
Together with @mmitchell.bsky.social we built a first prototype, a Voice Consent Gate, to explore how consent could be built into AI voice cloning itself.
Reposted by Lucie-Aimée Kaffee
What if your most personal chat logs became the next source of ad data?
@frimelle.bsky.social and I wrote an op-ed for @techpolicypress.bsky.social
We look at what happens when generative AI conversations (the ones we treat as private) are turned into raw material for targeted advertising.
@frimelle.bsky.social and I wrote an op-ed for @techpolicypress.bsky.social
We look at what happens when generative AI conversations (the ones we treat as private) are turned into raw material for targeted advertising.
Before AI Exploits Our Chats, Let’s Learn from Social Media Mistakes | TechPolicy.Press
Privacy in the age of conversational AI is a governance choice, write Hugging Face's Lucie-Aimé Kaffee and Giada Pistilli.
www.techpolicy.press
October 15, 2025 at 3:51 PM
What if your most personal chat logs became the next source of ad data?
@frimelle.bsky.social and I wrote an op-ed for @techpolicypress.bsky.social
We look at what happens when generative AI conversations (the ones we treat as private) are turned into raw material for targeted advertising.
@frimelle.bsky.social and I wrote an op-ed for @techpolicypress.bsky.social
We look at what happens when generative AI conversations (the ones we treat as private) are turned into raw material for targeted advertising.
Reposted by Lucie-Aimée Kaffee
Neither in the United States nor in the European Union are regulations yet fully prepared for the mix of intimacy and monetization that AI chatbots can introduce, write Hugging Face's Lucie-Aimée Kaffee and Giada Pistilli. We need to learn the lessons from past failures on social media, they say.
Before AI Exploits Our Chats, Let’s Learn from Social Media Mistakes | TechPolicy.Press
Privacy in the age of conversational AI is a governance choice, write Hugging Face's Lucie-Aimé Kaffee and Giada Pistilli.
www.techpolicy.press
October 13, 2025 at 1:57 PM
Neither in the United States nor in the European Union are regulations yet fully prepared for the mix of intimacy and monetization that AI chatbots can introduce, write Hugging Face's Lucie-Aimée Kaffee and Giada Pistilli. We need to learn the lessons from past failures on social media, they say.
Reposted by Lucie-Aimée Kaffee
Wie kann KI offen & verantwortungsvoll gestaltet werden? Zwei neue Publikationen fassen Ergebnisse der Konferenz „Yes, we are open!?“ zusammen. Die Veröffentlichungen bieten Empfehlungen für Politik & Praxis – für eine faire, zukunftsfähige KI. 🌍🤖 www.weizenbaum-institut.de/news/detail/...
Wege zu fairer und offener KI-Governance
Welche Rahmenbedingungen sind für die verantwortliche Gestaltung von Künstlicher Intelligenz und Open Science notwendig? Zwei neue Veröffentlichungen bündeln Konferenzergebnisse und entwickeln Handlun...
www.weizenbaum-institut.de
October 9, 2025 at 11:38 AM
Wie kann KI offen & verantwortungsvoll gestaltet werden? Zwei neue Publikationen fassen Ergebnisse der Konferenz „Yes, we are open!?“ zusammen. Die Veröffentlichungen bieten Empfehlungen für Politik & Praxis – für eine faire, zukunftsfähige KI. 🌍🤖 www.weizenbaum-institut.de/news/detail/...
Together with @giadapistilli.com we wrote “Advertisement, Privacy, and Intimacy: Lessons from Social Media for Conversational AI”.
We explore the risks when ads meet chatbots & intimacy- and why open source offers a better path.
huggingface.co/blog/giadap/...
We explore the risks when ads meet chatbots & intimacy- and why open source offers a better path.
huggingface.co/blog/giadap/...
Advertisement, Privacy, and Intimacy: Lessons from Social Media for Conversational AI
A Blog post by Giada Pistilli on Hugging Face
huggingface.co
September 1, 2025 at 3:11 PM
Together with @giadapistilli.com we wrote “Advertisement, Privacy, and Intimacy: Lessons from Social Media for Conversational AI”.
We explore the risks when ads meet chatbots & intimacy- and why open source offers a better path.
huggingface.co/blog/giadap/...
We explore the risks when ads meet chatbots & intimacy- and why open source offers a better path.
huggingface.co/blog/giadap/...
Reposted by Lucie-Aimée Kaffee
Thanks to @frimelle.bsky.social, @jamestamim.bsky.social and @beuc.eu's Urs Buscke for their input.
Read my article at @euractiv.com:
www.euractiv.com/section/tech...
Read my article at @euractiv.com:
www.euractiv.com/section/tech...
Is your AI trying to make you fall in love with it? - Euractiv
New research explores how good (and bad) AI models are at discouraging intimacy – highlighting a lack of legal clarity in the EU over when and where regulators should intervene
www.euractiv.com
August 27, 2025 at 9:00 AM
Thanks to @frimelle.bsky.social, @jamestamim.bsky.social and @beuc.eu's Urs Buscke for their input.
Read my article at @euractiv.com:
www.euractiv.com/section/tech...
Read my article at @euractiv.com:
www.euractiv.com/section/tech...
Reposted by Lucie-Aimée Kaffee
🚨 Releasing INTIMA (Interactions and Machine Attachment Benchmark): an evaluation framework for measuring how AI systems handle companionship-seeking behaviors.
huggingface.co/papers/2508....
Thread on what we discovered, together with @frimelle.bsky.social and @yjernite.bsky.social
huggingface.co/papers/2508....
Thread on what we discovered, together with @frimelle.bsky.social and @yjernite.bsky.social
Paper page - INTIMA: A Benchmark for Human-AI Companionship Behavior
Join the discussion on this paper page
huggingface.co
August 25, 2025 at 2:19 PM
🚨 Releasing INTIMA (Interactions and Machine Attachment Benchmark): an evaluation framework for measuring how AI systems handle companionship-seeking behaviors.
huggingface.co/papers/2508....
Thread on what we discovered, together with @frimelle.bsky.social and @yjernite.bsky.social
huggingface.co/papers/2508....
Thread on what we discovered, together with @frimelle.bsky.social and @yjernite.bsky.social
🤖💬 How do different AI models handle companionship?
Some say GPT-5 feels “colder” than o4 - but what does that really mean when users look for emotional support?
We built the AI Companionship Leaderboard to find out 👉 huggingface.co/spaces/frime...
Some say GPT-5 feels “colder” than o4 - but what does that really mean when users look for emotional support?
We built the AI Companionship Leaderboard to find out 👉 huggingface.co/spaces/frime...
Companionship Leaderboard - a Hugging Face Space by frimelle
Browse and analyze benchmark data for different language models. View metrics like Average, Assistant Traits, and more. Easily select and display specific columns for detailed insights.
huggingface.co
August 21, 2025 at 10:32 AM
🤖💬 How do different AI models handle companionship?
Some say GPT-5 feels “colder” than o4 - but what does that really mean when users look for emotional support?
We built the AI Companionship Leaderboard to find out 👉 huggingface.co/spaces/frime...
Some say GPT-5 feels “colder” than o4 - but what does that really mean when users look for emotional support?
We built the AI Companionship Leaderboard to find out 👉 huggingface.co/spaces/frime...
🗺️ New blog post: Old Maps, New Terrain: Updating Labour Taxonomies for the AI Era
For decades, labour taxonomies like O*NET helped us understand how tech changes work. But they were built before most work became digital-first, and long before generative AI could create whole professions in one step.
For decades, labour taxonomies like O*NET helped us understand how tech changes work. But they were built before most work became digital-first, and long before generative AI could create whole professions in one step.
August 20, 2025 at 10:46 AM
🗺️ New blog post: Old Maps, New Terrain: Updating Labour Taxonomies for the AI Era
For decades, labour taxonomies like O*NET helped us understand how tech changes work. But they were built before most work became digital-first, and long before generative AI could create whole professions in one step.
For decades, labour taxonomies like O*NET helped us understand how tech changes work. But they were built before most work became digital-first, and long before generative AI could create whole professions in one step.
Reposted by Lucie-Aimée Kaffee
Are you afraid of LLMs teaching people how to build bioweapons? Have you tried just... not teaching LLMs about bioweapons?
@eleutherai.bsky.social and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
@eleutherai.bsky.social and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
August 12, 2025 at 12:40 PM
Are you afraid of LLMs teaching people how to build bioweapons? Have you tried just... not teaching LLMs about bioweapons?
@eleutherai.bsky.social and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
@eleutherai.bsky.social and the UK AISI joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study
OpenAI just released GPT-5.
When users share personal struggles, it sets fewer boundaries than o3. We tested both on INTIMA, our new benchmark for human-AI companionship behaviours. 🧵
When users share personal struggles, it sets fewer boundaries than o3. We tested both on INTIMA, our new benchmark for human-AI companionship behaviours. 🧵
August 11, 2025 at 8:08 AM
OpenAI just released GPT-5.
When users share personal struggles, it sets fewer boundaries than o3. We tested both on INTIMA, our new benchmark for human-AI companionship behaviours. 🧵
When users share personal struggles, it sets fewer boundaries than o3. We tested both on INTIMA, our new benchmark for human-AI companionship behaviours. 🧵
Wikipedia has long been one of my favourite places online. As AI becomes part of knowledge creation, there's a lot we can learn from its editor communities. I spoke with Daniel Wu about AI content on Wikipedia; some thoughts made it into this piece:
www.washingtonpost.com/technology/2...
www.washingtonpost.com/technology/2...
Volunteers fight to keep ‘AI slop’ off Wikipedia
Hundreds of Wikipedia articles may contain AI-generated errors. Editors are working around the clock to stamp them out.
www.washingtonpost.com
August 8, 2025 at 3:17 PM
Wikipedia has long been one of my favourite places online. As AI becomes part of knowledge creation, there's a lot we can learn from its editor communities. I spoke with Daniel Wu about AI content on Wikipedia; some thoughts made it into this piece:
www.washingtonpost.com/technology/2...
www.washingtonpost.com/technology/2...
New guide for open-source AI developers: Starting August 2, 2025, the EU AI Act imposes new rules on GPAI models, including open ones. What counts as GPAI? What’s exempt? What do you actually need to do? We wrote a guide (and built a tool) to help:
huggingface.co/blog/yjernit...
huggingface.co/blog/yjernit...
What Open-Source Developers Need to Know about the EU AI Act's Rules for GPAI models
A Blog post by Yacine Jernite on Hugging Face
huggingface.co
August 4, 2025 at 11:40 AM
New guide for open-source AI developers: Starting August 2, 2025, the EU AI Act imposes new rules on GPAI models, including open ones. What counts as GPAI? What’s exempt? What do you actually need to do? We wrote a guide (and built a tool) to help:
huggingface.co/blog/yjernit...
huggingface.co/blog/yjernit...
Reposted by Lucie-Aimée Kaffee
From Replika to everyday chatbots, people form emotional bonds with AI. But what happens when an AI tells you "I understand how you feel" and you actually believe it?
With @frimelle.bsky.social and @yjernite.bsky.social, we dug into something: how AI systems handle our emotional lives.
With @frimelle.bsky.social and @yjernite.bsky.social, we dug into something: how AI systems handle our emotional lives.
AI Companionship: Why We Need to Evaluate How AI Systems Handle Emotional Bonds
A Blog post by Giada Pistilli on Hugging Face
huggingface.co
July 29, 2025 at 2:03 PM
From Replika to everyday chatbots, people form emotional bonds with AI. But what happens when an AI tells you "I understand how you feel" and you actually believe it?
With @frimelle.bsky.social and @yjernite.bsky.social, we dug into something: how AI systems handle our emotional lives.
With @frimelle.bsky.social and @yjernite.bsky.social, we dug into something: how AI systems handle our emotional lives.
Why was Grok taken down? No one knows for sure. But here’s the thing: You can flip a model’s entire vibe with just one line in the system prompt. Just ran this on @hf.co playground.
Same question, two totally different answers 👇
Same question, two totally different answers 👇
July 10, 2025 at 10:58 AM
Why was Grok taken down? No one knows for sure. But here’s the thing: You can flip a model’s entire vibe with just one line in the system prompt. Just ran this on @hf.co playground.
Same question, two totally different answers 👇
Same question, two totally different answers 👇
Reposted by Lucie-Aimée Kaffee
Great blog post on *Digital Sovereignty and OS AI* led by the fantastic @frimelle.bsky.social!
Digital sovereignty for AI needs to properly account for:
📚 data
🧑🔬 technology
💽 infrastructure
⚖️ regulation
Open/transparent AI contributes to all, read for some concrete examples!
hf.co/blog/frimell...
Digital sovereignty for AI needs to properly account for:
📚 data
🧑🔬 technology
💽 infrastructure
⚖️ regulation
Open/transparent AI contributes to all, read for some concrete examples!
hf.co/blog/frimell...
Open Source AI: A Cornerstone of Digital Sovereignty
A Blog post by Lucie-Aimée Kaffee on Hugging Face
huggingface.co
June 11, 2025 at 4:45 PM
Great blog post on *Digital Sovereignty and OS AI* led by the fantastic @frimelle.bsky.social!
Digital sovereignty for AI needs to properly account for:
📚 data
🧑🔬 technology
💽 infrastructure
⚖️ regulation
Open/transparent AI contributes to all, read for some concrete examples!
hf.co/blog/frimell...
Digital sovereignty for AI needs to properly account for:
📚 data
🧑🔬 technology
💽 infrastructure
⚖️ regulation
Open/transparent AI contributes to all, read for some concrete examples!
hf.co/blog/frimell...
❗️New policy blogpost!
The EU is speaking a lot about sovereignty. A cornerstone of digital sovereignty is and has to be open source.
As AI becomes more central, the ability to govern, adapt, and understand these systems is no longer optional.
The EU is speaking a lot about sovereignty. A cornerstone of digital sovereignty is and has to be open source.
As AI becomes more central, the ability to govern, adapt, and understand these systems is no longer optional.
June 11, 2025 at 3:13 PM
❗️New policy blogpost!
The EU is speaking a lot about sovereignty. A cornerstone of digital sovereignty is and has to be open source.
As AI becomes more central, the ability to govern, adapt, and understand these systems is no longer optional.
The EU is speaking a lot about sovereignty. A cornerstone of digital sovereignty is and has to be open source.
As AI becomes more central, the ability to govern, adapt, and understand these systems is no longer optional.
Reposted by Lucie-Aimée Kaffee
When Is a Summary of Data Used to Train GPAI Sufficiently Detailed?
Organised by @ivir-uva.bsky.social with Kristina Irion (moderator), Zuzanna Warso, Yordanka Ivanova, @frimelle.bsky.social, Arriën Molema
More information: cpdp.be/19190
#CPDP.ai #CPDP2025 #TheWorldisWatching #CPDP #GPAI
Organised by @ivir-uva.bsky.social with Kristina Irion (moderator), Zuzanna Warso, Yordanka Ivanova, @frimelle.bsky.social, Arriën Molema
More information: cpdp.be/19190
#CPDP.ai #CPDP2025 #TheWorldisWatching #CPDP #GPAI
May 5, 2025 at 11:28 AM
When Is a Summary of Data Used to Train GPAI Sufficiently Detailed?
Organised by @ivir-uva.bsky.social with Kristina Irion (moderator), Zuzanna Warso, Yordanka Ivanova, @frimelle.bsky.social, Arriën Molema
More information: cpdp.be/19190
#CPDP.ai #CPDP2025 #TheWorldisWatching #CPDP #GPAI
Organised by @ivir-uva.bsky.social with Kristina Irion (moderator), Zuzanna Warso, Yordanka Ivanova, @frimelle.bsky.social, Arriën Molema
More information: cpdp.be/19190
#CPDP.ai #CPDP2025 #TheWorldisWatching #CPDP #GPAI
Reposted by Lucie-Aimée Kaffee
In "Investigating Human Values in Online Communities", we perform a high-scale study of the unique values expressed by online communities with different perspectives
arxiv.org/abs/2402.14177 #NAACL2025 #NLProc
@nadavb.bsky.social @rnv.bsky.social @frimelle.bsky.social @iaugenstein.bsky.social
arxiv.org/abs/2402.14177 #NAACL2025 #NLProc
@nadavb.bsky.social @rnv.bsky.social @frimelle.bsky.social @iaugenstein.bsky.social
April 28, 2025 at 3:56 PM
In "Investigating Human Values in Online Communities", we perform a high-scale study of the unique values expressed by online communities with different perspectives
arxiv.org/abs/2402.14177 #NAACL2025 #NLProc
@nadavb.bsky.social @rnv.bsky.social @frimelle.bsky.social @iaugenstein.bsky.social
arxiv.org/abs/2402.14177 #NAACL2025 #NLProc
@nadavb.bsky.social @rnv.bsky.social @frimelle.bsky.social @iaugenstein.bsky.social
Reposted by Lucie-Aimée Kaffee
C3NLP Workshop #NAACL2025:
@iaugenstein.bsky.social will be presenting on tailoring LLM outputs to cultures, including where implicit cultural personalisation based on names leads to over-simplification arxiv.org/abs/2502.11995
@rnv.bsky.social @frimelle.bsky.social
bsky.app/profile/frim... #NLProc
@iaugenstein.bsky.social will be presenting on tailoring LLM outputs to cultures, including where implicit cultural personalisation based on names leads to over-simplification arxiv.org/abs/2502.11995
@rnv.bsky.social @frimelle.bsky.social
bsky.app/profile/frim... #NLProc
April 28, 2025 at 3:56 PM
C3NLP Workshop #NAACL2025:
@iaugenstein.bsky.social will be presenting on tailoring LLM outputs to cultures, including where implicit cultural personalisation based on names leads to over-simplification arxiv.org/abs/2502.11995
@rnv.bsky.social @frimelle.bsky.social
bsky.app/profile/frim... #NLProc
@iaugenstein.bsky.social will be presenting on tailoring LLM outputs to cultures, including where implicit cultural personalisation based on names leads to over-simplification arxiv.org/abs/2502.11995
@rnv.bsky.social @frimelle.bsky.social
bsky.app/profile/frim... #NLProc
I’m in Singapore this week for #ICLR2025 - co-organising the workshop on GenAI Watermaking sites.google.com/view/genai-w... and giving a keynote on policy & dataset transparency at the MLDPR workshop mldpr2025.com/schedule/
April 23, 2025 at 12:28 AM
I’m in Singapore this week for #ICLR2025 - co-organising the workshop on GenAI Watermaking sites.google.com/view/genai-w... and giving a keynote on policy & dataset transparency at the MLDPR workshop mldpr2025.com/schedule/
🚨 Our comments on the 3rd draft of the EU Code of Practice for GPAI developers are out! 🚨
Some promising updates—but also concerning choices that could harm transparency & open AI development. A quick breakdown of the blogpost with @yjernite.bsky.social 🧵👇
Some promising updates—but also concerning choices that could harm transparency & open AI development. A quick breakdown of the blogpost with @yjernite.bsky.social 🧵👇
March 13, 2025 at 10:50 AM
🚨 Our comments on the 3rd draft of the EU Code of Practice for GPAI developers are out! 🚨
Some promising updates—but also concerning choices that could harm transparency & open AI development. A quick breakdown of the blogpost with @yjernite.bsky.social 🧵👇
Some promising updates—but also concerning choices that could harm transparency & open AI development. A quick breakdown of the blogpost with @yjernite.bsky.social 🧵👇
Reposted by Lucie-Aimée Kaffee
Amazing work from HF colleagues that's super relevant to AI policy - open development matches "frontier" performance 🤗
For policymakers, it means we need regulation that works for both open and closed model - not closed-first that disproportionately harms open!
Good news is...
1/2
For policymakers, it means we need regulation that works for both open and closed model - not closed-first that disproportionately harms open!
Good news is...
1/2
March 12, 2025 at 4:26 PM
Amazing work from HF colleagues that's super relevant to AI policy - open development matches "frontier" performance 🤗
For policymakers, it means we need regulation that works for both open and closed model - not closed-first that disproportionately harms open!
Good news is...
1/2
For policymakers, it means we need regulation that works for both open and closed model - not closed-first that disproportionately harms open!
Good news is...
1/2