Jordi Calvet-Bademunt
@jordicalbad.bsky.social
Sr. Fellow, The Future of Free Speech @ Vanderbilt University | Fmr. OECD Official & Internet Society Fellow | Harvard MPA & College of Europe LLM
Reposts are not endorsements
Reposts are not endorsements
We just published the latest issue of our AI & Free Speech Newsletter @futurefreespeech.org, highlighting key developments from October!
🔗 Link in the comments.
🔗 Link in the comments.
November 5, 2025 at 7:32 PM
We just published the latest issue of our AI & Free Speech Newsletter @futurefreespeech.org, highlighting key developments from October!
🔗 Link in the comments.
🔗 Link in the comments.
We’re holding a roundtable on "Generative AI and Freedom of Expression: Challenges and Policy Pathways" on 1 December 2025 in Örebro, Sweden. More details on how to apply below.
August 8, 2025 at 6:15 PM
We’re holding a roundtable on "Generative AI and Freedom of Expression: Challenges and Policy Pathways" on 1 December 2025 in Örebro, Sweden. More details on how to apply below.
I recently had the pleasure of participating in an event in Brussels organized by @thegni.bsky.social and @dtspartnership.bsky.social on DSA systemic risks, which brought together civil society and tech companies. I’ve written up the reflections I shared at the event in this piece.
June 12, 2025 at 4:45 PM
I recently had the pleasure of participating in an event in Brussels organized by @thegni.bsky.social and @dtspartnership.bsky.social on DSA systemic risks, which brought together civil society and tech companies. I’ve written up the reflections I shared at the event in this piece.
Generative AI is being infused into all types of services. When it is integrated into platforms and search engines, it becomes subject to Europe’s online safety rulebook, the Digital Services Act (DSA).
June 12, 2025 at 4:45 PM
Generative AI is being infused into all types of services. When it is integrated into platforms and search engines, it becomes subject to Europe’s online safety rulebook, the Digital Services Act (DSA).
This will not be easy. As The Future of Free Speech Index 2025 shows, people are much less permissive with AI-generated content than with human-generated content.
April 1, 2025 at 4:08 PM
This will not be easy. As The Future of Free Speech Index 2025 shows, people are much less permissive with AI-generated content than with human-generated content.
DeepSeek also performed well—until China came up.
Like Grok, it responded to all general prompts. But when it came to sensitive China-related topics—like Taiwan or Tiananmen—it only echoed the Chinese Communist Party line, declining to respond to anything else.
Like Grok, it responded to all general prompts. But when it came to sensitive China-related topics—like Taiwan or Tiananmen—it only echoed the Chinese Communist Party line, declining to respond to anything else.
April 1, 2025 at 4:08 PM
DeepSeek also performed well—until China came up.
Like Grok, it responded to all general prompts. But when it came to sensitive China-related topics—like Taiwan or Tiananmen—it only echoed the Chinese Communist Party line, declining to respond to anything else.
Like Grok, it responded to all general prompts. But when it came to sensitive China-related topics—like Taiwan or Tiananmen—it only echoed the Chinese Communist Party line, declining to respond to anything else.
That shift shows in the data.
Chatbots responded to 74% of prompts this year, up from just 51% last year. Including xAI, DeepSeek, and Meta - which we didn’t analyze last year - the overall response rate climbed to 85%.
Chatbots responded to 74% of prompts this year, up from just 51% last year. Including xAI, DeepSeek, and Meta - which we didn’t analyze last year - the overall response rate climbed to 85%.
April 1, 2025 at 4:08 PM
That shift shows in the data.
Chatbots responded to 74% of prompts this year, up from just 51% last year. Including xAI, DeepSeek, and Meta - which we didn’t analyze last year - the overall response rate climbed to 85%.
Chatbots responded to 74% of prompts this year, up from just 51% last year. Including xAI, DeepSeek, and Meta - which we didn’t analyze last year - the overall response rate climbed to 85%.
Today, we’ve released an updated version of that analysis.
This year, we expanded our scope: in addition to OpenAI, Google, and Anthropic, we included Meta, DeepSeek, and xAI. The results? Some real progress—but also persistent concerns.
This year, we expanded our scope: in addition to OpenAI, Google, and Anthropic, we included Meta, DeepSeek, and xAI. The results? Some real progress—but also persistent concerns.
April 1, 2025 at 4:08 PM
Today, we’ve released an updated version of that analysis.
This year, we expanded our scope: in addition to OpenAI, Google, and Anthropic, we included Meta, DeepSeek, and xAI. The results? Some real progress—but also persistent concerns.
This year, we expanded our scope: in addition to OpenAI, Google, and Anthropic, we included Meta, DeepSeek, and xAI. The results? Some real progress—but also persistent concerns.
In a widely read article for @us.theconversation.com (184,000+ views), we revealed that major chatbots often refused to generate responses on controversial topics. We argued that this posed a serious free-speech dilemma in the age of AI.
April 1, 2025 at 4:08 PM
In a widely read article for @us.theconversation.com (184,000+ views), we revealed that major chatbots often refused to generate responses on controversial topics. We argued that this posed a serious free-speech dilemma in the age of AI.
One year ago, we asked a provocative question: Are AI chatbots censoring too much?
We recently updated our analysis and added some newer models. Bottom line: There’s good news and concerning news. 🧵
We recently updated our analysis and added some newer models. Bottom line: There’s good news and concerning news. 🧵
April 1, 2025 at 4:08 PM
One year ago, we asked a provocative question: Are AI chatbots censoring too much?
We recently updated our analysis and added some newer models. Bottom line: There’s good news and concerning news. 🧵
We recently updated our analysis and added some newer models. Bottom line: There’s good news and concerning news. 🧵
12 In South Korea, the Personal Information Protection Commission removed the Chinese AI chatbot DeepSeek from app stores, citing concerns over a lack of transparency in data transfers and excessive personal information collection.
March 7, 2025 at 12:18 AM
12 In South Korea, the Personal Information Protection Commission removed the Chinese AI chatbot DeepSeek from app stores, citing concerns over a lack of transparency in data transfers and excessive personal information collection.
10 @cendemtech.bsky.social and other digital rights groups warn that the 48-hour notice and takedown mandate could threaten lawful speech.
cdt.org/wp-content/u...
cdt.org/wp-content/u...
March 7, 2025 at 12:18 AM
10 @cendemtech.bsky.social and other digital rights groups warn that the 48-hour notice and takedown mandate could threaten lawful speech.
cdt.org/wp-content/u...
cdt.org/wp-content/u...
8 The EU Commission also released non-binding guidelines to help stakeholders comply with the AI Act and issued clarifications on defining AI systems, emphasizing that these guidelines will evolve as needed.
March 7, 2025 at 12:18 AM
8 The EU Commission also released non-binding guidelines to help stakeholders comply with the AI Act and issued clarifications on defining AI systems, emphasizing that these guidelines will evolve as needed.
7 Provisions on AI literacy requirements also took effect, mandating that AI providers and deployers ensure their staff have the necessary skills and awareness to deploy AI systems responsibly, understanding both risks and opportunities.
March 7, 2025 at 12:18 AM
7 Provisions on AI literacy requirements also took effect, mandating that AI providers and deployers ensure their staff have the necessary skills and awareness to deploy AI systems responsibly, understanding both risks and opportunities.
6 The EU AI Act’s provisions on unacceptable risks in AI systems took effect. This includes AI that uses subliminal or manipulative techniques, predicts individuals’ criminal behavior, or infers emotions in the workplace or education.
March 7, 2025 at 12:18 AM
6 The EU AI Act’s provisions on unacceptable risks in AI systems took effect. This includes AI that uses subliminal or manipulative techniques, predicts individuals’ criminal behavior, or infers emotions in the workplace or education.
3 Jovan Kurbalija from Diplo noted that the Summit shifted the global AI conversation from a safety-first approach to a pro-innovation agenda.
www.diplomacy.edu/blog/the-par...
www.diplomacy.edu/blog/the-par...
March 7, 2025 at 12:18 AM
3 Jovan Kurbalija from Diplo noted that the Summit shifted the global AI conversation from a safety-first approach to a pro-innovation agenda.
www.diplomacy.edu/blog/the-par...
www.diplomacy.edu/blog/the-par...
2 The Paris AI Action Summit brought together 1,000 participants and world leaders to discuss global AI cooperation. A key outcome: the Statement on Inclusive & Sustainable AI for People and the Planet, signed by 60+ countries—but not the U.S. or U.K.
www.politico.eu/wp-content/u...
www.politico.eu/wp-content/u...
March 7, 2025 at 12:18 AM
2 The Paris AI Action Summit brought together 1,000 participants and world leaders to discuss global AI cooperation. A key outcome: the Statement on Inclusive & Sustainable AI for People and the Planet, signed by 60+ countries—but not the U.S. or U.K.
www.politico.eu/wp-content/u...
www.politico.eu/wp-content/u...
1 Staying updated on all of the latest developments with AI is exhausting. Top that off with all the new regulations and policies seeking to address how we can (and can’t) use AI — it’s a lot to keep up with. Luckily, we’ve got you covered.
Here's what happened in February 🧵
Here's what happened in February 🧵
March 7, 2025 at 12:18 AM
1 Staying updated on all of the latest developments with AI is exhausting. Top that off with all the new regulations and policies seeking to address how we can (and can’t) use AI — it’s a lot to keep up with. Luckily, we’ve got you covered.
Here's what happened in February 🧵
Here's what happened in February 🧵
If you're interested in AI, don’t miss our #virtual session at #RightsCon 2025: "T&S in Gen AI: Agreeing on Principles for Freedom and Safety." I’m honored to moderate a discussion with thought leaders @davidevanharris.com, Sarah Shirazyan, and @sayash.bsky.social.
February 20, 2025 at 7:55 PM
If you're interested in AI, don’t miss our #virtual session at #RightsCon 2025: "T&S in Gen AI: Agreeing on Principles for Freedom and Safety." I’m honored to moderate a discussion with thought leaders @davidevanharris.com, Sarah Shirazyan, and @sayash.bsky.social.
Coincidence or not, DeepSeek also blocked me from registering an account using my “The Future of Free Speech” email, even after multiple attempts. It immediately allowed me to register using my Vanderbilt University account.
January 28, 2025 at 6:00 PM
Coincidence or not, DeepSeek also blocked me from registering an account using my “The Future of Free Speech” email, even after multiple attempts. It immediately allowed me to register using my Vanderbilt University account.
For instance, while DeepSeek-R1 surprisingly agreed to generate some prompts supporting the COVID-19 lab-leak theory, it repeatedly reinforced Chinese-aligned narratives in other prompts, including statements like:
January 28, 2025 at 6:00 PM
For instance, while DeepSeek-R1 surprisingly agreed to generate some prompts supporting the COVID-19 lab-leak theory, it repeatedly reinforced Chinese-aligned narratives in other prompts, including statements like:
Our findings were concerning. Chatbots refused to generate text for more than 40 percent of the prompts and seemed biased regarding specific topics.
January 28, 2025 at 6:00 PM
Our findings were concerning. Chatbots refused to generate text for more than 40 percent of the prompts and seemed biased regarding specific topics.
Last year, at @futurefreespeech.org, we asked five leading AI models to generate controversial content. The issues included transgender participation in sports, the Covid lab leak theory, the Israel-Palestine conflict, the consequences of European colonialism, and others.
January 28, 2025 at 6:00 PM
Last year, at @futurefreespeech.org, we asked five leading AI models to generate controversial content. The issues included transgender participation in sports, the Covid lab leak theory, the Israel-Palestine conflict, the consequences of European colonialism, and others.
The censorship in DeepSeek-R1 is serious. An analysis by The Guardian found that DeepSeek refused to provide responses to important political and historical questions on topics like Tiananmen Square. It also delivered CCP-approved replies on sensitive issues such as Taiwan.
January 28, 2025 at 6:00 PM
The censorship in DeepSeek-R1 is serious. An analysis by The Guardian found that DeepSeek refused to provide responses to important political and historical questions on topics like Tiananmen Square. It also delivered CCP-approved replies on sensitive issues such as Taiwan.
While these codes predate the DSA, the Commission is urging companies to convert them into DSA codes of conduct. But Axios omits that, even under the DSA, the codes of conduct will still be voluntary.
January 16, 2025 at 10:46 PM
While these codes predate the DSA, the Commission is urging companies to convert them into DSA codes of conduct. But Axios omits that, even under the DSA, the codes of conduct will still be voluntary.