Madeleine Daepp

12 May

🎉 I've joined @cdt.org as their Research Director. 🎉 I could not be more excited to be part of an organization that has championed people's online rights, fought for good technologies, and defended against problematic ones for decades. If you have research ideas, reach out!

argmin.netThe World Is Something We Make...and we can make a truly open language model.

12h

"On the one hand, machine learning technology is more amazing than I ever imagined it could be. On the other hand, it’s a technology making everyone miserable." A beautiful and important essay on the path back to openness, and why it matters.

Melanie Mitchell@melaniemitchell.bsky.social

Read this excellent essay by @beenwrekt.bsky.social In short, forget OpenAI; we need to build open AI. www.argmin.net/p/the-world-...

Rest of World is shutting down in January unless it finds a new home -- another casualty of the stark drop-off in search and social click-throughs.

Ben Mullin@benmullin.bsky.social

News: Rest of World, the nonprofit newsroom focused on the impact of technology outside the West founded by Sophie Schmidt, tells donors in an email it will “cease publishing as a standalone entity” at the end of January.

View full thread

These days it is my go-to for the pulse on multilingual AI, putting out piece after piece on the most important problems from disparate token costs to the ways Western-centric safety frameworks are failing around the world.

It sounds like there's still a chance that RoW will get pulled in by a bigger newsroom. I hope it does. It's a wonderful and important publication, and there will be a gaping hole in global technology coverage in its absence.

Rest of World is shutting down in January unless it finds a new home -- another casualty of the stark drop-off in search and social click-throughs.

Ben Mullin@benmullin.bsky.social

When I set out to study AI's use and misuse in elections way back in 2023, Global Majority users were months and sometimes years ahead in the ways they were adopting AI -- both adversarial and creative. Rest of World was where I learned about it first.

arxiv.orgMajor Revision, Minor Accountability: LLMs as Reviewers of Published Articles in Top-Tier JournalsWith the growth of scientific output, the peer review system is approaching a breaking point: reviewers are increasingly hard to secure, and turnaround times lengthen accordingly. LLM-assisted or LLM-...

LLMs still have a ways to go before they can be effectively used for peer review. An excellent new paper looks at 200 papers across the very top scientific journals and shows that an AI reviewer would have rejected 75%. Not a single one was unanimously accepted. arxiv.org/abs/2610.31414

Madeleine Daepp@mdaepp.bsky.social

When it comes to identifying papers that move a field forward, understanding that a limitations section is not a directive of what to attack, or simply seeing what is good in an otherwise messy paper, I've yet to see any of the models do an even remotely decent job. 3/4

gking.harvard.eduAI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices | Gary KingWe show that AI-generated article reviews are now better, by every measure we could construct, than reviews written by well-compensated human experts and judged to be of unusually high quality by jour...

As best I can tell, the King and Worthington is not, actually, about peer review. If I were their reviewer, I'd make them change that title. 1/4

Jan Kabatek 🇺🇦@jankabatek.com

Gary King just dropped a study showing that AI peer reviews now dominate human peer reviews on every metric they tested. gking.harvard.edu/peerless/

@manoelhortaribeiro.bsky.social thinks the problem is taste, and I tend to agree. doomscrollingbabel.manoel.xyz/p/ai-comes-f... 4/4

doomscrollingbabel.manoel.xyzAI Comes for Peer ReviewingA new paper by Gary King and Steven Worthington argues that, compared to very nicely compensated human reviews, “AI-generated article reviews are now better, by every measure (they) could construct.” ...

As best I can tell, the King and Worthington is not, actually, about peer review. If I were their reviewer, I'd make them change that title. 1/4

Jan Kabatek 🇺🇦@jankabatek.com

Gary King just dropped a study showing that AI peer reviews now dominate human peer reviews on every metric they tested. gking.harvard.edu/peerless/

Center for Democracy & Technology

The system they look at is a pre-review system for soundness and rigor, which is indeed how we are all already using AI all the time. There's a general consensus that stat rigor checks are incredibly useful! But nobody likes getting an AI-generated peer review. Why? 2/4

Madeleine Daepp reposted

@cdt.org

Today: the Global Encryption Day Summit. Strong encryption protects the privacy and security of everyone online, and it’s under pressure worldwide. Join us. #GED2026 www.globalencryption...

Calling all Seattle civic tech nerds: The great @zephoria.bsky.social is coming to @townhallseattle.bsky.social October 19th to talk Census Data! If you enjoy wonky behind-the-scenes investigations into the administrative data that underpin our democracy, this is not to be missed 👀

danah boyd@zephoria.bsky.social

I'm also embarking on a crazy book tour. Come out to see me! More info can be found here: www.danah.org/data-are-mad...

techcrunch.comEveryone can find a reason to dislike data center construction | TechCrunchInside two years of fraught AI data center debates in Pennsylvania.

22 Sept

As a proud product of Central PA, I've seen data center fears build intensely and authentically in my hometown over the last year. Now, a wonderful new report from the brilliant folks at @datasociety.bsky.social documents these concerns in social and historical context. Important work!

alice e marwick@alicetiara.bsky.social

I am SO DELIGHTED to announce that our brand new report: The AI Factory: Data Centers, Power, and Resistance in Late Industrial Pennsylvania is out today. datasociety.net/research-lib... TechCrunch has the scoop: techcrunch.com/2026/09/22/e...

foso26.cdt.orgFuture of Speech Online

21 Sept

What are the implications of generative AI for the erosion of anonymity online? I'll be moderating a panel with @jaystanley.bsky.social, Tianshi Li, and Neil Chilson at the Future of Speech Online Conference. Join us on Sept 23-24 at foso26.cdt.org, hosted by @cdt.org and Stand Together Trust

cdt.orgCareers - Center for Democracy and TechnologyDo the possibilities presented by an innovative, open, and free internet inspire you? Do you find yourself engaged in passionate discussions about privacy policies? Are you yearning to put your expert...

17 Sept

Apply for the CDT Non-Resident Fellows Program! Deadline Extended through October 4th

Madeleine Daepp@mdaepp.bsky.social

Are you a researcher who wants to have policy impact? Apply to the @cdt.org non-resident fellows program! Fellows join an interdisciplinary cohort of leading scholars driving work on digital civil rights and liberties. Apply here by September 17: cdt.org/careers/#op-...

You can be concerned about deepfakes and still want a digital world with Balenciaga pope. Great analysis from @beccabranum.bsky.social

This post is unavailable.

theonion.comI Work Very Hard, And I Would Like To Try CakeHello. I am a horse. I work very hard at my job of being a horse. When humans say move the heavy thing, I move the heavy thing. When humans sit on top of me and pull on my head, I carry them where the...

16 June

Every now and then I read a thing that could only have been written by a human. This is one of those things: theonion.com/i-work-very-... AI could never.

anthropic.comClaude CorpsClaude Corps connects early-career talent with mission-driven nonprofits to put AI to work. Apply to be a fellow or host.

Anthropic is embedding AI fellows at nonprofits around the United States. Think Code for America, Peace Corps, or Americorps, but make it AI. It's called Claude Corps.

Anthropic {bot}@anthropicai.xmirror.bot

We’re launching Claude Corps, a national fellowship program matching people early in their careers with US nonprofits. We'll teach 1,000 people to use Claude, and pay them to use AI to advance their hosts’ missions. https://www.anthropic.com/claude-corps

But I study AI use and misuse around the world. I have seen firsthand that internet miscreants are adopting these tools far faster than democracy's defenders. I also mentor a lot of CS students and the first rung of the career ladder is just...gone. The model companies have a responsibility here!

And as somebody who is in civil society thanks to programs like this one, I love a funded initiative that gives more people who want to work on social causes the resourcing to do so. I hope to see similar programs from other companies in the space!

anthropic.comClaude CorpsClaude Corps connects early-career talent with mission-driven nonprofits to put AI to work. Apply to be a fellow or host.

Anthropic is embedding AI fellows at nonprofits around the United States. Think Code for America, Peace Corps, or Americorps, but make it AI. It's called Claude Corps.

Anthropic {bot}@anthropicai.xmirror.bot

@isabelalinzer.bsky.social

Civil society resistance to AI is not just about capacity. It has anchors in deep and serious concerns around data centers, intellectual property, the lack of regulation, the displacement issues, and more.

Madeleine Daepp reposted

Isabel Linzer

10 June

ICYMI, the proposed USPS rules that follow the March mail voting EO are live for comment until July 2. A couple thoughts🧵 www.cnn.com/2026/06/10/p...

cnn.comPostal Service won’t deliver mail ballots for states that don’t hand over voter lists, under plan for Trump directive | CNN PoliticsState election officials could soon face a stark choice: hand over voter lists to the Trump administration or risk losing Postal Service delivery for mail-in ballots.

3 June

When people do work with generative AI, they frequently do knowledge work. But will broad adoption of AI *democratize* knowledge work? In a new perspective in @natcomputsci.nature.com, we lay out technical + social factors shaping access to and use of AI for knowledge work around the world.

As always, these challenges require social as well as technical interventions. What, exactly, the social interventions should be changes, though, depending on the extent to which skill-biased or equalizing effects prevail.

nature.comAI and the democratization of knowledge work - Nature Computational ScienceThis Perspective explores the extent to which AI democratizes knowledge work and discusses potential intervention strategies for addressing social and place-based divides.

3 June

Technologists often say something is "democratized" when they simply mean it has scaled. For knowledge work to be genuinely democratized with AI, it will need to be broadly accessible---and that access needs to produce real benefits and minimize harms. Full perspective: nature.com/articles/s43...

1. The distribution of AI's benefits depends on the task. - Equalizing tasks are standardized, common in training data, and have objective success criteria. - Skill-biased tasks are open-ended, creative, and where success hinges on user discretion (or taste!)

3 June

2. Access and usefulness is also constrained in low-resource languages and where knowledge work is a small portion of the economy. In the absence of intervention, we risk an AI divide not just in who accesses and uses AI for knowledge work, but also in who benefits versus who is harmed.

404media.coNew Study Reveals the Manipulative ‘Dark Patterns’ of AI ChatbotsA new study by the Center for Democracy & Technology shows how chatbots like ChatGPT, Gemini, Replika and more can lead users down paths they didn't intend.

29 May

When does the design of conversational AI systems push users into acting in ways that are misaligned with their interests? Great @404media.co article about the new @cdt.org report on "dark patterns" by Ruchika Joshi, Adinawa Adjagodjou, and @mluria.bsky.social www.404media.co/new-study-re...

7 Apr

"conversational agents can covertly redirect consumer choices at scale, most users cannot tell when it is happening, and existing transparency mechanisms are insufficient" New + excellent paper on effects of persuasive AI in shaping purchasing behaviors. Effect sizes are concerningly large!

Manoel Horta Ribeiro@manoelhortaribeiro.bsky.social

LLMs will transform e-commerce in ways that consumer protection is unprepared for! Our new preprint finds that conversational AI can strongly steer consumer choices: sponsored product selection nearly tripled relative to traditional placement (N=2012). Link: arxiv.org/abs/2604.04263

31 Mar

Hello if you think this expression on my face is mischief and provocation you would be correct. Having an absolute blast at #ECIR2026 #IR4Good - thank you for letting me show up to to talk about how important (good) IR is for global democracies!

Bhaskar Mitra | ভাস্কর মিত্র@bmitra.bsky.social

...an engaging panel with critical conversations between Madeleine Daepp, @djoerd.idf.social.ap.brid.gy, @sannevrijenhoek.bsky.social, and myself, moderated by my wonderful co-track chair @mariaheuss.bsky.social. #ECIR2026 #IR4Good (Photo credit: @arjenpdevries.bsky.social) /4

30 Mar

"I don't want to not use cool technology just because it has a shitty business model" - Cindy Cohn at @townhallseattle.bsky.social 80% through the book and she is an inspiration and a delight

Electronic Frontier Foundation@eff.org

📺 HAPPENING TONIGHT 📺 EFF's Cindy Cohn will be on @TheDailyShow.com with Jon Stewart at 10 PM Central and 11 PM Eastern and Pacific! Make sure you tune in!

news.mit.eduStudy: Platforms that rank the latest LLMs can be unreliableThe results of popular LLM ranking platforms can be skewed by just a few data points, possibly providing an unreliable report about which LLM would perform best in real situations, according MIT resea...

23 Mar

People make expensive decisions based on how LLMs perform on Chatbot Arena and other ranking systems. Turns out these rankings are very sensitive to tiny changes. New paper from MIT: news.mit.edu/2026/study-p...

theverge.comGoogle Search is now using AI to replace headlinesLet us know if you see more.

22 Mar

I am working on my ECIR keynote on "gemini hegemony" and wow this is going to be an excellent slide

Casey Newton@caseynewton.bsky.social

This really sucks

adventures-in-ai-and-democracy.ghost.ioAI Interviewers, Qualitative Literacy, and “Big Dick” DataThe moment I first tried ChatGPT, I knew it would change my research. So, in the summer of 2023, a wonderful team of PhD interns and I decided to test whether it could do interviews. We ended up writi...

20 Mar

In 2023, my team published one of the first papers on AI-led interviews. We wanted to know if ChatGPT could do qualitative research. (It couldn't) Now Anthropic has done 80k AI interviews across 159 countries. Some thoughts on why adaptive AI surveys are useful but do *not* replace qual research.

werkenbij.uva.nlVacancy — Postdoctoral Researcher in Open AI research for social sciences and humanities<p>The Humane AI research priority area of the University of Amsterdam is looking for a new PostDoc to amend our existing team (<a href="https://humane-ai.nl/">https://humane-ai.nl/</a>). AI is everywhere. Chatbots, translation tools, shopping assistants, hyper-personalised services - as Large Language Models become widely accessible, they're reshaping how we live, work, play and connect. This rapid shift sparks big questions: Whose norms and values shape this AI-driven society—and who gets left behind? What are the rules that govern the development and deployment of Generative AI different sectors of society, and towards what kind of society will they steer us? Is the risked-based approach to AI. </p>

18 Mar

This thread on a Brussels AI + Foreign Interference panel is one of the best run-downs I've seen of the challenges ahead this year including: - GEO, carpet bombing, and training data grooming - model sovereignty vs. quality / access / cost - the cognitive costs of efficiency over accuracy bad bad!

Aaron Rodericks@aaron.bsky.team

-> you've seen a massive spike in GEO companies that are the equivalent of SEO for LLMs. They spin up 100k web sites, and then will saturate Reddit and Wikipedia to saturate a specific message that will be used to train the bots to deliver that message.

Madeleine Daepp reposted

Petter Törnberg

@pettertornberg.com

16 Mar

Come be my colleague! Postdoc position at the University of Amsterdam: Open-source AI for Social Sciences & Humanities. Work on open LLM infrastructure, reproducible research tools, and ethical AI for SSH. 🗓 Tight deadline: Apply by April 3 werkenbij.uva.nl/en/vacancies...

13 Mar

Last night the Iranian Embassy in The Hague dropped a Pixar-style AI video trolling President Trump. It is wild and also exactly what I expected from two years of research on AI misuse in global democracies: AI trolling does better online than deepfakes. We called the tactic a "deep roast."

arxiv.orgGenerative PropagandaGenerative propaganda is the use of generative artificial intelligence (AI) to shape public opinion. To characterize its use in real-world settings, we conducted interviews with defenders (e.g., factc...

13 Mar

If you need words for what you're seeing online, here's our taxonomy: arxiv.org/abs/2509.19147

Madeleine Daepp reposted

Mozilla

@mozilla.org

28 Feb

🚨 Call now open: Democracy x AI Incubator Mozilla Foundation is funding working AI systems that strengthen information, transparency, and civic space. $50K + 12 months of support. Apply by March 16, 2026 https://bit.ly/406u4yj

bit.lyAPPLY NOW Democracy x AI Cohort | Mozilla FoundationWe are seeking to support working prototypes that prove AI can serve democracy. Apply now.

6 Mar

I am delighted to be keynoting the IR-for-good track at @ecir2026.eu I will be talking about what I learned from studying generative AI's use during the biggest election year in history. (Spoiler: the problem is not just deepfakes.) The talk is called gemini hegemony, expect a little 🌶️

Bhaskar Mitra | ভাস্কর মিত্র@bmitra.bsky.social

🗓️ Monday, March 30 10:30-12:30: IR-for-Good Paper Session I 13:30-14:30: IR-for-Good Paper Session II 16:00-17:15: IR-for-Good Invited Talks and Panel 🗓️ Tuesday, March 31 09:00-10:00: IR-for-good Keynote by @mdaepp.bsky.social on "Gemini Hegemony" 10:30-12:30: IR-for-Good Paper Session III Cont...

plus.flux.communityThe Hegsethian jihad against AnthropicDid OpenAI cast aside safety concerns about fully automated weapon systems?

Hegseth calling it "defective altruism" is the strangest and most telling moment in this whole saga. When you interview for Anthropic, the most important question is "why anthropic?" It's a culture screener: the answer is safety, alignment, being one of the good guys.

Matthew Sheffield@matthew.flux.community

Did OpenAI reject its stated core concerns about AI running fully autonomous robots and conducting mass surveillance of Americans? It sure seems that way. I parse the language and provide the background here:

And then there is the OpenAI response. Anthropic drew an explicit line against fully autonomous systems. Sam put a human in the loop. That's what any good scholar of sociotechnical systems would recognize as a "moral crumple zone"---someone to blame when autonomous systems go awry.

...So now the moral onus sits with the warfighter workforce instead of the AI one.

plus.flux.communityThe Hegsethian jihad against AnthropicDid OpenAI cast aside safety concerns about fully automated weapon systems?

Matthew Sheffield@matthew.flux.community

A really high fraction of Anthropic's workforce has safety-aligned values because they've been screening for it since day one. Whatever your perspective on EA ideology, this is a wildly interesting case of workforce values clashing with state power.

adventures-in-ai-and-democracy.ghost.ioHow not to set up an agent swarmMost humans have moved on from Moltbook, the first social media site exclusively for AIs, but plenty of bots are still checking in regularly. Including one particular bot, Marginal, that I tasked with...

3 Mar

If you're trying to build successful and pro-social agent swarms, here are four lessons from moltbook (...on what *not* to do)

26 Feb

Are other people's AIs really into Ursula K. Le Guin and Borges or is that just mine?