Sign in

Madeleine Daepp

@mdaepp.bsky.social
128 followers 126 following 111 posts

Civic technologist studying AI + democracy. Research Director @cdt.org, currently on civic leave from Microsoft Research. PhD MIT. Concerned about generative propaganda but have handed my entire life over to Claude. Views my own. madeleinedaepp.com

PostsRepliesMedia
Madeleine Daepp @mdaepp.bsky.social · 04/10/2026
This talk was amazing! Inspiring to hear how you brought an research-based moderation taxonomy to the community, had democratic systems so ppl could vote on what to implement, and now run it irl. I felt like I had stepped into an alternate timeline where social media could be liberatory again.
000
Madeleine Daepp @mdaepp.bsky.social · 01/10/2026
Sometimes I wonder why we invested millions of dollars to create self-driving cars when we could just have good public transit...and then i visit Palo Alto
010
Madeleine Daepp @mdaepp.bsky.social · 24/09/2026
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 👀
011
Madeleine Daepp @mdaepp.bsky.social · 23/09/2026
i'm so stoked for you but also we've started an STS reading group at work so i will come knocking when we get to nancy baym week 👀
100
Madeleine Daepp @mdaepp.bsky.social · 23/09/2026
"At a time of austerity and belt-tightening, the MIT administration loves us enough to spend millions of dollars to watch over us as closely as if we were maximum security prison inmates, or bugs under microscopes." Amazing.
010
Madeleine Daepp @mdaepp.bsky.social · 23/09/2026
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!
062
Reposted by Madeleine Daepp
🦇🎃💀 Riana-mator 💀🎃🦇 @riana.bsky.social · 17/09/2026
The Non-Resident Fellowship at @cdt.org is open for applications! Deadline is Oct. 3. I'm one of this year's cohort and it's been a great opportunity to highlight the civil liberties underpinnings of my work. If you're an academic with a digital rights bent, please apply! cdt.org/careers/#op-...
cdt.org
Careers - Center for Democracy and Technology
Do 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...
067
Madeleine Daepp @mdaepp.bsky.social · 21/09/2026
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
foso26.cdt.org
Future of Speech Online
010
Madeleine Daepp @mdaepp.bsky.social · 17/09/2026
Apply for the CDT Non-Resident Fellows Program! Deadline Extended through October 4th
000
Madeleine Daepp @mdaepp.bsky.social · 15/09/2026
I do think there is something fundamentally different about the 1-1 nature of chatbot interactions versus the 1-many nature of socials. We expect humans in social content moderation, but folks have always been weird about e.g. having humans see their search histories.
000
Madeleine Daepp @mdaepp.bsky.social · 15/09/2026
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-...
cdt.org
Careers - Center for Democracy and Technology
Do 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...
063
Madeleine Daepp @mdaepp.bsky.social · 17/06/2026
You can be concerned about deepfakes and still want a digital world with Balenciaga pope. Great analysis from @beccabranum.bsky.social
010
Madeleine Daepp @mdaepp.bsky.social · 17/06/2026
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.
theonion.com
I Work Very Hard, And I Would Like To Try Cake
Hello. 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...
000
Madeleine Daepp @mdaepp.bsky.social · 12/06/2026
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!
000
Madeleine Daepp @mdaepp.bsky.social · 12/06/2026
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!
110
Madeleine Daepp @mdaepp.bsky.social · 12/06/2026
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.
100
Madeleine Daepp @mdaepp.bsky.social · 12/06/2026
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.
100
Madeleine Daepp @mdaepp.bsky.social · 12/06/2026
I'm seeing so many wonderful super-smart early career people who are struggling to find a first job, would be great if this helps them + gives civil society folks a little more org capacity.
100
Reposted by Madeleine Daepp
Isabel Linzer @isabelalinzer.bsky.social · 10/06/2026
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.com
Postal Service won’t deliver mail ballots for states that don’t hand over voter lists, under plan for Trump directive | CNN Politics
State 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.
265
Madeleine Daepp @mdaepp.bsky.social · 08/06/2026
I've always loved the Moving to Opportunities study. The quant eval was mostly null, but Briggs, Popkin and Goering led a qual evaluation that found big reported mental health benefits among other things. Decades later, Chetty brings in tax records + finds big fx for affected kids as adults.
010
Madeleine Daepp @mdaepp.bsky.social · 03/06/2026
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...
nature.com
AI and the democratization of knowledge work - Nature Computational Science
This Perspective explores the extent to which AI democratizes knowledge work and discusses potential intervention strategies for addressing social and place-based divides.
010
Madeleine Daepp @mdaepp.bsky.social · 03/06/2026
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.
100
Madeleine Daepp @mdaepp.bsky.social · 03/06/2026
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.
100
Madeleine Daepp @mdaepp.bsky.social · 03/06/2026
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!)
100
Madeleine Daepp @mdaepp.bsky.social · 03/06/2026
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.
111
Madeleine Daepp @mdaepp.bsky.social · 29/05/2026
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...
404media.co
New Study Reveals the Manipulative ‘Dark Patterns’ of AI Chatbots
A 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.
063
Madeleine Daepp @mdaepp.bsky.social · 25/05/2026
I think you may have been the one who recommended that paper to me :) Would be suuuuuuper interested to know whether + how this effect shows up in non-US politics!
010
Madeleine Daepp @mdaepp.bsky.social · 25/05/2026
It has been so easy to dismiss these stories and blame the users. It's sobering to revisit these with the agent meltdown framework in mind. Feels like a reminder to track and address near-misses before they become true misses, per the great STS text mitpress.mit.edu/978026254699...
mitpress.mit.edu
Rational Accidents
Nuclear reactors, deep-sea drilling platforms, deterrence infrastructures—these are all complex and formidable technologies with the potential to fail cata...
110
Madeleine Daepp @mdaepp.bsky.social · 25/05/2026
@jacyanthis.bsky.social @pettertornberg.com As somebody who is about to be doing much more audit work, the inferred-auditor framing is super useful. Pairs nicely with arxiv.org/abs/2503.02080 as a mechanism for how sycophancy might be steering ideology
arxiv.org
Linear Representations of Political Perspective Emerge in Large Language Models
Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly ab...
120
Madeleine Daepp @mdaepp.bsky.social · 14/05/2026
Amazing and depressing work! What a bummer update to the amazing and more hopeful authoritarian data problem work you were doing a few years ago
120
Madeleine Daepp @mdaepp.bsky.social · 13/05/2026
When I worked at @sfiscience.bsky.social, lunches and tea breaks were the best part of the job: One good chat with a smart person would save weeks of going down the wrong research rabbit hole 🐰 Knowledge work requires knowledge sharing, and you lose so much when productivity is the only metric.
010
Madeleine Daepp @mdaepp.bsky.social · 12/05/2026
🎉 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!
2112
Madeleine Daepp @mdaepp.bsky.social · 17/04/2026
Lance Armstrong on Revisionist History: "Cars freak me out. Being out on the road with cars and drivers and distracted drivers, it scares the shit out of me. But, you know, Malcolm, I actually now prefer the mountain bike just so I can be totally away from cars." Same.
020
Madeleine Daepp @mdaepp.bsky.social · 15/04/2026
Yes we should all be talking about this! But tbh i worry that making it about AI makes people think we need to fix *the AI* when really we just need to finally pass a proper privacy law + close the data broker loophole.
010
Madeleine Daepp @mdaepp.bsky.social · 15/04/2026
One of my mentors taught me that you have to advocate for your research, that if you have a finding people should know about you have to go out and tell them. But it's remarkable how important this is versus how little infra / training / reward there is for it at universities.
010
Madeleine Daepp @mdaepp.bsky.social · 15/04/2026
This is definitely a major prio for the labs. A key breakthrough in ChatGPT was post-training for user / assistant interactions; Opus 4.6 was agent / tool. Almost surely folks are pushing on RLVR for agent swarms, it's just not clear yet what good reward functions are for group interactions.
010
Madeleine Daepp @mdaepp.bsky.social · 09/04/2026
I worked for U of Minnesota the summer their state govt shut down (a decade ago now). People there hated all government *except* ag extension. Far more disciplines should have extension services. People like you more if you listen to + work with them! And it's satisfying + meaningful!
010
Madeleine Daepp @mdaepp.bsky.social · 07/04/2026
"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!
011
Madeleine Daepp @mdaepp.bsky.social · 31/03/2026
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!
022
Madeleine Daepp @mdaepp.bsky.social · 30/03/2026
the opposing lawyers who make “let me tell you about your case, little girl” calls -- and, in so doing, reveal their entire legal strategies! the cryptographers who got NSA letters to prevent them from publishing, and who wore their code on tshirts and as tattoos in protest! this book is a treat.
000
Madeleine Daepp @mdaepp.bsky.social · 30/03/2026
"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
030
Madeleine Daepp @mdaepp.bsky.social · 23/03/2026
This is great - A major pet peeve of mine is when researchers say they "prompted ChatGPT" instead of (a) specifying which model and (b) being clear that it was the api, not the web interface. These are v different things!
020
Madeleine Daepp @mdaepp.bsky.social · 23/03/2026
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...
news.mit.edu
Study: Platforms that rank the latest LLMs can be unreliable
The 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...
000
Madeleine Daepp @mdaepp.bsky.social · 22/03/2026
Wow i want to go to this waffle house and be part of this crew.
101
Madeleine Daepp @mdaepp.bsky.social · 22/03/2026
Really important to know if these headlines get changed *before* or *after* gemini consumes it for the summary
000
Madeleine Daepp @mdaepp.bsky.social · 22/03/2026
I am working on my ECIR keynote on "gemini hegemony" and wow this is going to be an excellent slide
000
Madeleine Daepp @mdaepp.bsky.social · 22/03/2026
Yes! When the science paper on how AI could durably reduce conspiracy theory beliefs came out, my first thought was: okay but that's not how any actual conspiracy theorists will use it.
000
Madeleine Daepp @mdaepp.bsky.social · 22/03/2026
So...how are you feeling about the new anthropic work in this space?
100
Madeleine Daepp @mdaepp.bsky.social · 20/03/2026
I attempted to build an AI interviewer three years ago (it did not work) and then did a v different study with in-person interviews in Taiwan and India and the disembodied and decontextualized critiques are *so* useful in explaining why the real-world work felt like such better research.
010
Madeleine Daepp @mdaepp.bsky.social · 20/03/2026
Hello I included your post in my blog post about this! adventures-in-ai-and-democracy.ghost.io/ai-interview... It is a cool study but also feels like one of those situations where we think we are talking about AI but actually we are all talking about capitalism.
adventures-in-ai-and-democracy.ghost.io
AI Interviewers, Qualitative Literacy, and “Big Dick” Data
In the summer of 2023, a wonderful team of PhD interns and I decided to test whether AI could do interviews. We ended up writing one of the first papers on the topic.  Now, Anthropic has scaled the a...
010