Who produces hate speech? And how does that matter for content moderation?
We show that across different countries and platforms, a relatively small share of users are responsible for a very large share of hate - overall, 5% write 83-100% of hateful content.
www.cambridge.org/core/journal...
The big opportunity I see here (and perhaps that's me being an optimist) is to imagine ways to scaffold LLM use in ways that preserve or increase epistemic vigilance.
If the early 2020s showed us something, it is that the bar is not that high 😅
But also, I wonder the extent to which the argument holds in an increasingly agentic world? LLMs these days can provide you with a full reasoning path backed by credible sources (and links!).
Does this change the game somehow? Can this enable reasonable epistemic vigilance while using LLMs?
Thought-provoking thread by @mjcrockett.bsky.social
Some additional thoughts: maybe the reason why math is where LLMs seem to be making most progress is that that's where verification is cheaper and more definitive (tuhs, sources matter less)? E.g., 17-year-olds proving conjectures...
🧵on my new paper "Synthetic personas distort the structure of human belief systems" w Roberto Cerina I'm v excited about...
🚨 Do synthetic samples look like human samples?
We compare 28 LLMs to the 2024 General Social Survey (GSS) to find out + develop host of diagnostics...
Afaiu they did not
This is such a banger
LLMs make statistical metascience easier. LLMs don't increase the validity of statistical metascience. www.argmin.net/p/devezers-urn
New paper! The Linear Representation Hypothesis is a powerful intuition for how language models work, but lacks formalization. We give a mathematical framework in which we can ask and answer a basic question: how many features can be stored under the hypothesis? 🧵 arxiv.org/abs/2602.11246
Congratz Maria!!!
Aren't there two possible meanings of "derogatory" here? If meant as belittling, then I'd agree with the OP. If meant as an insult, then I'd agree with you! It is just a rhetorical shorthand.
That said, the meaning may blur past its initial intention given Bender’s stance of LLMs as a dead-end...
In the post, I shortly lay out three arguments against this conclusion: 1) taste isn’t steered by decree; 2) every new medium looks like cheating at first; and 3) I'm pretty sure there will be some process when doing AI art.
And he ends with a tempting conclusion: since art is partly what we collectively decide to treat as art, the “solution” to AI art is… collective refusal. Just decide it’s not worth pursuing. I’m sympathetic. But I don’t fully buy it...
Sanderson starts questioning his own stance: Am I just doing the “new medium isn’t real art” thing? (Think Ebert on video games) Then he tries to argue that his discomfort is different. He argues AI art collapses “art” into product, but that the point of art is also the process.
I wrote about Brandon Sanderson’s take on AI art. It is one of the most thoughtful and intellectually humble pieces I've seen on the subject.
TL;DR: I agree with the spirit, but not the conclusion.
doomscrollingbabel.manoel.xyz/p/brandon-sa...
New research on Deepfake NCII: if I'm reading this correctly, legislation and site shutdowns are not enough to curb the behavior and demand for deepfaked pornography. It basically shifts those users to other spaces.
h/t @tristanl.ee
arxiv.org/abs/2602.02754
Thank you!!!
We argue that these findings have the potential to guide policy to address the proliferation of such content. If the goal is to reduce prevalence, we likely need enforcement and cross-platform coordination to avoid playing whack-a-mole.
Our main finding is that there is no sustained decline after interventions. On the contrary, there is a substantial growth in deep fake-related content that more than offsets all that was shared in Mr. Deepfakes. E.g., 4chan saw a 3k+ increase in deepfake requests per week!
Within each of these websites, we use a synthetic control approach to estimate how the sharing of deepfakes would have progressed in the absence of the compound shock. We use various subforums within each website as controls. (SNCEI is the term we use to refer to deepfakes).
To answer that, we tracked weekly activity across three other sites that host this type of material. We considered a variety of outcomes: new posts including deepfake content, newly active contributors, and even requests for deepfakes.
We ask whether this kind of shock led to the suppression of Deepfake content online. Did this result in fewer people sharing deepfakes? Or did this simply lead to a restructuring of the ecosystem?
On April 28, 2025, the U.S. House passed the TAKE IT DOWN Act. Within a week of that vote, MrDeepfakes (a major hub for synthetic non-consensual explicit imagery) announced it was shutting down. We treat these closely timed events as a compound shock to the deepfake ecosystem
Deepfake pornography isn’t going away just because we are passing laws and taking down a couple of big websites.
Our new pre-print, led by @aedcv.bsky.social suggests that the sharing of this material continued to prosper even after platform and policy shocks.
arxiv.org/abs/2602.02754
A search for factors for algorithm understanding results in multiple terms displayed as documents, including available, compact, and aligned. These are shown to be necessary and sufficient. Other, similar terms are shown in the background faded, like intuitive, rule-based, grounded, modular, linear, decomposable, accurate, symbolic, causal, and personalized.
Is the only way we can create algorithms that people understand to make them trivially simple? We argue, no.
People can predict the behavior of algorithms that are arbitrarily complex, if and only if they are available, compact and aligned.
arxiv.org/abs/2601.18966
CS ArXiv recently banned “review and position” papers, but what are those? Do they include more generated content? Who is most affected by this change? @yanai.bsky.social and I dug into the data to find out!
Nearly 50% of Computers & Society papers might be censored, vs 3% of Computer Vision ‼️
This was incompatible with ACM template, but what did work was:
\title{My Amazing Paper \texorpdfstring{🤖}{} with Emoji}
Folks who have papers with emojis / images in the title! How do you make the pdf metadata title pretty?
