iiinova — where your digital self keeps evolving

We also find that platform governance is inconsistent. Civitai displays a 'real person likeness' notice on 86% of SFW deepfake bounties, but that figure drops to just 58% for NSFW bounties.

$Stacked horizontal bar chart showing the distribution of platform content marking for deepfake bounties classified as SFW versus NSFW. Each bar represents 100\% of bounties within that category, divided into ``Not marked'' (light purple) and ``Marked'' (dark purple) segments. For SFW deepfake bounties, 85.6\% were marked by the platform while 14.4\% remained unmarked. In contrast, NSFW deepfake bounties show a more balanced distribution, with 58.3\% marked and 41.7\% not marked. The substantially higher marking rate for SFW deepfakes (85.6\%) compared to NSFW deepfakes (58.3\%) suggests that the platform's content moderation systems more effectively detect deepfakes when they violate deepfake policies in non-adult content contexts. The near-even split for NSFW deepfakes indicates potential detection challenges when deepfake content overlaps with adult content categories, where the presence of NSFW elements may obscure or complicate deepfake identification. This differential marking effectiveness has significant implications for platform governance, suggesting that moderation strategies may need category-specific approaches to achieve consistent enforcement across different content types.$

100

Participation in requesting bounties is concentrated, with the top 20% of requesters driving ~50% of all bounties.

$Lorenz curves illustrating the distribution of bounty requests across supporters (requesters) for three content categories: SFW (blue line, Gini = 0.395), NSFW (red line, Gini = 0.380), and Deepfake (green line, Gini = 0.452). The x-axis represents the cumulative share of supporters sorted by request volume (from highest to lowest), while the y-axis shows the cumulative share of bounties created. The diagonal dashed gray line represents perfect equality, where each supporter would contribute an equal share of requests. All three curves bow above the equality line, indicating concentration of requesting activity among smaller groups of users. Deepfake requests exhibit the highest concentration, with approximately 10\% of supporters responsible for 25\% of all deepfake bounties, compared to 20\% for SFW and 18\% for NSFW at the same percentile. The higher Gini coefficient for deepfakes (0.452) compared to SFW (0.395) and NSFW (0.380) indicates that deepfake requesting is significantly more concentrated among a small group of intensive users, suggesting potentially systematic rather than casual requesting behavior. This concentration pattern may serve as a behavioral indicator for targeted harassment, non-consensual intimate imagery production, or commercial exploitation operations.$

110

Deepfakes are a non-negligible slice of the marketplace. We identify hundreds of bounties requesting human deepfakes of real people, including some that are explicitly NSFW. Targets are overwhelmingly women (90% for SFW; 96% for NSFW) and skew towards public figures (actors/actresses/influencers).

$Two-panel figure analyzing deepfake bounties on Civitai. Panel (a) shows gender distribution split into two horizontal bar charts. The top chart displays NSFW deepfake bounties with Female at 95.8\% and Male at 4.2\%. The bottom chart shows SFW deepfake bounties with Female at 89.6\%, Male at 10.1\%, and Non-binary at 0.3\%. Panel (b) presents a stacked horizontal bar chart showing the profession distribution across all deepfakes, with bars color-coded for NSFW (red) and SFW (blue) categories. Actor/actress leads at 39.3\% of deepfakes, followed by Influencer at 13.9\%, Adult worker at 13.0\%, Artist at 10.2\%, Miscellaneous at 8.4\%, Model at 7.4\%, Athlete at 5.3\%, Self/spouse at 1.2\%, Reporter at 0.9\%, and Producer at 0.3\%. The x-axis shows the number of bounties ranging from 0 to 140.$

100

We find this marketplace is dominated by "LoRAs" (low-rank adaptations): lightweight, reusable adapters that "steer" model outputs towards a specific "look" (e.g., a person). Additionally, demand for NSFW content has been growing and has made up a majority of requests since September 2024.

100

Civitai runs a monetized feature called Bounties: users post paid requests—via "buzz" which is worth real money—and others compete to fulfill them. We analyze all available bounty requests collected over a 14-month period following the platform's launch—especially NSFW content and deepfakes.

100

Matt DeVerna @matthewdeverna.com · 29/12/2025

🚨 New working paper 🚨 What happens when you add money + competition to generative AI creation? On Civitai—a prominent platform for gen-AI content/tools w. millions of users—you get a growing marketplace for NSFW requests and a nontrivial stream of deepfakes. Preprint: arxiv.org/abs/2601.09117 🧪

$Four-panel figure showing Civitai bounty statistics. Panel (a) shows a horizontal bar chart of bounty counts by type, with LoRA having the most bounties at 2,984, followed by Image at 1,475, Model at 138, and decreasing to Video at 35. Panel (b) displays a horizontal bar chart of bounty themes by proportion, with NSFW being the largest category at approximately 0.43 (comprising 2,286 bounties split between platform-labeled and GPT-4.1 classified NSFW shown with diagonal hatching), followed by Fictional Characters at 1,463 bounties. Panel (c) presents stacked horizontal bars showing NSFW versus SFW proportions for each bounty type, with Video being 69\% NSFW and Embed being 82\% SFW. Panel (d) shows a time series from November 2023 to January 2025 displaying the proportion of SFW (decreasing trend line) and NSFW (increasing trend line) bounties over time, with 95\% confidence intervals shown as shaded regions around each trend line and vertical dashed lines marking the start of 2024 and 2025.$ $Two-panel figure analyzing deepfake bounties on Civitai. Panel (a) shows gender distribution split into two horizontal bar charts. The top chart displays NSFW deepfake bounties with Female at 95.8\% and Male at 4.2\%. The bottom chart shows SFW deepfake bounties with Female at 89.6\%, Male at 10.1\%, and Non-binary at 0.3\%. Panel (b) presents a stacked horizontal bar chart showing the profession distribution across all deepfakes, with bars color-coded for NSFW (red) and SFW (blue) categories. Actor/actress leads at 39.3\% of deepfakes, followed by Influencer at 13.9\%, Adult worker at 13.0\%, Artist at 10.2\%, Miscellaneous at 8.4\%, Model at 7.4\%, Athlete at 5.3\%, Self/spouse at 1.2\%, Reporter at 0.9\%, and Producer at 0.3\%. The x-axis shows the number of bounties ranging from 0 to 140.$

2173

I love the Claude — and now Codex — API “gifts” but they are for the only time period of the years when I try to truly stop working and don’t have the flexibility to develop stuff… 🤦🏻‍♂️

Screenshot of an email from the OpenAI Codex team informing Codex users that they have doubled API limits.

000

Matt DeVerna @matthewdeverna.com · 10/12/2025

The incredible @yyahn.bsky.social is building something special at UVA. 🚨 If you're a fit for this job, I highly recommend applying! Find more details via the link below. www.linkedin.com/posts/yyahn_...

Job posting on LinkedIn for a Research Scientist position in YY Ahn's group at UVA.

042

Exploring GPT citation patterns... Cited sources were mostly fact-checking outlets, mainstream news, and government sites. They have high reliability scores (NewsGuard) and tend to align with the political left.

181

Reasoning didn’t help much. Web search improved GPT models, but Gemini saw no benefit—likely because it failed to return sources for most queries. GPT models often return citations and many are the PolitiFact article containing the fact check. Again, curated info helps a lot.

1121

"Standard" models—i.e., models that do not leverage reasoning or web-search abilities—perform poorly when predicting PolitiFact's fact-checking label, with macro F1 typically ranging from 0.1 to 0.3. However, when we give models curated fact-checking evidence, performance improves dramatically.

181

Matt DeVerna @matthewdeverna.com · 22/11/2025

🚨 New working paper 🚨 Can LLMs with reasoning + web search reliably fact-check political claims? We evaluated 15 models from OpenAI, Google, Meta, and DeepSeek on 6,000+ PolitiFact claims (2007–2024). Short answer: Not reliably—unless you give them curated evidence. arxiv.org/abs/2511.18749

1621463

Bunch of postdoc opportunities. www.uni-konstanz.de/zukunftskoll...

Screenshot of text that reads as follows:

14 Postdoctoral Fellowships (2 years)
Fellowship Period: between July 2026 – November 2028
Application Deadline: 9 January 2026, 11:00 AM (CET)

The fellowships are available to researchers from all over the world and from any discipline represented at one if the three universities with a minimum of one year and a maximum of seven years of post-doctoral experience, who cannot conduct or continue their work in the USA
appropriately because of actual political pressure. The aim is to provide time and space for their excellent research and open up international prospects for their future career. The programme
offers a two-year research position at one of the three universities, preferred by the applicant, and local disciplinary cooperation, as well as support, resources, networks and exchange within the respective Institute for Advanced Studies with its international and interdisciplinary community of fellows.

031

Matt DeVerna @matthewdeverna.com · 18/04/2025

@bgoldberg.bsky.social and a rockstar team of technologists in industry and academia share some thoughts on 🚨AI and the Future of the Public Square🚨 Deserves a bit more attention (imho). arxiv.org/abs/2412.09988

041

Matt DeVerna @matthewdeverna.com · 11/04/2025

While writing my dissertation, I was reminded of this recent preprint led by @ryanmoore.bsky.social. The parallels between deception research and misinformation research are quite incredible. osf.io/preprints/ps...

Screenshot of a title page. The title is: Extending Truth-Default Theory to Misinformation: Lessons for Misinformation Scholarship from Deception Research. The authors are Ryan C. Moore and Jeffrey T. Hancock from Stanford University.

Screenshot of an abstract that reads: This paper connects deception research, specifically truth-default theory (TDT), to misinformation research. Through TDT’s propositions and supporting empirics, we demonstrate how the central questions in misinformation research can benefit from deeper engagement with deception scholarship and in particular TDT. Findings in the misinformation literature that have been surprising or unexpected (e.g., relatively few people are exposed to and share misinformation online) are predicted and explained by TDT. Robust conclusions from TDT about deception in human communication (e.g., people are truth-biased) are challenged by findings from misinformation research, prompting deeper investigation. Deception detection research paradigms have important lessons for misinformation research about the design of misinformation-detection tasks that should be considered when interpreting results (e.g., true-false base rates, prompted vs. unprompted veracity judgments). Analytic approaches from TDT (e.g., analyzing accuracy for truthful and deceptive messages independently) provide a framework for misinformation research to better understand the detection of misinformation and more holistically evaluate the outcomes of misinformation interventions, and TDT’s lessons on how to improve deception detection accuracy also suggest strategies for interventions which improve the detection of misinformation without undermining trust in credible news and information.

100

Matt DeVerna @matthewdeverna.com · 25/03/2025

www.nytimes.com/2025/03/24/o...

091

Matt DeVerna @matthewdeverna.com · 31/01/2025

www.washingtonpost.com/politics/202...

A screenshot of a Washington Post fact-checking article.

Headline: "$50 million for condoms in Gaza? There’s no evidence for the White House claim."

Subheading: "The figure would rain 1.5 billion condoms on an area only double the size of the District of Columbia."

Date: Updated January 29, 2025

Byline: Analysis by Glenn Kessler

Image Description: A photograph of White House Press Secretary Karoline Leavitt speaking at a press briefing. She has blonde hair, is wearing a magenta blazer, and appears to be mid-speech with a serious expression. The background is blurred, suggesting a formal press setting. The photo is credited to Jabin Botsford of The Washington Post.

052

Matt DeVerna @matthewdeverna.com · 29/01/2025

💥 Don’t miss our next OSoMe Awesome Speaker: @dfreelon.bsky.social! 💥 He'll be discussing 🚨 Computational Research in the Post-API Age 🚨 — which seems more and more relevant as time goes by... 📅 Feb 6, 12 PM ET | 💻 Online 🔗 Register: iu.zoom.us/meeting/regi...

010

Matt DeVerna @matthewdeverna.com · 22/12/2024

Want to understand your Google Cloud costs? It will only take THREE AND A HALF HOURS... 🙄

080

Matt DeVerna @matthewdeverna.com · 04/12/2024

Join us virtually 🚨TODAY @ 12 Eastern🚨 for a talk by @mrjimmyblack.com. "Understanding the Prominence of Alternative Social Media Platforms" Register here: iu.zoom.us/meeting/regi... Learn about the series here: osome.iu.edu/events/speak...

041

Matt DeVerna @matthewdeverna.com · 25/11/2024

@jay.bsky.team any plans to allow folks to get notifications when selected accounts post? Used this a lot in the other place. Would be wonderful if there were options for “all” vs “popular” posts, which did not exist in the other place.

A cellphone screenshot showing the functionality on X that allows users to get notified when a specific account posts. In this case it is shown for the official CNN account.

070

Matt DeVerna @matthewdeverna.com · 22/11/2024

This paper is super cool and deserves more attention. Plurals: A System for Guiding LLMs Via Simulated Social Ensembles (arxiv.org/abs/2409.17213) Comes with a Python package to guide LLM agents. Even allows incorporating nationally representative ensembles. Package: github.com/josh-ashkina... 🧪

Screenshot demonstrating how to use the Plurals Python library to create a nationally representative ensemble of agents modeled by different large language models (LLMs). Includes installation instructions (pip install plurals), API key setup, and example Python code to ask agents a task question about improving bike commuting.

Diagram and code demonstrating how to create a directed acyclic graph (DAG) of agents for story development using the Plurals Python library. The diagram includes SettingAgent, CharacterAgent, PlotAgent, and Moderator nodes with directional arrows indicating dependencies. The example code shows how to define a story prompt about a mystery set in 1920s Paris, focusing on the theft of an artwork from the Louvre, and specifies agents with unique roles (plot, character, and setting) using GPT-4 to collaboratively generate story elements.

Python code example showing how to create and process a directed acyclic graph (DAG) for collaborative story development using the Plurals library. The code defines a creative writing moderator with GPT-4, tasked with synthesizing plot, character, and setting elements into a cohesive story. It specifies edges representing the interaction pattern between agents ('setting' to 'character', 'setting' to 'plot', and 'character' to 'plot') and creates a Graph object with agents, edges, a story prompt, and the moderator. The DAG is processed to generate a final story outline, which is printed at the end.

1192

Matt DeVerna @matthewdeverna.com · 12/11/2024

Make sure you join the Observatory on Social Media TOMORROW for Kristina Lerman's talk, "Collective Psychology of Social Media: Emotions, Conflict & Mental Health in the Digital Age" at 12pm ET Register: iu.zoom.us/meeting/regi... Abstract/Series info: osome.iu.edu/events/speak...

Flyer for the 'Awesome Speakers' virtual event hosted by the Observatory on Social Media (OSoMe) at Indiana University. The flyer features the event title, date (November 13, 2024), and time (12 PM - 1 PM ET). It describes the event as a platform for scholars to discuss social media manipulation and misinformation. The featured speaker is Kristina Lerman from the University of Southern California. Attendees are encouraged to register and view the full list of speakers at osome.iu.edu/events.

021

We use a traditional method as well as PDI to reconstruct the same cascade in various ways and explore how much the distribution of cascade properties are affected by different approaches, finding substantial differences.

140

For example, *half* of the content in a user's "For you" feed on X comes from accounts they do not follow! (blog.x.com/engineering/...) so it seems unlikely that these methods are still reliable.

170

If the reconstruction had no effect, the set of influential accounts in both networks would be identical. Instead, we find the overlap is only 10-35%! (Fig c and f below)

160

Through case studies on Twitter and Bluesky, we find that something as simple as identifying influential accounts is massively affected when using networks constructed from reconstructed versus naïve (platform-provided) cascades.

160

Traditional analyses often assume that reshared content directly stems from the original poster. But the reality is messier—users encounter posts in many different ways, so we suspected this simplification might distort our understanding of what takes place on these platforms.

180