Sign in

Abhisek Dash

@adash93.bsky.social
67 followers 28 following 34 posts
PostsRepliesMedia
Abhisek Dash @adash93.bsky.social · 29/09/2026
I’m happy to share that earlier this month I have started as a Research Group Leader at MPI-SWS! As part of this position, my group aims at bridging the AI Computing and Society Initiative (aics.mpi-sws.org) and SOUNDS - Societal Observatory Using Novel Data Sources initiative (lnkd.in/eekwusVu).
001
Abhisek Dash @adash93.bsky.social · 12/09/2026
@pushpdeep.bsky.social, @ingmarweber.de.
000
Abhisek Dash @adash93.bsky.social · 12/09/2026
Joint work with Ayan Majumdar, Shounak Paul, Pushpdeep Singh, Ines Abdelaziz, Sayeh Jarollahi, Seungeon Lee, Krishna P. Gummadi, Ingmar Weber.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
Our results suggest that they are a promising building block, rather than a replacement for human moderators. But important challenges remain (e.g., biases, scalability etc.) to be explored in deeper details.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
Another alternative is to provide the VLMs with prior moderation decisions as examples. The best example-driven approach achieves nearly identical F1 performance. This parity comes at a severe practical cost in latency and expenditure.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
Within the instruction-driven paradigm, granu- lar policy details are critical in boosting moderation consistency and effectiveness: roughly halving flag- ging rates and improving F1 scores by up to 25%, indicating the importance of better harness or policy for better operationalization.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
When we provide VLMs with detailed moderation policy, the instruction-driven foundation models outper- form the current Bluesky moderation system, achieving nearly triple its F1 score on Random Posts (0.60 vs. 0.22). Open-weight models deliver competitive performance to their frontier models.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
Key findings: Our proposed ModerationBench exposes a significant coverage gap in #Bluesky 's deployed moderation system: human annotators found nearly 35% of posts semantically similar to moderated content were unsafe and missed by the platform.
130
Abhisek Dash @adash93.bsky.social · 12/09/2026
Following our ICWSM work, in the next work, we ask the question Can foundation models moderate online content? We present MODERATIONBENCH a benchmark of multimodal online posts and @bluesky moderation decisions. Arxiv: arxiv.org/abs/2609.10410… Project page: moderation-bench.github.io
arxiv.org
141
Abhisek Dash @adash93.bsky.social · 12/09/2026
Question: Now that we know the blindspot of a deployed moderation system, how do we improve the operationalization of the complicated moderation policies to plug the gap?
010
Abhisek Dash @adash93.bsky.social · 12/09/2026
Takeaway: the biggest takeaway comes from the open protocol and how it enables researchers to conduct research on prevalent harms on the platform. This could be something for the @EU_Commission to look at for streamlining operationalization of audit requirements of the Digital Services Act.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
With those drawbacks, still the highly precise moderation system flags a lot of NSFW, Graphic and Self harm related contents, along with hateful, profane and content promoting violence and threat.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
This lack of coverage, can be attributed to two primary reasons. A) It's infeasible and cognitively challenging for human moderators to operationalize long and complicated moderation policies at scale. B) Rule based automated pipeline is brittle where many potentially harmful posts don't get flagged
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
In terms of effectiveness, we find BMS is highly precise. In a random sample of moderated posts, human moderators agree that 84% of the posts moderated by BMS needed moderation. However, on a random sample, we worryingly find human annotators flag 4 times more posts to be harmful than BMS.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
We find that the moderation pipeline is a Human-AI collaboration with automated systems used for detecting graphic and sexual content in seconds and more nuanced harms (eg, hatespeech or profanity) being detected with significantly more human oversight.
110
Abhisek Dash @adash93.bsky.social · 12/09/2026
In this work, we ask three basic questions about the #Bluesky #Moderation system. 1. How does Bluesky moderate the online content? 2. How effective is Bluesky Moderation Service? 3. What's the landscape of harms it detects?
120
Abhisek Dash @adash93.bsky.social · 12/09/2026
The preprint and projects pages are up. Arxiv: arxiv.org/abs/2609.113... Codebase: github.com/iampushpdeep... Project page: bms-audit.github.io
arxiv.org
Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms
Empirical research on content moderation is fundamentally constrained by the opaque deployment of moderation systems on major social media platforms. To this end, the recent emergence of decentralized...
120
Abhisek Dash @adash93.bsky.social · 17/07/2026
Joint work with @pushpdeep.bsky.social, Sayeh Jarollahi, Ayan Majumdar, Vabuk Pahari, Abhijnan Chakraborty, Krishna P. Gummadi and @ingmarweber.de. Preprint, Codebase and more details: Coming Soon. Stay Tuned!
010
Abhisek Dash @adash93.bsky.social · 17/07/2026
🥳🥳🥳 Pleased to share that our work on characterizing #ContentModeration on @bsky.app has been accepted in @icwsm.bsky.social 2027. In this work, we investigate three aspects of @moderation.bsky.app (a) its degree of automation, (b) its effectiveness and (c) the landscape of harm it detects.
131
Abhisek Dash @adash93.bsky.social · 06/02/2026
Huge shoutout to the team : Soumi Das, Elisabeth Kirsten, Qinyuan Wu, Sai Keerthana Karnam, Krishna Gummadi, Thorsten Holz, Muhammad Bilal Zafar, zsavvas90.bsky.social
010
Abhisek Dash @adash93.bsky.social · 06/02/2026
We show that personalization now reaches into users’ goals, desires, and beliefs, is largely system-driven, often involves sensitive information, yet is mostly grounded in user input. #ConversationalAI #Privacy #Memory #ResponsibleAI #GDPR #Agency
110
Abhisek Dash @adash93.bsky.social · 06/02/2026
In Bowling with ChatGPT, we showed how people increasingly use conversational AI as a companion. In our new paper (accepted at #TheWebConf2026), we ask what happens to all that self-disclosure-- and study ChatGPT’s Memory feature. Arxiv : arxiv.org/abs/2602.01450
arxiv.org
The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT
To enable personalized and context-aware interactions, conversational AI systems have introduced a new mechanism: Memory. Memory creates what we refer to as the Algorithmic Self-portrait - a new form ...
110
Reposted by Abhisek Dash
Abhisek Dash @adash93.bsky.social · 03/02/2026
#BowlingAlone documented a quiet shift from collective to solitary practices. In Bowling with ChatGPT ( #TheWebConference2026), we study whether a similar shift is unfolding in interactions with conversational AI. arxiv link: arxiv.org/abs/2602.01114
arxiv.org
Bowling with ChatGPT: On the Evolving User Interactions with Conversational AI Systems
Recent studies have discussed how users are increasingly using conversational AI systems, powered by LLMs, for information seeking, decision support, and even emotional support. However, these macro-l...
254
Abhisek Dash @adash93.bsky.social · 03/02/2026
Using 825K real interactions donated using #GDPR Right of Access, we show that over time, #ChatGPT is increasingly engaged not just as a tool, but in more social and relational ways. Joint work with : Sai Keerthana Karnam, Krishna Gummadi, Animesh Mukherjee, @ingmarweber.de, @zsavvas90.bsky.social
010
Abhisek Dash @adash93.bsky.social · 03/02/2026
#BowlingAlone documented a quiet shift from collective to solitary practices. In Bowling with ChatGPT ( #TheWebConference2026), we study whether a similar shift is unfolding in interactions with conversational AI. arxiv link: arxiv.org/abs/2602.01114
arxiv.org
Bowling with ChatGPT: On the Evolving User Interactions with Conversational AI Systems
Recent studies have discussed how users are increasingly using conversational AI systems, powered by LLMs, for information seeking, decision support, and even emotional support. However, these macro-l...
254
Abhisek Dash @adash93.bsky.social · 18/01/2026
Interesting. But, would not it make the whole user experience worse when ads show up? Do you know how ads show up in ChatGPT UI?
110
Abhisek Dash @adash93.bsky.social · 14/01/2026
Paper #3 : The Algorithmic Self-Portrait:Deconstructing Memory in ChatGPT #ChatGPT #Memory #Personalization #ConversationalAI
000
Abhisek Dash @adash93.bsky.social · 14/01/2026
Paper #2 : Bowling with ChatGPT: On the Evolving User Interactions with Conversational AI Systems #GDPR #RightOfAccess #ChatGPT #ConversationalAI
100
Abhisek Dash @adash93.bsky.social · 14/01/2026
Paper #1 : Does Ad-Free Mean Less Data Collection? An Empirical Study of Platform Data Practices and User Expectations #Instagram #Facebook #X #DataCollection #GDPR
100
Abhisek Dash @adash93.bsky.social · 14/01/2026
Excited to share that three papers from our group have been accepted at The Web Conference 2026 (www2026.thewebconf.org) Huge shout out to Sepehr Mousavi, Sai Keerthana Karnam, Qinyuan Wu and Elisabeth Kirsten for leading the works. Stay tuned for more details. #TheWebConf2026
www2026.thewebconf.org
The Web Conference 2026 - Dubai, UAE
100
Reposted by Abhisek Dash
Saarland Informatics Campus @saarland-informatics-campus.de · 09/01/2026
Congrats to Prof. Dr. @ingmarweber.de on his AAAI Fellowship! The fellowship honors individuals with significant contributions to #AI. Working in societal computing, he bridges AI and the social sciences, analyzing social platform data for valuable insights on society. 👏 Read more: sic.link/weber
image displays the logos for the Saarland Informatics Campus (SIC) and Saarland University. Below the logos is a professional headshot of Prof. Dr. Ingmar Weber with glasses and a grey button-down shirt, looking directly at the camera against a blurred indoor background.At the bottom, white text on a dark blue banner reads: "AI for Society: Ingmar Weber elected as an AAAI fellow".
052
Reposted by Abhisek Dash
Ingmar Weber @ingmarweber.de · 23/07/2025
In today's @ic2s2.bsky.social #IC2S2 poster session:: "Studying Behavioral Addiction by Combining Surveys and Digital Traces: A Case Study of TikTok" Joint work with Cai Yang, Sepehr Mousavi, @adash93.bsky.social and Krishna Gummadi. Find me at poster #37. Full paper: ojs.aaai.org/index.php/IC...
1284
Abhisek Dash @adash93.bsky.social · 30/06/2025
Who would know NeurIPS review deadline may also be detrimental for submission platforms 😜
000
Abhisek Dash @adash93.bsky.social · 26/06/2025
Congratulations to all...❤️👏
000
Abhisek Dash @adash93.bsky.social · 25/06/2025
2. DSA has enabled multiple data reporting which could complement each other once the delegated act is passed and researchers get a chance to request for private data under DSA Article 40 (4). #DSA #ICWSM #DataAccess
010
Abhisek Dash @adash93.bsky.social · 25/06/2025
My top 2 key takeaways 1. Data scraping can be legally justified if you satisfy certain conditions of eligibility. Under DSA Article 40 (12) platforms are providing publicly accessible data too. Although, many studies question their reliability now. @icwsm.bsky.social
110
Abhisek Dash @adash93.bsky.social · 24/06/2025
Fascinating discussion on data access under Digital Services Act to make platforms more "Auditable" to assess and mitigate systemic risks.
030