Abhisek Dash

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

Joint work with Ayan Majumdar, Shounak Paul, Pushpdeep Singh, Ines Abdelaziz, Sayeh Jarollahi, Seungeon Lee, Krishna P. Gummadi, Ingmar Weber.

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

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.

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

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.

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

The preprint and projects pages are up. Arxiv: arxiv.org/abs/2609.113... Codebase: github.com/iampushpdeep... Project page: bms-audit.github.io

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.

The preprint and projects pages are up. Arxiv: arxiv.org/abs/2609.113... Codebase: github.com/iampushpdeep... Project page: bms-audit.github.io

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

The preprint and projects pages are up. Arxiv: arxiv.org/abs/2609.113... Codebase: github.com/iampushpdeep... Project page: bms-audit.github.io

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.

The preprint and projects pages are up. Arxiv: arxiv.org/abs/2609.113... Codebase: github.com/iampushpdeep... Project page: bms-audit.github.io

🥳🥳🥳 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.

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

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

#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

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

Paper #2 : Bowling with ChatGPT: On the Evolving User Interactions with Conversational AI Systems #GDPR #RightOfAccess #ChatGPT #ConversationalAI

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

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