Sign in

Neil Sehgal

@nsehgal.bsky.social
151 followers 89 following 30 posts

CS PhD @upenn.bsky.social Computational Social Science @WorldBank Harvard, Brown alumn sehgal-neil.github.io

PostsRepliesMedia
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
Also tagging some people who may be interested or whose great prior work helped inspire this project! @seanjwestwood.bsky.social, @dgrand.bsky.social, @emollick.bsky.social, @joinprolific.bsky.social, @spur.science, @davmicrot.bsky.social, @manoelhortaribeiro.bsky.social, @veniamin.bsky.social
220
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
Takeaway: safeguards should look beyond free text and consider repeated use across studies. A small group accounts for most observed assistance, so worker-level signals may help. Paper: arxiv.org/abs/2610.09279 w/ @manueltonneau.bsky.social, Dunigan Folk, Lyle Ungar, & Sharath Chandra Guntuku
arxiv.org
AI-Assisted Submissions in Online Research Are Rare and Highly Concentrated but Routinely Approved
Online research platforms underpin much of what science claims about people, on the assumption that a human produced each response. Generative AI threatens that assumption by letting participants dele...
111
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
One important caveat: we’re specifically focusing on the threat of AI assistance, not autonomous agents. We observed ChatGPT histories, not every AI system or unobserved use, and our sample consisted of weekly ChatGPT users. These could bias our estimates in either direction.
110
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
So is AI an existential threat to online research right now? Our results don't support that broad view. The overall rate we observed was low and stable. But near-universal approval of assisted work and gaps in detector coverage show clear weaknesses in current safeguards.
100
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
One last worrying finding: AI-assisted submissions were approved 99.6% of the time, basically as often as unassisted submissions (99.7%). In 60 cases where we directly observed instructions explicitly banning AI, all 60 were still approved and paid.
101
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
This isn't only an open-ended-text problem. 53% of assisted submissions involved bounded responses (e.g., choices, ratings, exact answers). This is important because Prolific's optional LLM assistance detector only screens free text questions.
101
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
We also found no detectable upward trend over more than 3 years. And use was often localized: 40% of assisted submissions involved just one ChatGPT exchange. In the rest, messages spanned a median of 29% of the study's duration.
100
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
And that 1% was VERY concentrated. Just 5% of workers accounted for 2/3 of assisted submissions; the top 10% accounted for almost 80%! Excluding the top 5% dropped the assistance rate from 1% to 0.40% of submissions.
111
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
First surprise: 68% of these weekly ChatGPT users had used it on a study at least once. Yet AI assistance appeared in only 1% of their submissions. So using AI at least once was common; using it on a substantial share of studies was not.
110
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
That gave us 700k+ submissions across 127k+ studies, from Dec. 2022 to Feb. 2026. We counted a submission as AI-assisted when we observed ChatGPT’s output answering or materially contributing to a study response.
100
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
So we tried something different. We first surveyed 2,500 Prolific workers about their AI use, then directly observed AI assistance by linking previously donated ChatGPT histories with Prolific submission records from 408 workers who used ChatGPT at least weekly.
100
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
The concern is simple: online research assumes a human being produced each response. LLMs make it easy to outsource part of a task. But how often does that actually happen? Estimates on Prolific range from 0% to 35%, often from single tasks, self-reports, or imperfect detectors.
110
Neil Sehgal @nsehgal.bsky.social · 08/10/2026
🚨 New paper 🚨 Are your participants using AI to take your study? We conducted the first study directly observing AI assistance in online research, spanning 700K+ Prolific submissions. We found very rare, concentrated AI use, but gaps in safeguards. arxiv.org/abs/2610.09279
13620
Reposted by Neil Sehgal
JAMA Network Open @jamanetworkopen.com · 09/06/2026
Chatbots supporting #HPV vaccine promotion matched public health materials for short-term intent but showed less durable effects and did not increase vaccination uptake. ja.ma/4uTd50g
RCT study on HPV vaccine intent. Chatbot conversations increased immediate parental intent for children's HPV vaccination versus no intervention. Chatbots didn't outperform public health materials. A graph shows changes in intent for conversational/default chatbots, public health, & no intervention groups.
022
Reposted by Neil Sehgal
Manuel Tonneau @manueltonneau.bsky.social · 27/01/2026
Demographic cues (eg, names, dialect) are widely used to study how LLM behavior may change depending on user demographics. Such cues are often assumed interchangeable. 🚨 We show they are not: different cues yield different model behavior for the same group and different conclusions on LLM bias. 🧵👇
1189
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
@pennldi.bsky.social @pennengineering.bsky.social @upenn.edu @pennchibe.bsky.social @pennmedicine.bsky.social @pminnovation.bsky.social
000
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
We hope these findings help health systems design more effective & scalable outreach to close preventive care gaps. Thoughts welcome! w/ @manueltonneau.bsky.social, @alison-buttenheim.bsky.social, @sharathg.bsky.social + team
100
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
💡 Bottom line: 🔹 LLMs can generate short, tailored, clinically appropriate messages that move intent particularly for lower-barrier behaviors. 🔹 These messages can fit into portals, texts, or mailed materials. 🔹 They’re low-cost & scalable. Read more: arxiv.org/abs/2507.08211
arxiv.org
Effect of Static vs. Conversational AI-Generated Messages on Colorectal Cancer Screening Intent: a Randomized Controlled Trial
Large language model (LLM) chatbots show increasing promise in persuasive communication. Yet their real-world utility remains uncertain, particularly in clinical settings where sustained conversations...
100
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
📈 Results: ✅ Both AI formats significantly boosted stool-test intent (+13 pts) over expert material. 🩺 For colonoscopy, no AI advantage over expert material. Surprisingly: single AI message ≈ chatbot – despite participants choosing to spend 3.5 minutes longer with the chatbot!
100
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
🧪 In a randomized trial (n=915), we compared: 1️⃣ No intervention 2️⃣ Expert-written patient materials 3️⃣ Single AI message 4️⃣ AI chatbot using motivational interviewing techniques Outcome: intent to screen (stool test & colonoscopy) over 12 months.
100
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
🩺 Why it matters: Colorectal cancer is the 2nd leading cause of cancer death in the US – but ~1/3 of eligible adults aren’t screened. We need scalable, persuasive tools to close this gap. Can AI help?
100
Neil Sehgal @nsehgal.bsky.social · 14/07/2025
🚨 New study! We tested whether AI-generated messages – single static messages vs. conversations – can boost intent to screen for colorectal cancer. Turns out: short, tailored AI messages outperform expert-written materials & match conversations, at a fraction of the time! 🧵👇
131
Neil Sehgal @nsehgal.bsky.social · 01/05/2025
tagging some others who may be interested! @chrisbail.bsky.social @hugoreasoning.bsky.social @tnfalpha.bsky.social @emollick.bsky.social @jennyallen.bsky.social @noelbrewer.bsky.social @julieleask.bsky.social @pminnovation.bsky.social @susanmichie.bsky.social
110
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
@pennldi.bsky.social @pennengineering.bsky.social @upenn.edu @pennchibe.bsky.social @pennmedicine.bsky.social
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
Shout-outs to inspiring work: @gordpennycook.bsky.social @dgrand.bsky.social @tomcostello.bsky.social, @jeffhancock.bsky.social @kobihackenburg.bsky.social on AI persuasion & others pushing this field forward 🙌
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
Thoughts welcome! w/ @manueltonneau.bsky.social, @sharathg.bsky.social, @alison-buttenheim.bsky.social +team
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
Take-home: chatbots can nudge short-term intent, but add little over high-quality public-health materials. AI looks best as an add-on, not a replacement, in vaccine communications. Link here: arxiv.org/abs/2504.20519
arxiv.org
Conversations with AI Chatbots Increase Short-Term Vaccine Intentions But Do Not Outperform Standard Public Health Messaging
Large language model (LLM) based chatbots show promise in persuasive communication, but existing studies often rely on weak controls or focus on belief change rather than behavioral intentions or outc...
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
In a 15-day follow-up, gains from the reading arm stuck (+7 pts) while chatbot effects faded to ≈0. We also found no spill-over to flu/COVID or general vaccine hesitancy
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
In an RCT with 930 parents (US/CA/UK, with kids old enough for the HPV vaccine): chatbots raised vaccine intent vs. no intervention—but neither variant beat simply reading official public-health materials, with the conversational chatbot doing significantly worse.
100
Neil Sehgal @nsehgal.bsky.social · 30/04/2025
🚨 New preprint on AI persuasion and public health 🚨 A 3-min conversation with GPT-4o nudged HPV-vax-hesitant parents (who obv knew it was AI & consented!)—BUT reading standard public-health material still outperformed chatbots in impact and longevity. Details below 👇
2147
Neil Sehgal @nsehgal.bsky.social · 21/04/2025
🚨 New CHI'25 EA paper! 🚨 How can we design culturally sensitive mental health chatbots for Indian adolescents? 🇮🇳📱 Our mixed-methods study reveals key design insights—from stigma to personalization. Read it here: arxiv.org/abs/2503.08562 #CHI2025 #HCI #MentalHealth #AIforGood #India
arxiv.org
Exploring Socio-Cultural Challenges and Opportunities in Designing Mental Health Chatbots for Adolescents in India
Mental health challenges among Indian adolescents are shaped by unique cultural and systemic barriers, including high social stigma and limited professional support. Through a mixed-methods study invo...
071
Neil Sehgal @nsehgal.bsky.social · 12/03/2025
Thanks for compiling this! Would you be able to add our paper on designing mental health chatbots for Indian adolescents? arxiv.org/abs/2503.08562
arxiv.org
Exploring Socio-Cultural Challenges and Opportunities in Designing Mental Health Chatbots for Adolescents in India
Mental health challenges among Indian adolescents are shaped by unique cultural and systemic barriers, including high social stigma and limited professional support. Through a mixed-methods study invo...
110
Reposted by Neil Sehgal
Penn LDI @pennldi.bsky.social · 03/03/2025
LDI Senior Fellows Neil Sehgal, Anish Agarwal, Raina Merchant, Sharath Chandra Guntuku, and colleagues analyzed Yelp reviews of health care facilities to asses how patient sentiment toward changed before and after COVID-19.
161