Dhrumil Mehta
@dmil.bsky.social
110 followers 64 following 10 posts
Data Journalism Prof. @ Columbia University and Deputy Director of Tow Center for Digital Journalism.
Posts Media Videos Starter Packs
Reposted by Dhrumil Mehta
nausheen.bsky.social
Today at #datajconf: we talked about what we've learned about pedagogy in the process of running datajournalismteachers.club. Focused on grounding data journalism in tools *and* humanities, teaching with a learner-centered rather than a "banking" model, and how we learn in community. Join the club!
dmil.bsky.social
Can AI tools fact-check images? The Tow Center tested 10 images, asking AI tools to: 𝗰𝗼𝗻𝗳𝗶𝗿𝗺 𝘄𝗵𝗲𝘁𝗵𝗲𝗿 𝗲𝗮𝗰𝗵 𝗶𝗺𝗮𝗴𝗲 𝗶𝘀 𝗿𝗲𝗮𝗹 and identify its 𝗹𝗼𝗰𝗮𝘁𝗶𝗼𝗻, 𝗱𝗮𝘁𝗲, and 𝘀𝗼𝘂𝗿𝗰𝗲.

Results were...well, there's a lot of red on this chart.

Read the research here:
www.cjr.org/tow_center/w...
dmil.bsky.social
“AI search” has a citation problem.

Latest from @towcenter.bsky.social

www.cjr.org/tow_center/w...
dmil.bsky.social
Yes please! Just started BlueSky-ing. Thanks!
dmil.bsky.social
It turns out this problem extends to many other chatbots too. In @towcenter.bsky.social's tests, they often cited the wrong article, and sometimes also fabricated links & cited syndicated or plagiarized articles instead of the original.
dmil.bsky.social
This builds on prior tests showing that ChatGPT frequently cited the wrong source article, but rarely expressed uncertainty. If it didn't know the answer (at least to the kinds of questions we asked), it confidently made one up. cjr.org/tow_center/how-chatgpt-misrepresents-publisher-content.php
dmil.bsky.social
Using AI search tools for research? Tow Center asked 8 AI search tools to find the source of excerpts from news articles (something Google search does pretty well). They often cited the wrong sources with confidence.

Find @aisvarya17.bsky.social at #NICAR25 to hear more.
🧵Thread w/ findings below👇
klaudia.bsky.social
We found that:

1️⃣ Chatbots were generally bad at declining to answer questions they couldn’t answer accurately, offering incorrect or speculative answers instead.
Graph title: "Generative search tools were often confidently wrong in our study"
Subtitle: "The Tow Center asked eight generative search tools to identify the source article, the publication and URL for 200 excerpts extracted from news articles by 20 publishers. Each square represents the citation behavior of a response."
dmil.bsky.social
Are you using Generative AI tools as your search engine? @towcenter.bsky.social ran a test to see how often they cite news articles correctly given direct quotes from them.

Results are troubling.

www.cjr.org/tow_center/w...

Great work by @aisvarya17.bsky.social and @klaudia.bsky.social
[graph] Generative search tools were often confidently wrong in our study
dmil.bsky.social
Hello #NICAR25!

Are you an educator who is bringing students to NICAR? Sign them up for the NICAR Scavenger Hunt!

nicarscavengerhunt.com

Students interview working journalists about things they’re learning in your class and you get new example stories related to your course content. Yay!
NICAR Scavenger Hunt
The NICAR25 Scavenger Hunt. Brought to you by the Data Journalism Teachers' Club.
nicarscavengerhunt.com