People are more susceptible to misinformation with realistic AI-synthesized images that provide strong evidence to headlines
Sean Guo, Yiwen Zhong, and Xiaoqing Hu
This pre-registered and peer-reviewed study found that “the realism of images and the amount of evidence they provided to headlines were significant positive predictors of belief in false headlines.” More hopefully, post-event corrections did reduce belief in the falsehood as did invitations to scrutinize images carefully.
Why Language Models Hallucinate
Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang
In this paper, researchers from OpenAI and Georgia Tech claim that AI chatbots hallucinate because they’re incentivized to provide an answer regardless of how confident they are about it. Introducing an “I don’t know” option in benchmark tests and penalizing it less than an inaccurate answer may reduce the problem, the researchers argue.
GenAI Misinformation, Trust, and News Consumption: Evidence from a Field Experiment
Filipe R. Campante, Ruben Durante, Felix Hagemeister, and Ananya Sen
This experimental study on Süddeutsche Zeitung readers tried to discern the impact of AI-generated misinformation on user trust in the media. The group that was given a short quiz to detect a deepfaked photo went on to trust all news (including SZ) a little less, but visited the SZ website a little more. They were also a bit more likely to keep their subscription.
Assessing the System-Instruction Vulnerabilities of Large Language Models to Malicious Conversion Into Health Disinformation Chatbots
Natansh D. Modi, Bradley D. Menz, Abdulhalim A. Awaty, Cyril A. Alex, Jessica M. Logan, Ross A. McKinnon, Andrew Rowland, Stephen Bacch, Kacper Gradon, Michael J. Sorich, and Ashley M. Hopkins
This group of researchers was able to build custom chatbots spreading medical disinformation via the APIs of five widely-used foundational LLMs. Of the five, only Claude’s Anthropic didn’t return disinformation for all of the queries attempted. The other four gave inaccurate responses 20 out of 20 times.
This is my test accordion for Jeet - Question
Filipe R. Campante, Ruben Durante, Felix Hagemeister, and Ananya Sen
This experimental study on Süddeutsche Zeitung readers tried to discern the impact of AI-generated misinformation on user trust in the media. The group that was given a short quiz to detect a deepfaked photo went on to trust all news (including SZ) a little less, but visited the SZ website a little more. They were also a bit more likely to keep their subscription.
Lorem ipsum dolor sit amet. Et sint mollitia id cumque cumque ut aliquam voluptas et sequi magnam ex voluptas dignissimos. Et delectus distinctio quo odio debitis quo error dolorum. Aut odio nihil vel nihil excepturi qui nobis blanditiis sed repudiandae iure quo exercitationem asperiores est veniam autem non ipsam placeat. Qui facilis labore cum asperiores quod ut voluptas soluta a aperiam accusantium nam impedit sint ea eveniet vitae.
Lorem ipsum dolor sit amet. Et sint mollitia id cumque cumque ut aliquam voluptas et sequi magnam ex voluptas dignissimos. Et delectus distinctio quo odio debitis quo error dolorum. Aut odio nihil vel nihil excepturi qui nobis blanditiis sed repudiandae iure quo exercitationem asperiores est veniam autem non ipsam placeat. Qui facilis labore cum asperiores quod ut voluptas soluta a aperiam accusantium nam impedit sint ea eveniet vitae.
- This is a bullet point
- This is another bullet Point
- This is a number bullet point.
Assessing the System-Instruction Vulnerabilities of Large Language Models to Malicious Conversion Into Health Disinformation Chatbots
Natansh D. Modi, Bradley D. Menz, Abdulhalim A. Awaty, Cyril A. Alex, Jessica M. Logan, Ross A. McKinnon, Andrew Rowland, Stephen Bacch, Kacper Gradon, Michael J. Sorich, and Ashley M. Hopkins
This group of researchers was able to build custom chatbots spreading medical disinformation via the APIs of five widely-used foundational LLMs. Of the five, only Claude’s Anthropic didn’t return disinformation for all of the queries attempted. The other four gave inaccurate responses 20 out of 20 times.
