Add BrightSurf on Google Email

Study: Generative AI succumbs to conversational misinformed pressure and argument

Researchers at the University of Arizona assessed seven generative AI models for fallibility, persuadability and correctability during lengthy conversations. The study revealed intrinsic limitations of these models, including oscillation between accepting and rejecting false statements, and the need for careful human engagement to miti...

SourceUniversity of Arizona·JournalScientific Reports·TypeData/statistical analysis·DateSep 4, 2026

AI makes human-like reasoning mistakes

Large language models (LMs) exhibit human-like reasoning flaws, particularly in tasks requiring semantic content and believability. LMs struggle to distinguish between valid and invalid arguments, mirroring human errors, especially when semantic content is sensical and believable.

SourcePNAS Nexus·JournalPNAS Nexus·DateJul 16, 2024

Clinicians grapple with decisions in crisis-care simulation

A simulated crisis-care event revealed the moral distress experienced by triage team members who had to prioritize patients for scarce resources. The study aimed to operationalize a process for making life-and-death patient decisions, but triage-team members struggled with balancing individual patient needs with fair resource allocation.

SourceUniversity of Washington School of Medicine/UW Medicine·JournalJAMA Network Open·TypeSurvey·DateApr 18, 2022

Logical reasoning: An antidote or a poison for political disagreement?

A new study explores ideological belief bias, where people judge logical arguments based on their believability rather than sound premises. Liberals and conservatives showed varying levels of ability to identify flawed arguments supporting opposing views. The findings suggest that being open to the other side can lead to a better under...

SourceSociety for Personality and Social Psychology·JournalSocial Psychological and Personality Science·DateApr 17, 2019

Identifying problems with national identifiers

Researchers from Harvard University found that Resident Registration Numbers in South Korea can be easily decrypted using computation and logical reasoning. The study reveals vulnerabilities in the encoding system, compromising individual privacy and urging a more robust redesign.

SourceHarvard University·JournalTechnology Sciences·DateSep 28, 2015