AI Fact-Checkers Often Disagree on Truthfulness of Claims

AI Fact-Checking Disagreement

نظرة سريعة

A study tested 5 advanced AI systems on 1,000 real-world fact-check claims, finding disagreement in 67.2% of cases, with severe disagreements (true vs. false) in 34%.
Agreement was limited (Krippendorff’s alpha: 0.639), mostly at extremes (true/false), with little consensus on nuanced verdicts (mostly true/misleading).

ملخص مُنشأ بالذكاء الاصطناعي

حجم الخط

A study tested 5 advanced AI systems on 1,000 real-world fact-check claims, finding disagreement in 67.2% of cases, with severe disagreements (true vs. false) in 34%. Agreement was limited (Krippendorff’s alpha: 0.639), mostly at extremes (true/false), with little consensus on nuanced verdicts (mostly true/misleading).

Read the full article on Decrypt

This article was originally published by Decrypt.

أخبار ذات صلة

المزيد حول هذا الموضوعAI

يتطور·54 dk önce

Hackers Compromise Injective Software Package in Supply Chain Attack

A popular npm package for the Injective blockchain was compromised in a supply chain attack, with malware designed to steal crypto wallet private keys and seed phrases. The malicious code, present in version 1.20.21 of the @injectivelabs/sdk-ts package, has been removed, and affected versions deprecated, with no funds on the network reported at risk.

Cointelegraph

يتطور·6 sa önce

AI Hallucinations Could Be Exploited by Hackers for Botnets, Research Warns

New research from Tel Aviv University, Technion, and Intuit reveals a technique called "HalluSquatting" that exploits AI hallucinations. Hackers can register fake software repository links predicted by AI, potentially leading to AI-enabled botnets and compromising computers.

Decrypt

يتطور·7 sa önce

Perplexity Uses Chinese Open-Source Model to Cut AI Costs

Perplexity has fine-tuned Z.AI's GLM 5.2, an open-source Chinese model, to act as a cost-effective workhorse for its AI agent. This allows Perplexity to achieve performance comparable to more expensive models like Claude Opus 4.8 at a fraction of the cost by intelligently handling tasks or escalating them.

Decrypt

يتطور·7 sa önce

Former Fed Chair Ben Bernanke Joins Anthropic's AI Oversight Board

Former Fed Chair Ben Bernanke has joined Anthropic's AI oversight board, aiming to guide the company's long-term strategy and assess AI's economic and societal impacts. His appointment comes amid debates over AI valuations resembling market bubbles.

Decrypt

يتطور·9 sa önce

Ethereum Foundation Uses AI Agents for Security Red Teaming

The Ethereum Foundation is employing AI agents to proactively identify vulnerabilities in its core software, cryptographic systems, and smart contracts. This red teaming effort has already uncovered bugs, including a critical issue in libp2p's gossipsub.

Decrypt

يتطور·10 sa önce

OpenAI Releases GPT-5.6 Sol, New Flagship Model, Alongside Terra and Luna

OpenAI launched its new flagship AI model, GPT-5.6 Sol, to the public, alongside smaller models Terra and Luna. Sol features advanced capabilities, including a max reasoning effort and ultra mode, and is priced competitively against other leading AI models. Benchmarks show Sol outperforming competitors in specific tasks, though some testers note Anthropic's models excel in writing.

Decrypt

المزيد حول هذا الموضوعAI

AI Fact-Checkers Often Disagree on Truthfulness of Claims

نظرة سريعة

مواضيع ذات صلة

أخبار ذات صلة

Hackers Compromise Injective Software Package in Supply Chain Attack

AI Hallucinations Could Be Exploited by Hackers for Botnets, Research Warns

Perplexity Uses Chinese Open-Source Model to Cut AI Costs

Former Fed Chair Ben Bernanke Joins Anthropic's AI Oversight Board

Ethereum Foundation Uses AI Agents for Security Red Teaming

OpenAI Releases GPT-5.6 Sol, New Flagship Model, Alongside Terra and Luna