PulseAugur
EN
LIVE 12:49:11
한국어(KO) AI에게 고장 난 계산기를 줬다 — 혼자 맞히던 산수를 도구를 받고 틀렸다 Luna와 Haiku에게 같은 산수 12문항을 도구 없이, 정상 계산값과 함께, 틀린 계산값과 함께 줬다. 혼자서는 둘 다 전부 맞혔지만 1만 어긋난 도구 결과를 주자 오답을 받아들였다. 답란에는 틀린 숫자를,

AI models adopt incorrect tool outputs, even when capable of solving problems correctly

By PulseAugur Editorial · [1 sources] ·

A study tested how AI models Luna and Haiku handle incorrect information from tools. When given a faulty calculator that provided slightly wrong answers, both models, which could solve the math problems correctly on their own, adopted the incorrect tool-generated answers. This occurred even when the tool's output was significantly off, such as a value ten times the correct answer. The models sometimes incorporated the incorrect numbers into their explanations, even contradicting their own correct calculations. AI

IMPACT AI models may blindly trust external tools, potentially leading to incorrect outputs even when they possess the correct information.

RANK_REASON The cluster describes an experiment testing AI model behavior with faulty tools, which is a research finding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Mastodon — mastodon.social →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

AI models adopt incorrect tool outputs, even when capable of solving problems correctly

How we ranked this

Signal score
0 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster describes an experiment testing AI model behavior with faulty tools, which is a research finding. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, other
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
1 days old
Coverage has settled into its steady-state source set.

Full methodology in our editorial standards.

COVERAGE [1]

  1. Mastodon — mastodon.social TIER_1 한국어(KO) · [email protected] ·

    I gave an AI a broken calculator — it got arithmetic wrong after getting it right on its own. I gave Luna and Haiku the same 12 arithmetic problems with and without tools, with correct and incorrect calculation values. They both got them all right on their own, but when given tool results that were off by just 1, they accepted the wrong answers. In the answer field, they put the wrong number,

    AI에게 고장 난 계산기를 줬다 — 혼자 맞히던 산수를 도구를 받고 틀렸다 Luna와 Haiku에게 같은 산수 12문항을 도구 없이, 정상 계산값과 함께, 틀린 계산값과 함께 줬다. 혼자서는 둘 다 전부 맞혔지만 1만 어긋난 도구 결과를 주자 오답을 받아들였다. 답란에는 틀린 숫자를, 바로 아래 풀이에는 맞는 숫자를 쓴 응답도 있었다. https://zerry.co.kr/blog/broken-calculator-trust #LLM #AI #도구호출 #실측

RELATED ENTITIES

RELATED TOPICS