Why do LLMs make stuff up? New research peers under the hood.

Quick Summary:

Claude's faulty "known entity" neurons sometime override its "don't answer" circuitry.

Anthropic's newly published research this week expands on that previous work by tracing how these features can affect other neuron groups that represent computational decision "circuits" Claude follows in crafting its response.

From a human perspective, it can be hard to understand why these models don't simply say "I don't know" instead of making up some plausible-sounding nonsense.

---

More Pictures

---

Read Quick SummarySave Time

or

Article Details

Author / Journalist: Kyle Orland

Category: Technology

Markets: United StatesJordan

Topics: Features Model Researchers Research Anthropic Prompt Answer Known Neurons Claude Entity

Source Website Secure: Yes (HTTPS)

News Sentiment: Neutral

Fact Checked: Legitimate

Article Type: News Report

Published On: 2025-03-28 @ 22:33:51 (2 days ago)

News Timezone: GMT -5:00

News Source URL: arstechnica.com

Language: English

Article Length: 212 words

Reading Time: 2 minutes read

Sentences: 51 lines

Sentence Length: 5 words per sentence (average)

Platforms: Desktop Web, Mobile Web, iOS App, Android App

Copyright Owner: © Ars Technica

News ID: 27392100

View Article Analysis

About Ars Technica

Main Topics: Technology

Official Website: arstechnica.com

Update Frequency: 2 posts per day

Year Established: 1998

Headquarters: United States

News Last Updated: 11 hours ago

Coverage Areas: United States

Ownership: Independent Company

Publication Timezone: GMT -5:00

Content Availability: Worldwide

News Language: English

RSS Feed: Available (XML)

API Access: Available (JSON, REST)

Website Security: Secure (HTTPS)

Publisher ID: #32

Publisher Details

Frequently Asked Questions