惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

V
Visual Studio Blog
N
Netflix TechBlog - Medium
GbyAI
GbyAI
大猫的无限游戏
大猫的无限游戏
博客园 - 三生石上(FineUI控件)
T
Tailwind CSS Blog
IT之家
IT之家
博客园 - Franky
雷峰网
雷峰网
博客园 - 聂微东
腾讯CDC
M
MIT News - Artificial intelligence
B
Blog RSS Feed
博客园_首页
罗磊的独立博客
S
SegmentFault 最新的问题
I
InfoQ
博客园 - 叶小钗
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
阮一峰的网络日志
阮一峰的网络日志
D
Docker
宝玉的分享
宝玉的分享
B
Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报

Latest from TechRadar in News

VodafoneThree gets Ofcom approval to bring satellite connectivity to your smartphone NYT Connections today – my hints and answers for April 16 (#1040) Quordle hints and answers for Thursday, April 16 (game #1543) NYT Strands hints and answers for Thursday, April 16 (game #774) Is this the tipping point for AI at work? New Gallup survey finds half of all US employees now use it in some way Allbirds — the shoe viral company — just pivoted into AI, and I wish this were an Onion headline 'Every Apple user needs to know about this nasty scam': Fake warnings tell users their iCloud data will be… 'Makes it even more disappointing': Microsoft backs fossil fuel big time with $7 billion deal in race for AI… 'Maybe it’s not science fiction': Solar panels are causing rainwater to fall in one of the driest places… Maine becomes first US state to pass data centre construction ban Dozens of WordPress plugins hijacked to target thousands of sites Drone-killing laser weapons greenlit for use in US airspace – FAA and Defense Department say high-energy weapons are ‘ready to protect all air travelers from illicit drone use’ despite airspace restrictions and friendly-fire incidents 'We are currently being extorted' — crypto giant Kraken says it is facing extortion attack, here's… McGraw Hill becomes latest to see its Salesforce data hacked Looking for a new PC? Now might be great time to upgrade, as Gartner figures claim shipments are rising — while… Farewell Surface Hub — Microsoft kills off its super-sized touchscreen displays, but you might still be able to get one if you act fast 'We have no interest in patient data in the UK': Palantir UK head defends record as criticisms rise Amazon’s new AI Bio Discovery tool can provide ‘every researcher’ with ‘lab-in-the-loop drug discovery’ – 40+ AI biology models can filter 300,000 novel antibody candidates down to the top results for testing in just weeks Over 100 Chrome Web Store extensions found stealing user data from thousands of accounts OpenAI reveals its Mythos rival designed for cybersecurity pros NYT Connections hints and answers for Tuesday, April 14 (game #1038) Forget Dr Doolittle, study finds animals might not only want to use tech, but they also want to talk to us with it… 'The decision is deeply troubling': Tesla gets a green light for Full Self-Driving in Europe — but not… OpenAI flags third-party data issue — all macOS users should update now Microsoft says Copilot is for ‘entertainment' not work, Meta’s Muse Spark and 7 other AI stories you… Man Utd vs Leeds Live Streams: How to watch Premier League 2025/26 from anywhere in the world, team news What is the release date for Invincible season 4 episode 7 on Prime Video? Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… The Lenovo Legion Go 2 handheld costs more than two Nvidia RTX 5080 GPUs — and that's genuinely absurd Secretlab is launching its first Diablo desk, with a design that 'traces the infernal history' of the series
AIs like ChatGPT fall apart in classic 'Stroop' psycholog...
Darren Allan · 2026-06-05 · via Latest from TechRadar in News
A robot standing thoughtfully in front of a giant digital display with code on it
(Image credit: Getty Images)

  • A new study tasked AIs with tackling the 'Stroop' test
  • GPT and Claude performed very poorly compared to humans
  • There are nuances here, but broadly, the researchers argue that improving this side of AIs is crucial for achieving artificial general intelligence

A freshly published study has pointed out a limitation of big-name AI models such as ChatGPT, albeit causing some controversy as the primary piece of research uses now outdated versions of those models – but there are nuances therein, and this doesn't make the findings irrelevant.

I'll go into that more shortly, but first, let's look at the study itself, which was highlighted on Reddit ('New study reveals top AI models completely fail the classic 'Stroop' psychological attention test') and published via the Oxford University Press in the journal PNAS Nexus.

The research consists of testing the so-called 'Stroop effect' with GPT-4o and Claude 3.5 Sonnet. As noted, these aren't the cutting-edge versions of those AIs (Large Language Models, or LLMs) – but they were at the time the initial study was carried out.

The Stroop effect refers to the phenomenon whereby the human brain gets confused when asked to name the color of the ink used to write a word, when that word can be the written version of another (incongruent) color in some cases. So, if the word 'red' is written in blue ink, that'll cause a slower response – or possibly a wrong response, where the viewer will accidentally say "red" rather than the actual color of the ink, which is blue.

This is because the brain is trying to juggle two different tasks – reading comprehension and color recognition – and so cognitive interference arises. Overriding the compulsion to read the word and say the color instead requires "executive control of attention," and this is what the authors were testing in the AI models. Both color-naming and word-reading were tested in shorter and longer lists of words (5, 10, 20, and 40 words).

The study observes: "Like humans, both LLMs [GPT and Claude] showed relatively high accuracy on the word-reading task and performed worse in the incongruent condition [where the word doesn't match the color] than in the congruent and neutral conditions for the color-naming task."

For color naming, humans maintain around 95% accuracy even in very long tests (up to an hour), but LLMs' accuracy declined very swiftly with longer word lists under the incongruent condition (mismatched color and word name). GPT-4o was 91% accurate in a five-word test, but dropped off to 57% with 10 words, and fell away completely to 22% with 20 words (and was only 15% accurate at 40 words).

Sign up for breaking news, reviews, opinion, top tech deals, and more.

Claude 3.5 Sonnet did better, staying 76% accurate at 20 words, but again fell hopelessly to 24% in the longest test of 40 words.

The authors conclude: "The significant degradation pattern of the two LLMs suggests fundamental limitations compared with human attention."


Analysis: another necessary step on the path to AGI?

An AI face in profile against a digital background.

(Image credit: Shutterstock / Ryzhi)

If you've scanned through the Reddit thread, you doubtless noticed that, as mentioned at the outset, there's a lot of flak fired at this study by commenters due to the usage of outdated models of GPT and Claude.

Indeed, these older LLMs are called "state of the art" at one point by the authors – and of course, as already noted, they were cutting-edge when the main study was conducted. Still, this is unfortunate phrasing that should've been updated and tweaked now that the paper has just been published (after peer review and so forth).

However, the researchers did conduct tests on GPT-5, Claude Opus 4.1, and Gemini 2.5 Pro in September 2025, although this is somewhat buried in the paper. That more recent testing found that these models offered only "slight" improvements on their predecessors, and that they still exhibited "ongoing executive attention deficiencies, consistent with our comprehensive analysis of earlier transformer models" (as did Gemini 2.5 Pro, which was a new introduction here).

Granted, a smaller sample size was used, but the researchers still argue that overall, their study reflects a fundamental limitation which is "inherent to the architectural constraints of transformer-based LLMs".

The authors note that a caveat is that GPT-5 in 'Thinking' mode can write and then run code to ensure it performs the Stroop test flawlessly – and similar functionality can be utilized by other LLMs – but this is essentially the AI (cleverly) fudging around its inadequacies. It isn't changing the way it works or reasons more broadly, of course.

The researchers note that transformer architecture innovations for LLMs are focused on enhancing memory capabilities, which fail to address the "core limitations of attention mechanisms, specifically the need for sophisticated alerting, orienting, and executive control networks to enable cognitive flexibility."

The ultimate aim is effective goal-directed behavior, and the study observes: "Future [LLM] development might benefit from implementing more sophisticated executive control systems that can handle decision conflicts through structured, goal-directed processing rather than relying solely on enhanced memory capabilities."

The authors argue that "incorporating executive control mechanisms akin to those in biological attention is crucial for achieving artificial general intelligence [AGI]."


Google logo on a black background next to text reading 'Click to follow TechRadar'

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.


An Apple MacBook Air against a white background

Darren is a freelancer writing news and features for TechRadar (and occasionally T3) across a broad range of computing topics including CPUs, GPUs, various other hardware, VPNs, antivirus and more. He has written about tech for the best part of three decades, and writes books in his spare time (his debut novel - 'I Know What You Did Last Supper' - was published by Hachette UK in 2013).