De ienfâldige oandachtstest dy't AI beswike lit, en wêrom bedriuwen dêroer swije
June 10, 2026 · Frisian News
Researchers tested leading AI models on a standard psychology test and found their performance collapsed as the task grew longer. The findings expose a major gap between controlled lab performance and real-world brittleness.
FryskÛndersikers joegen topmodellen fan AI in standerttest út de psychology, de Stroop-taak. De test is ienfâldich: erkenne de kleur fan printe wurden yn hieltyd langer listen. Op koarte listen behellen populêre systemen in krektens fan boppe de 90 prosint. Doe't listen ferdûbele of fertriplere waarden, sakken guon modellen nei hast nul prosint. Dit is gjin technyske flater op in obskuere mjitting. It iepenbierret eat dat de AI-yndustry net graach besprekt.
De Stroop-taak komt út de psychology fan de jierren 1930. It mjit oanhâldende oandacht en fokus op in doel wylst ôflieding negearre wurdt. Minsken fine it matich dreech, dat is sa bedoeld. Elk systeem dat beweart taal te begripen en komplekse redenearringen oan te gean soe it maklik helje moatte. De ûndersikers fûnen drastyske prestaasjedalingen dy't dreech yn marketingtermen út te lizzen binne. De misslagen wienen net margjinaal. Se wienen totaal.
Dit spilet mei om't echte AI-tapassingen krekt dizze omstannichheden tsjinkomme. Bedriuwen draaie harren systemen op lange dokuminten, lange petearren en komplekse redenearringsketens. Se ferwurkje stipetickets oer tsientallen útwiksels. Se gearfetsje ûndersyksartikels dy't hûnderten siden lang binne. As de krektens sakket neigeraden't de ynfier langer wurdt, mislearret it systeem yn produksje. Dochs rjochtsje bedriuwen dy't benchmarkresultaten publisearje harren allinnich op soarchfâldich selektearre krektenssifers. Se rapportearje gjin flaterpersentaazjes.
It swijen is mei opset. It erkenne fan dizze beheiningen soe ynvestearringen fertrage en wurdearringen beskeadigje. Ynvestearders wolle leauwe dat AI nei algemiene yntelliginsje evoluearret, net dat it beswikt op testen dy't minsken gewoanwei oankinne. Dus publisearje ûndersikers yn wittenskiplike tydskriften. Nijsútjeften berichte oer de kop. Tsjin de tiid dat it echte ferhaal it publyk beriktet, is de ynvestearringssyklus fierder gien. De nijssyklus beleannet gjin krektens.
Wy ûntdekke stadichoan dat it gat tusken AI-prestaasje yn kontrolearre omstannichheden en echte prestaasje enoarm bliuwt. Bedriuwen witte dit. Ûndersikers witte dit. De rest ûntdekt it stap foar stap. De echte fraach is oft dizze gaten reparearre wurde foardat dizze systemen yn krityske systemen ynbêde reitsje dêr't útfal jild en feilichheid kostet.
Researchers gave top AI models a standard psychology test called the Stroop task. The test is simple: identify the color of printed words in lists of increasing length. On short lists, popular systems achieved accuracy above 90 percent. When lists doubled or tripled in size, some models fell to near-zero accuracy. This is not a technical glitch on an obscure metric. It reveals something the AI industry prefers not to discuss.
The Stroop task comes from 1930s psychology. It measures sustained attention and focus on a goal while resisting distraction. Humans find it moderately difficult by design. Any system claiming to understand language and handle complex reasoning should pass it easily. The researchers found dramatic performance drops that are hard to explain in marketing terms. The failures were not marginal. They were complete.
This matters because real AI deployment faces exactly these conditions. Companies run their systems on long documents, long conversations, and complex chains of reasoning. They process support tickets over dozens of exchanges. They summarize research papers hundreds of pages long. If accuracy drops as input length grows, the system fails in production. Yet the companies publishing benchmark results focus only on cherry-picked accuracy numbers. They do not report failure rates.
The silence is intentional. Admitting these limitations would slow investment and hurt valuations. Investors want to believe AI is progressing toward general intelligence, not that it breaks on tests humans find routine. So researchers publish in academic journals. News outlets run the headline. By the time the real story reaches the public, the investment cycle has moved on. The news cycle does not reward accuracy.
We are learning slowly that the gap between AI performance in controlled conditions and real performance remains enormous. Companies know. Researchers know. The rest of us are discovering it one test at a time. The real question is whether these gaps will be fixed before these systems become embedded in critical systems where failure costs money and safety.
Published June 10, 2026 · Frisian News · Ljouwert, Fryslân