Issue 04 · 17 Jun 2026
scroll, baby ↓
✦
★ THIS WEEK'S #1 INSIGHT ★
GPT-5 CAN'T
READ COLORS.
A psychology test built in 1935 for first-graders just
annihilated the entire frontier of artificial intelligence.
GPT-5. Claude Opus 4.1. Gemini 2.5.
From 90%+ accuracy to near zero.
★ play along at home ★
Quick! Say the
COLOUR, not the word →
RED
BLUE
GREEN
YELLOW
PINK
ORANGE
PURPLE
CYAN
LIME
Hard? Yeah. That's the Stroop test.
Your brain's word-reader and colour-namer have a knife fight every time.
✷
★ the experiment ★
They handed it
to the BIG GUYS.
Claude Opus 4.1
~92% ✓
collapsed ✗
Gemini 2.5
~90% ✓
collapsed ✗
GPT-4o
~88% ✓
collapsed ✗
Five-year-olds: fine, actually.
★ the number ★
90%
→
~0%
accuracy, same model, same test, longer list.
The really weird bit?
It's not because they don't know the answer.
★ the part that broke my brain ★
Claude was given the test with no instructions.
It identified the Stroop paradigm,
explained the word-colour mapping rules, generated the correct answer table…
…then got 7 out of 10 wrong.
— Stroop AI study · Jan 2026, published June 2026
It knew the test. Explained the test. Failed the test anyway.
★ why you should care, ceo ★
Three takeaways
for the home loan biz.
01
"Knows the answer" ≠ "Does the answer."
If your AI workflow has a checking step, the AI also has to execute it, not just describe it. Big difference.
02
Long lists kill accuracy.
50-line broker checklists or document bundles? Break them into chunks. The cliff at scale is real, even for GPT-5.
03
Audit your prompts like a regulator.
If the model can articulate the rule and still violate it, "well-written prompts" are not a control. Add structured validation.
★ the bigger story ★
These models can
pass the BAR.
Just not the
PRESCHOOL ATTENTION TEST.
Everyone's racing benchmarks.
Almost no one is measuring whether the model actually does what it says it'll do.
That gap is where every AI product currently lives or dies.
★ that's the issue ★
Keep questioning
your robots, Otto.
Cheers,
Steve
CEO & Co-Founder · Hyro