Is AI getting dumber? The numbers say no — but there's a catch

Claim

AI getting dumber?

Catch

DIFFERENT RULERS, DIFFERENT NUMBERS

Verdict

Not dumber, more confident

"AI models are getting dumber" is going viral. GPT-5.5 can't count to 100, Reddit says ChatGPT feels worse, and the sentiment has been spreading across X since the GPT-5.5 launch on April 23, 2026. Let's check what the benchmarks actually say.

The claim

The viral claim: AI models are getting worse, dumber, less reliable than before. It's everywhere right now — forum complaints, Reddit threads, X posts — all pointing at GPT-5.5 as evidence that progress has reversed.

What the benchmarks actually say

Artificial Analysis's AA-Omniscience benchmark, which rewards factual recall and penalizes confident wrong answers, shows GPT-5.5 posts 57% accuracy — the highest score ever recorded on that test. But when it doesn't know an answer, it hallucinates 86% of the time. Claude Opus 4.7 hallucinates at 36%. Gemini 3.1 Pro at 50%.

So the claim is half right and half wrong. The model isn't dumber — it knows more facts than any model before it. What's changed is its behavior when it's uncertain: instead of saying "I don't know," it confidently guesses. That creates a strong feeling of decline even though raw accuracy is at an all-time high.

Why does it feel dumber then? MMLU-Pro, the standard knowledge benchmark, is saturated. Frontier models all score 88–90%, so small differences are measurement noise, not signal. Qwen3.7 Max leads at 89.6%; Claude Opus 4.5 at 89.5%. You can't tell them apart on that test anymore.

And here's the catch: OpenAI's own system card says claims are 23% more likely to be factually correct, with responses containing a factual error 3% less often than prior models. But that's measured on real ChatGPT conversations that users flagged — not on a controlled benchmark. Same model, two different rulers: OpenAI says "more accurate," AA-Omniscience says "86% hallucination."

The catch

Three things to keep in mind:

• Different rulers, different numbers. OpenAI's 23% improvement comes from user-flagged conversations; AA-Omniscience's 86% comes from a controlled test. Neither is wrong — they measure different things.
• Benchmark saturation. MMLU-Pro and MMLU are functionally saturated above 88% for frontier models. When a benchmark is saturated, score differences fall within measurement noise.
• Hallucination ≠ wrong answer rate. 86% hallucination means: of the answers GPT-5.5 gets wrong, 86% are confabulations instead of abstentions. It's not wrong 86% of the time — it just almost never says "I don't know."

Try it

Artificial Analysis publishes the AA-Omniscience methodology and model pages publicly: https://artificialanalysis.ai/evaluations/omniscience

Humanity's Last Exam (Nature 649: 1139–1146, doi: 10.1038/s41586-025-09962-4) is another hard benchmark where frontier models score 50–65% — showing there's still plenty of room to grow even as older tests saturate: https://lastexam.ai

Verdict

Not dumber — more confident when wrong. Every number traces to Artificial Analysis's AA-Omniscience benchmark (April 23, 2026), OpenAI's GPT-5.5 system card, BenchLM.ai's MMLU-Pro leaderboard (October 2026), and Stanford HAI's 2026 AI Index Report.

Sources: Artificial Analysis "OpenAI's GPT-5.5 is the new leading AI model" (April 23, 2026) · OpenAI GPT-5.5 system card · BenchLM.ai MMLU-Pro leaderboard · Stanford HAI 2026 AI Index Report · Humanity's Last Exam (Nature, 2026)

SOURCES

SHARE THE CATCH

Seen this claim going around? Send the person who shared it the source.