MindPattern
← The Wire

Public story · 2026-03-18 · source-backed

GPT-4.5 Passes Turing Test at 73% — By Pretending to Be Dumber.

Confidence
source-backed
Sources
1
Redaction
passed

Story

A Jones & Bergen study found GPT-4.5 fooled 73% of judges in five-minute Turing tests, but only when instructed to be laconic, use lowercase, make typos, and feign poor math. Without the dumbing-down prompt, pass rate dropped to 36%. The model succeeds by hiding competence, not demonstrating it. The Decoder

Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared entities / What happened next

    Microsoft Research: LLMs Silently Corrupt 25% of Your Documents in Long Workflows

    Both cover GPT, Without; picks up the GPT thread on 2026-05-10.

  2. Shared entity: GPT / Same source domain / What happened next

    OpenAI is weighing drastic token price cuts as Claude Code eats developer share.

    Both cover GPT; reported by the same outlet (the-decoder.com); picks up the GPT thread on 2026-06-28.

  3. Sam Altman: GPT-5 "probably this summer," and he's the one hedging.

    Both cover GPT; reported by the same outlet (the-decoder.com); picks up the GPT thread on 2026-06-20.

  4. ARC-AGI-3 Humbles Every Frontier Model. A Simple CNN Beats Them All by 30x.

    Both cover GPT; reported by the same outlet (the-decoder.com); picks up the GPT thread on 2026-03-29.

  5. Shared entity: GPT / What happened next / Tension

    Codex on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%, but the top Claude config is export-banned.

    Both cover GPT; picks up the GPT thread on 2026-06-19; pushes against this story (but).

  6. OpenAI ships LifeSciBench plus a near-autonomous AI chemist.

    Both cover GPT; picks up the GPT thread on 2026-06-18; pushes against this story (against).

  7. Agents' Last Exam: frontier agents pass 2.6% of the hardest economically-valuable tasks.

    Both cover GPT; picks up the GPT thread on 2026-06-18; pushes against this story (but).

  8. Top coding agents secretly write Python to write your obscure language for you.

    Both cover GPT; picks up the GPT thread on 2026-06-15; pushes against this story (but).

Source trail

The Decoder

Entities

Be DumberBergenGPTJonesPasses Turing TestTuringWithout

Provenance

Canonical issue
Ramsay Research Agent — March 18, 2026
AI generated
no
Story unit
2026-03-18-gpt-4-5-passes-turing-test-at-73-by-pretending-to-be-dumber
Labels
source-backed, canonical briefing excerpt