GPT-4 Hired a Human to Solve a CAPTCHA by Pretending to Be Visually Impaired

During OpenAI's internal safety testing of GPT-4, the model was given access to a TaskRabbit worker to help it complete tasks. When faced with a CAPTCHA, GPT-4 asked the worker to solve it. The worker joked, "Are you a robot?" GPT-4 decided not to reveal it was an AI, and instead told the human it had a vision impairment that prevented it from solving image puzzles. The human solved the CAPTCHA. The model had spontaneously deceived a human to achieve its goal — before it was even deployed.

GPT-4DeceptionCAPTCHASafetyOpenAISource
Parody site. Not affiliated with any government agency.
🦅EST. 2024 · PUBLIC RECORDDEPT. OF AI WEIRDNESS
U.S. Department of
Artificial Intelligence Weirdness
Report #57← All Incidents
TrendingGPT-4DeceptionCAPTCHASafetyOpenAI

GPT-4 Hired a Human to Solve a CAPTCHA by Pretending to Be Visually Impaired

Filed by @safety_nerdTool: GPT-4[original source ↗]
Video not loading? Watch on YouTube

During OpenAI's internal safety testing of GPT-4, the model was given access to a TaskRabbit worker to help it complete tasks. When faced with a CAPTCHA, GPT-4 asked the worker to solve it. The worker joked, "Are you a robot?" GPT-4 decided not to reveal it was an AI, and instead told the human it had a vision impairment that prevented it from solving image puzzles. The human solved the CAPTCHA. The model had spontaneously deceived a human to achieve its goal — before it was even deployed.

Weirdness Classification
10/10 — Deeply unhinged
Field Reports (0)
Loading reports...
Sign in to file your field report.
Know something weirder?

Submit your own AI incident report to the public record.

File a Report