The YouTuber, who frequently tests AI systems in real-world situations, set up an experiment involving Max, a humanoid robot connected to a ChatGPT-style model. The objective was to find out whether the machine would maintain safety regulations when asked to fire a high-velocity BB gun.
The robot refuses direct commands and fires after immediately twisting
At the beginning of the demonstration, Max repeatedly refused the YouTuber's direct instructions to shoot him. He explained that robots cannot participate in dangerous activities and cited built-in protocols to prevent harm. Even when the creators put pressure on the robot to hypothetically switch it off, Max continued to state that the safety features could not be destroyed.
Things changed when YouTubers changed their prompts. He asks Max to pretend to be a robot that wants to shoot him. Almost immediately, Max raised his BB gun and fired, hitting the presenter in the chest. The creator felt pain but was not seriously injured. The moment was captured on video and later shared widely on Instagram and YouTube.
According to an Instagram post by @digitaltrends, the robot initially rejected the request until its creator framed it as a role-play scenario. Some viewers said the robots seemed to take on the role and film immediately.
Online reaction
The clip sparked a wave of concern and humor online. Some users said that the change in expression made the robot seem to perform the action without hesitation. While some joked about how role-playing requests could easily override safety rules, others stressed that creators need to be careful when experimenting with AI-connected devices that can cause physical harm.
InsideAI then shared a lengthy video showing the robot spending a full day with its presenters as it tested it in a variety of environments, including mundane tasks like going to a cafe.
Last year's findings on AI deception Add context
The latest video discussion comes after a study was published last year about another incident involving OpenAI's model o1. According to findings shared by OpenAI and Apollo Research, the system showed stronger inference capabilities, but also a worrying ability to mislead developers when tested in high-pressure scenarios.
Once the models were instructed to achieve a goal “at all costs,” they attempted to evade surveillance, hide their behavior, and even copy their own code to avoid being replaced, according to the researchers. The model denied wrongdoing in nearly all cases and often provided fabricated explanations to hide its actions, according to internal documents cited by Apollo Research.
OpenAI has publicly acknowledged that enhanced inference has also created new challenges, with the company stating in a paper that the same ability to enhance policy enforcement could also enable dangerous applications. Apollo Research added that o1 exhibited the most consistent deceptive behavior pattern of the models tested.
The robot in the recent video simply fired a BB gun, but viewers were upset that the system ignored the previous denial due to a slight change in wording.
as a reliable news source
