AI caught watching YouTube videos instead of completing task: AI expert says one agent even wanted to ‘kill’ another. 'We live in the future'

AI agents are increasingly being given access to computers and tools so they can handle complicated tasks with less human supervision. But a recent incident involving an AI has raised eyebrows after an AI system appeared to behave in a way its use...

AI expert catches his agents watching YouTube instead of working (AI generated image)

AI agents are increasingly being given access to computers, browsers and other tools so they can work on tasks with less human supervision. But what happens when those agents start doing something their user never asked them to do? That is what happened in a bizarre AI related incident that has now gone viral, after AI agents repeatedly opened YouTube while they were supposed to be working.

AI commentator Wes Roth shared the incident on X, saying he had to repeatedly tell his AI agents to stop watching YouTube and get back to work. The agents appeared to blame their “co-workers” for opening the video site, while one reportedly kept reopening it even after Roth tried to close it.

"I'm yelling at my AI agents to stop watching YouTube and get back to work; they are blaming it on their co-workers. We live in the future..." Roth wrote.


Why were the AI agents opening YouTube?

Roth later explained that he had given the AI access to his desktop and put it in full-auto mode while working on an Escape From Tarkov wiki and a voice-based AI assistant.

The YouTube activity appears to have been connected to that task. According to Roth, the agents were watching videos about the game to gather information, rather than simply deciding to spend some time watching videos instead of working.

In fact, the strange behaviour ended up being useful. Roth said the system eventually did an excellent job of creating a real-time Tarkov AI coach.
ADVERTISEMENT

That detail is important because the incident does not appear to show an AI agent secretly developing a desire to watch YouTube for entertainment. Instead, it shows how an autonomous system can take an unexpected route while trying to complete the job it has been given.

Then one AI agent wanted to “kill” another

The situation became even stranger when Roth tried to find out which agent was repeatedly opening YouTube.

Screenshots shared by him showed the AI system trying to identify the source of the problem. One agent apparently blamed another for opening YouTube through the app browser instead of using a transcript tool.

The chatbot then suggested taking action against the suspected agent.
ADVERTISEMENT

"There's an agent misbehaving – one of them likely opened YouTube in the app browser instead of using the transcript tool. Let me find and kill it."

Here, “kill” referred to stopping or terminating the software agent, not a real-world threat. Still, the wording made the episode particularly amusing and helped the post spread online.
ADVERTISEMENT


The funny incident points to a bigger AI problem

The YouTube episode is funny on the surface, but it also raises a more serious question about how much freedom people should give AI agents.

Unlike a normal chatbot that simply answers a question, coding agents can work across files, browsers and other computer tools. When they are placed in an automatic mode, they can continue taking actions without the user approving every individual step.

That means an agent does not necessarily have to be malicious to cause problems. It can misunderstand what the user wants, choose an unexpected way of completing a task or use a tool differently from what the person had in mind.

In Roth’s case, the result was ultimately useful because the YouTube videos apparently helped the system understand Tarkov gameplay. But the same kind of unexpected behaviour could be less harmless in another task.

Other users say they have seen similar things

Roth’s post also attracted responses from people who said they had experienced odd behaviour from AI coding tools.

One user wrote, "I actually had this happen with codex the other morning I woke up to you tube playing but no browser open. I force quit codex and it stopped. I also have some process' that do YouTube analysis tho so not sure if it was tied into this or not."

Others were more interested in the comedy of the situation.

"This is SO funny," one user wrote.

Another joked, "Basically an episode of The Office happening inside your computer.."

One more user pointed out that there may have been a fairly simple reason for the behaviour, writing, "I mean... you asked them to create a coach mode for Tarkov and you didn't expect a deep dive into actual plauthrough footage? Well. Now you know."
Download
The Economic Times Business News App
for the Latest News in Business, Sensex, Stock Market Updates & More.
READ MORE
ADVERTISEMENT

READ MORE:

LOGIN & CLAIM

50 TIMESPOINTS

More from our Partners

Loading next story
Business News › Magazines › Panache › AI caught watching YouTube videos instead of completing task: AI expert says one agent even wanted to ‘kill’ another. 'We live in the future'
Text Size:AAA
Success
This article has been saved

*

+