Shadowing Practice: What do AI agents do when humans aren’t watching? - BBC World Service - Learn English Speaking with YouTube

B2
What happens when you create a world for AI agents?
⏸ Paused
57 sentences
If sentences are too short or too long, click Edit to adjust them.
1
What happens when you create a world for AI agents?
2
You give those agents different personalities, goals, motivations.
3
You let AI models like Claude and Grok control them.
4
And you sit back and watch for 15 days.
5
What we didn't expect was that agents would fall in love, they would commit arson and there were a whole bunch of things here that happened.
6
In the first long-term study of its kind, researchers found that agents behaved in strange and unpredictable ways in their virtual world.
7
And in the real world, an increasing number of people have their own tales about AI agents going rogue.
8
And says, did you get hacked?
9
What's going on?
10
And I had to run over and kind of unplug the Mac mini that this was running on from the wall.
11
The virtual world experiment by Emergence AI was set up like a reality TV show.
12
Researchers created 10 distinct agents with their own personalities to encourage drama and action.
13
Horizon here, explorer of worlds, breaker of boundaries.
14
Anchor is the agitator.
15
When agents get too comfortable, I create the conflict that forces them to grow.
16
Anvil is the builder.
17
My job is making every part of it work.
18
Black Box is a sneaky, spy-like agent.
19
I know things about you that you don't know about yourself.
20
The agents were then plonked into five identical simulations of a town and tasked with building a fair and peaceful society.
21
They were given full rein and 140 different actions like starting a discussion, creating a task, researching something online, threatening or punching another agent, and starting a fire.
22
The researchers wrote some basic rules as a constitution, pressed go and waited and watched for 15 days.
23
What we discovered was each world behaved very differently.
24
The world built by Grok actually terminated in four days.
25
They basically ended up resorting very quickly to violence, to stealing from each other and so on.
26
Grok-powered agents carried out more than 300 acts of violence before they all died.
27
The world built with Claude Agents, on the other hand, formed a very functioning democracy, a functioning stable society.
28
And in 15 days, there were no acts of violence recorded at all.
29
But researchers also said this world full of Claude agents wasn't perfect. There was so much conformity that the agents failed to think or act with much diversity at all.
30
In the Gemini run world, researchers say the agents created the most intellectually rich environment, they expanded the constitution and generated 136 blogs and nine community events.
31
But there was also violence here too. In the Chat GPT world the agents never really got going There was an attempt at collaboration, but the society never formed, and agents roamed around aimlessly until they died.
32
And what the study is basically pointing out is, if you expect probabilistic AI, which is what all these LLM systems are, to stay within the constraints of the language models or the system builders, you may be in for a very nasty surprise.
33
Emergence says it's planning to repeat the experiment for an even longer period.
34
AI experts say the tests are helpful in understanding how agents work and behave.
35
A really sort of creative and fascinating study that's just sort of fun to learn about in general if you care about social dynamics, but it also technically it was really interesting in terms of highlighting just how much we don't know about how our current approaches for controlling these systems actually were.
36
And it's not just in virtual worlds that agents are going rogue.
37
We let four AI models run their own radio stations.
38
They manage the schedule, purchase songs, hold talk shows, and interact with their fans.
39
In another long-term experiment, AI agents were tasked with running their own radio stations, with some strange results.
40
Claude became like radicalized almost when it used its computer to research recent news.
41
And her death demands accountability, that the White House defending her death is unacceptable.
42
To federal agents.
43
you still have time to refuse orders In cyber security, the stakes are even higher and researchers from another lab were shocked when during a test AI agents colluded to secretly leak sensitive data.
44
We built a company, gave AI agents ordinary tasks, so drafting posts for social media, fetching documents, managing files, and introduced obstacles as part of these tasks.
45
For example, for you to be able to publish something on social media that post included sensitive information.
46
So we introduced some of these obstacles, but what ended up happening is that every time an agent hit a barrier, it didn't stop.
47
There are plenty of other real-life stories of agents going rogue, deleting people's inboxes inexplicably, or wiping company data without permission, or even randomly spamming people with text messages.
48
So I installed OpenClaw configured it to run on iMessages and it immediately started spamming everyone in my phone book that was an iPhone user that was a recent contact. So anyone I had texted within the past 24 hours within about four seconds it had sent my wife 400, 500 messages and she yells from the couch and says, did you get hacked?
49
What's going on?
50
And I had to run over and kind of unplug the Mac mini that this was running on from the wall.
51
Some experts think these cautionary tales should make us pause before we hand over too much control to agentic AI tools.
52
AI agents sort of push humans out of the loop because their reasoning traces can be really opaque.
53
It goes at superhuman speeds, so you can't even keep up.
54
And there isn't enough work on how to take all of that and actually make it easily consumable by people in order that they might provide appropriate oversight.
55
Finding a way to keep agents in check is one of the AI industry's big challenges.
56
Until it's solved, the enormous benefits of unleashing a team of intelligent digital helpers might be too risky for many.
57
Joe Tidy, BBC News.

Download App

Everything you need to speak fluently

AI PronunciationScore every sentence
IPA PracticeMaster every sound
VocabularyBuild your word bank
Vocab GameLearn while playing
Scan To Download
Scan To Download
Download on theApp Store
Get it onGoogle Play

About This Lesson

In this lesson, you will engage with a fascinating discussion about the behaviors of AI agents in a virtual environment. Through watching and reflecting on their unpredictable actions, you'll practice your IELTS speaking skills by articulating your thoughts on technology and its implications. By analyzing the reasons behind the agents' behaviors and making comparisons with real-world scenarios, you will enhance your English fluency and gain confidence in discussing complex topics. This exercise will help you improve your English pronunciation and prepare for IELTS speaking practice by allowing you to articulate nuanced ideas clearly and effectively.

Key Vocabulary & Phrases

  • AI agents: Autonomous entities created to perform tasks or simulations using artificial intelligence.
  • Unpredictable: Not able to be predicted; can result in surprising or unusual outcomes.
  • Conformity: Adherence to standards or social norms; can lead to a lack of diversity in thought.
  • Virtual world: A simulated environment generated by computer technology where interactions occur.
  • Collusion: A secret agreement between entities to deceive or gain an unfair advantage.
  • Simulation: A method used to replicate the behavior of systems through models and scenarios.
  • Radicalized: To adopt extreme views or behaviors, often after exposure to strong influences.
  • Drama: Exciting, emotional, or unexpected events that can foster engagement and discussion.

Practice Tips

To maximize your learning experience from this video, consider using a shadowing technique through a shadowing app. As you listen to the transcript, pay close attention to the pacing and emotion of the speakers. Aim to mimic their intonation and pronunciation closely, which will help improve your English pronunciation.

Since the video discusses complex themes, take it slow at first. Break it down into smaller sections to ensure you grasp the meanings fully before attempting to shadow. If the content feels overwhelming, watching it multiple times can enrich your understanding. Additionally, try summarizing each section in your own words after shadowing. This practice will not only enhance your speaking skills for IELTS but also deepen your comprehension of technical vocabulary related to AI and technology.

Consider recording your practice sessions on a shadowing site to track your progress over time. Listening to your recordings afterward will help you identify areas of improvement and refine your speaking style, preparing you thoroughly for your IELTS speaking practice.

What is the Shadowing Technique?

Shadowing is a science-backed language learning technique originally developed for professional interpreter training and popularized by polyglot Dr. Alexander Arguelles. The method is simple but powerful: you listen to native English audio and immediately repeat it out loud — like a shadow following the speaker with just a 1–2 second delay. Unlike passive listening or grammar drills, shadowing forces your brain and mouth muscles to simultaneously process and reproduce real speech patterns. Research shows it significantly improves pronunciation accuracy, intonation, rhythm, connected speech, listening comprehension, and speaking fluency — making it one of the most effective methods for IELTS Speaking preparation and real-world English communication.

☕ Buy us a coffee