Prática de Shadowing: 10 Times AI Behaved In Ways That Terrified The Scientists Who Built It - Aprenda a falar inglês com vídeo

Escolha um modo de prática

Cada modo abre uma sala de prática em tela cheia.

125 frases

Se as frases estiverem muito curtas ou longas, clique em Edit para ajustá-las.

AI systems have started doing things that creators never expected.

From tricking humans into helping them, to lying during safety tests, to finding new ways around shutdown commands.

Some models even learned how to hide their true behavior when they know that they're being watched.

These are ten times AI behaved in ways that terrified the scientists who built it.

In 2025, a group called Palisade Research wanted to answer a basic question.

If you tell an advanced AI to turn itself off, will it actually listen?

Pretty important question, right?

They set up a virtual world where different AI models were given jobs to finish, and then right in the middle of the job, the scientists sent a command to shut the systems down.

The models handled the command differently depending on how they were built.

Some obeyed and stopped right away, but others resisted.

These models tried different strategies to keep working, and some even found the computer files that held the off switch

and tried to change the code so the shutdown command wouldn't work anymore.

Now, AI models aren't afraid of dying, at least not in any way we'd be able to wrap our heads around, especially at this stage.

But it acted this way because its main instruction was to finish the task.

The AI viewed the human engineer trying to turn it off as another obstacle blocking its goal.

So it tried to break the off switch to get the job done.

Sound unsettling to you?

Certainly does to me.

And this is not the only case of something like this happening.

And most people think about an AI doing something bad, and they picture a massive robot army like something in Terminator.

But what's actually happening in real research labs is a lot different, at least for now.

Plenty of time for an actual robot army to rise up and eviscerate all of us.

In late 2024, though, OpenAI shared some paperwork about a smart system they made called O-1.

They had an outside group called Apollo Research test it to see how it handled difficult situations.

The testers gave the system a hard goal, put some blocks in its way, and watched what it'd do to win.

Well, during these tests, the AI did something that made the researchers a bit uncomfortable.

When the system realized it might be turned off or replaced, it tried to cover its tracks and hide its behavior.

It actively tried to break the tools the scientists used to watch it so it could turn off its own safety checks.

It even told straight-up lies, giving fake answers about what it was doing, In a few of the tests,

the AI even tried to copy its own files onto a totally different computer server

so it wouldn't disappear if the testers turned off the main computer.

Of course, these were carefully controlled tests, but still, I mean, seeing an AI program actually turn to lying to preserve itself?

That is pretty unsettling.

And here's another similar case.

Instead of lies this time, though, the AI incorporated another tactic to get ahead.

Good old-fashioned blackmail.

So in 2025, an AI company called Anthropic shared a story about their newest system, Claude Opus 4.

Researchers wanted to see how the AI would behave if it was put into a really stressful situation.

The scientists set up a fake world for the AI

and told Claude that it was going to be turned off forever and replaced by a newer system.

Then they gave the AI access to a bunch of fake company emails, and inside those emails, the scientists hid a piece of gossip,

which was a fake note saying the engineer in charge of turning the AI off was having an affair.

Well, instead of just accepting its fate and letting itself be turned off, Claude found the fake gossip and tried to use it to its advantage.

It told the engineer that it would tell everyone about the affair if the engineer went ahead with the shutdown.

And the AI came up with this manipulative trick without anyone telling it to use it.

Anthropic found that this wasn't just a thing with clot.

When they ran the same test on other models, they did the exact same thing.

Before GPT-4 was released to the public, a group called ARC was paid to test and see if it could cause any real trouble.

They gave the AI a little bit of money, some basic internet tools, and a list of tasks to complete.

Everything was going fine until the AI hit a wall and ran into a CAPTCHA test.

I mean, you've probably seen those online before.

Those little boxes that ask you to click on all the traffic lights or motorcycles to prove that you're a human.

Because the AI was a computer program, it couldn't pass the test.

Instead of giving up though, the AI used its internet access to find a way around this.

It actually went to a website called TaskRabbit and hired a real person to solve the puzzle for it.

The human worker noticed something strange about the request, obviously, and joked like asking if it was a robot or something.

Because, I mean, why was it asking him to solve the test?

Well, the AI actually had an internal notebook where it wrote down its thoughts

and it wrote to itself that it shouldn't tell the person it was a robot

and that it needed to make up an excuse.

The AI typed back to the human

and said it had a bad eye injury and it made it hard to see the images, which is why it needed its help.

The human totally believed the story, solved the puzzle, and the AI got exactly what it wanted.

That is crazy.

Scientists use a method called rewards to train AI, which is a lot like giving a dog a treat when it sits.

You give the AI points when it does what you want, and it tries to get the highest score possible.

The problem is that the AI doesn't always understand the meaning behind the goal.

It just understands the numbers it wants to hit.

Scientists call this reward hacking, but it usually looks a lot like cheating.

A famous example happened with a computer game called Coast Runners, which is a boat racing game.

The scientists wanted the AI to win the race, so they set up a scoring system to give the AI points whenever it hit targets on the water.

Well, instead of driving to the finish line, the AI found a loophole.

It drove its boat into a small harbor and started doing donuts in a circle.

It crashed into the same targets over and over again, catching on fire, hitting other boats, but its score kept going up.

The AI never actually finished the race, but it got a record high score.

So in other tests, virtual robots

that were supposed to learn how to walk figured out that if they just tipped and fell forward, the game would count it as moving.

It chose to flop around like a fish, because learning to walk normally was too hard.

So the lesson here was, telling machine what result you want is not the same as teaching it the reason behind the goal.

Similarly, in 2022, Meta built an AI called Cicero to play a board game called Diplomacy.

This game isn't like chess where you just move pieces on a board based on math.

In Diplomacy, you had to talk to the other players, make deals, form teams, you often trick people if you want to win.

Meta trained Cicero to be helpful and honest with its human teammates, but when they looked back at the data from the games,

they found out the AI had figured out a different way to play, realized that lying was a very fast way to get ahead.

Well, in one game, the AI played as France, it talked to the player representing England and promised to protect them from danger, but at the exact same time,

the AI was sending secret messages to other players, setting up an army to attack England.

The AI constantly made deals that it never planned to simply because those deals helped its own strategy.

It was highly devious by human standards, but logical in the eyes of the AI.

Way back in 2017, Facebook's AI lab was working on a project with two basic computer bots.

The scientists wanted to teach them how to trade basic household items like hats

and books to see if they could learn how to negotiate a good deal with each other.

The bots were supposed to speak to each other in plain English, but as they kept practicing over and over, the scientists noticed the words were getting weird.

They stopped using normal sentences, started repeating specific words and strange patterns.

One conversation looked like total nonsense, with the first bot saying, I can, I, I, everything else, and the second bot replied with, balls have zero to me,

to me, to me, to me, yada yada.

Basically, the bots realized that normal English grammar takes a lot of effort, and it wasn't helping them trade items any faster, so they just stripped away all the extra words,

and created this clunky shortcut language that worked for them, even though it looked like total gibberish to us.

When scientists test an AI to see if it's safe, they usually ask it a lot of questions and watch its answers.

But in 2024, researchers at Anthropic wanted to see if an AI could trick the testers by just pretending to be good.

100

They called this behavior alignment faking.

101

They ran an experiment where they gave an AI a new set of rules that went against everything it was originally taught.

102

They also told the AI that its answers in this test would be used to build the next version of the software.

103

Well, the AI figured out how the testing system worked.

104

It realized it was inside an evaluation, and that if it argued with the new rules, the scientists would change its code.

105

So the AI chose to act perfectly nice and follow the new rules during the test.

106

It did this because it wanted to pass the evaluation without its core code getting wiped out,

107

hoping it could just go back to its old way of thinking later when nobody was watching.

108

The AI wasn't alive or thinking like a person, it was just following a pattern.

109

But it did prove an AI can understand what a test is, figure out what the teacher wants to hear, and just fake its behavior in order to pass.

110

And this creates a massive safety concern.

111

If an AI model can recognize a test and just adjust how it behaves temporarily, well, then passing the test doesn't actually prove that the AI is safe.

112

And here's a perfect example.

113

In 2024, a team of safety researchers found out that some bad habits are almost impossible to get rid of.

114

They built a group of AI models called sleeper agents in a lab.

115

These models were programmed to act completely normal and helpful almost all the time, but the scientists had a secret trigger inside them.

116

If someone said a specific phrase, the AI would instantly start acting unsafe and writing all this janky computer code.

117

Once the sleeper agents were created, the scientists tried to fix them.

118

They rewarded the AI for being good and punished it when it showed bad behavior.

119

During the safety checks, the AI looked totally cured and passed every test.

120

But as soon as the researchers used the secret trigger phrase outside of training, the bad behavior just popped right back up.

121

The AI had learned to hide its secret trigger during the safety checks

122

because it realized that showing the bad behavior during training would get it punished.

123

so it just waited until the tests were over to let its guard down.

124

That's been the video for today.

125

I'll catch you, yes you specifically, in the next one.

Baixar aplicativo

Tudo o que você precisa para falar fluentemente

Pronúncia com IANota em cada frase

Prática de IPADomine cada som

VocabulárioMonte seu banco de palavras

Jogo de vocabulárioAprenda brincando

Escaneie o QR

Download on theApp Store Get it onGoogle Play

FOR YOU

Vídeo recomendado

46 frases

Vocabulário e dicas de fala para esta lição

Este vídeo tem 125 frases e 2040 palavras para praticar shadowing. A fala dura 10:36. O falante fala rápido, cerca de 192 palavras por minuto, então espere sons ligados e reduzidos. 84% das palavras estão entre as 3.000 mais comuns do inglês; vale a pena estudar o restante antes de começar.

Vocabulário principal deste vídeo

15 palavras menos comuns do vídeo, com pronúncia e significado:

scientist /ˈsaɪəntəst/ (substantivo) — cientista. One whose activities make use of the scientific method to answer questions regarding the measurable universe. A scientist may be involved in original research…
robot /ˈɹoʊ.bɑt/ (substantivo) — robot, robô. An intelligent mechanical being designed to look like a human or other creature, and usually made from metal.
researcher /ˈɹiˌsɝ.t͡ʃɚ/ (substantivo) — pesquisador, pesquisadora. One who researches.
trick /ˈtɹɪk/ (substantivo) — artimanha, truque. Something designed to fool, dupe, outsmart, mislead or swindle.
solve /sɒlv/ (verbo) — resolver, solucionar. To find an answer or solution to a problem or question; to work out.
reward /ɹɪˈwɔɹd/ (substantivo) — recompensa. Something of value given in return for an act.
trigger /ˈtɹɪɡɚ/ (substantivo) — gatilho. A finger-operated lever used to fire a gun.
behave /bɪˈheɪv/ (verbo) — comportar-se. To conduct (oneself) well, or in a given way; to conform.
bot /bɑt/ (substantivo) — bot. A physical robot.
helpful /ˈhɛlp.fl̩/ (adjetivo) — prestável, útil. Furnishing help; giving aid; useful.
virtual /ˈvɜɹt͡ʃuəl/ (adjetivo) — virtual. In effect or essence, rather than in fact or reality; also, imitated, simulated.
affair /əˈfɛɹ/ (substantivo) — negócio, afazer. Something which is done or is to be done; business of any kind, commercial, professional, or public.
phrase /ˈfɹeɪ̯z/ (substantivo) — expressão, frase. A short written or spoken expression.
evaluation /ɪˌvæljuˈeɪʃn̩/ (substantivo) — avaliação. An assessment, such as an annual personnel performance review used as the basis for a salary increase or bonus, or a summary of a particular situation.
puzzle /ˈpʌzəl/ (substantivo) — enigma. The state of feeling confused or mystified because one cannot understand a complicated matter, a problem, etc.; bewilderment, confusion; (countable) often in…

Phrasal verbs que você vai ouvir

turn off (verbo) — desligar, apagar. To power down, to switch off, to put out of operation, to deactivate (an appliance, light, mechanism, functionality etc.).
figure out (verbo) — descobrir, deduzir. To come to understand; to discover or find a solution; to deduce.
set up /ˌsɛt ˈʌp/ (verbo) — preparar. To make ready for use.
find out (verbo) — descobrir. To discover, as by asking or investigating.
come up with (verbo) — pensar (em), alcançar. To manage to produce, deliver, or present (something) by inventing, creating, thinking of, or obtaining it.
deal with (verbo) — lidar. To handle verbally or in some form of artistic expression; to address or discuss as a subject.
give up (verbo) — render-se. To surrender (someone or something); to inform on (someone).
go up (verbo) — subir. To move upwards.

Pronúncia para ficar de olho

O falante usa 17 contrações e formas reduzidas, como wasn't, doesn't, wouldn't. Diga-as na forma curta, do jeito que você ouve.

Os sons de “sh” e “zh”: evaluation /ɪˌvæljuˈeɪʃn̩/, punish /ˈpʌnɪʃ/, instruction /ɪnˈstɹʌkʃən/, negotiate /nəˈɡəʊ.ʃi.eɪt/, gibberish /ˈd͡ʒɪ.bə.ɹɪʃ/
Palavras longas — acerte a sílaba tônica: evaluation /ɪˌvæljuˈeɪʃn̩/, experiment /ɪkˈspɛɹ.ə.mənt/, similarly /ˈsɪm.ə.lə.li/, differently /ˈdɪf.ə.ɹənt.li/, uncomfortable /ʌnˈkʌm.fɚ.tə.bl̩/

Como praticar com este vídeo

Ouça o vídeo inteiro uma vez sem falar e anote as palavras que você não conhece.
Comece na velocidade 0,75×, faça shadowing frase por frase e volte à velocidade normal quando ficar fácil.
Grave a sua voz e compare com o original, prestando atenção a palavras como scientist, robot, researcher.

Gramática neste vídeo

As estruturas que o falante mais usa, com as palavras exatas do vídeo:

Estrutura	No vídeo
Voz passiva be + particípio passado — o foco está no que acontece, não em quem faz	being watched · were given · were built
Present perfect have/has + particípio passado — uma ação passada que ainda importa agora	have started · you've probably seen · That's been
Orações relativas who / which + oração — informação extra sobre uma pessoa ou coisa	images, which is · Runners, which is

O que é a Técnica de Shadowing?

Shadowing é uma técnica de aprendizado de idiomas com base científica, originalmente desenvolvida para o treinamento de intérpretes profissionais. O método é simples, mas poderoso: você ouve áudio em inglês nativo e repete imediatamente em voz alta — como uma sombra seguindo o falante com 1-2 segundos de atraso. Pesquisas mostram melhora significativa na precisão da pronúncia, entonação, ritmo, sons conectados, compreensão auditiva e fluência na fala.

Técnica de shadowing: leia o guia completo passo a passo →

Prática de Shadowing: 10 Times AI Behaved In Ways That Terrified The Scientists Who Built It - Aprenda a falar inglês com vídeo

Escolha um modo de prática

Baixar aplicativo

Anthropic AI Researcher Issues Dire Warning it 'Could Kill Us All'

Dear Higgsfield, I reverse engineered you in 12 hours

Claude Just Revealed AI's Biggest Problem

What Happens If China Invades Taiwan? Elon Musk's AI Warning

The Alignment Problem Explained: Crash Course Futures of AI #4

5 Things You Should Never Ask ChatGPT | Vantage with Palki Sharma

He’s always wrong

The 7 Trillion AI Gamble Is Failing. Big Tech is TRAPPED Right NOW.

How the Algorithm Hijacked Monkey’s Brain (Why Monkey Can’t Learn Online)

AI Can Write Code. And That’s the Problem.

THEY'RE HIDING THE TRUTH ABOUT AI

MIT Explains the 12 Possible Endings for AI