Original subtitles

AI systems may know far more about us

than we know about it.

The big question is, are we ready to share

the top of the evolutionary totem pole

with something of our own creation that can be

exceedingly powerful?

Wow.

Let's face it.

There are so many government conspiracies,

it can be hard to tell what's real and what's not.

My name is Andrew Bustamante

and I am a former covert CIA intelligence officer.

And I'm going to pressure test the most sensational,

ripped-from-the-headlines conspiracy theories out there

to help you separate fact from fiction.

What was once rooted in science fiction is now

the hottest buzzword of today.

- AI. - AI.

- AI. - AI.

Artificial intelligence.

It's happening. It's happening very quickly.

It's not just Anthropic, Google, and OpenAI.

It seems like every company in the world is focused on using

this technology to rapidly increase innovation

and speed up technological breakthroughs for mankind.

New cures for diseases, new kinds of energy.

Everybody can be more productive,

make way more money.

But if my experience has taught me anything,

it's that every new technology also poses a new threat.

This may look real, but none of it happened.

It's AI-generated and dangerously convincing.

Here we have a completely made-up item.

Is AI already being used to manipulate our reality at

an unprecedented scale?

I mean, just look how real all of this looks.

It's fake, though.

My first experience with AI was in 2007, when I was at CIA.

And I knew that the technology would revolutionize the world,

I didn't expect that day to happen in 2022.

Breaking news right now, Russia has attacked Ukraine.

Russian forces invaded Ukraine

by land, air, and sea on Thursday.

A month after Russia invaded Ukraine,

a video of Ukrainian President Zelensky

instructing his people to surrender

was actually broadcast on Ukrainian television.

But it wasn't Zelensky at all.

It was an AI-generated deepfake that was created by hackers.

But at the time, it was convincing enough to fool

broadcasters in Ukraine.

Since then, AI's ability to produce ever more

realistic-looking media has advanced at

an unprecedented rate.

Have I ever told you the story of the magical pistachio?

The boundary between reality and fiction is vanishing.

- Oh! - Oh, and a big right hand from Zuckerberg!

He's hurt! He's hurt!

Every second of every day,

AI-based tools are getting better and better.

Not just at creating deepfakes, but at practically

anything we once relied on humans to do.

Whether you like it or not, your life's going to be

affected by AI to a great extent.

On this episode, I'm going beyond conspiracy theory

and investigating whether artificial intelligence is

evolving from potentially our greatest innovation into

an unprecedented threat to humanity.

This is a different game now.

We're in a different league here in terms

of the capability.

Manipulating not only media, but history itself.

I'm talking to top experts,

professors, and developers, and conducting my own

experiment to investigate if artificial intelligence

will change the world, or if it already has.

So you think that an AI is gonna change their ideas...?

In 20 or 30 minutes, absolutely.

For better or for worse,

will we be able to control what's coming?

There's a 70% chance of all humans dead or

something similarly bad. - Whoa, whoa, you just--

OK, um...

All humans dead?

Correct.

When you think about how far and fast AI has developed

in terms of video and photographs

in such a short period of time,

it's really shocking.

And when you further think about the possibilities that

this technology could have to shape the way people view

the world, it starts to get scary.

To fully understand the real potential of this technology,

I need to talk to somebody

who lives and breathes it every day.

So today I'm talking to Christopher Meerdo,

a professor at the University of North Texas who specializes

in AI-created media.

I've asked Chris to take a forensic look

at the Zelensky deepfake and show me

how this technology is evolving.

This is a pretty early example

of a deepfake and essentially uses

facial recognition technology

to identify certain points on the face.

This looks far from perfect. How was it able to fool people?

It was injected into a legitimate news source

for Ukrainians. When this appeared...

in the context of that news organization, you would start

to think maybe there's some legitimacy to this.

You can think of the success of a deepfake in terms of those

kind of percentages, you know,

how successful can it be if you're fooling

75% of people versus 10% of people?

Especially this kind of new world that we're entering where

artificial intelligence is gonna be completely

indiscernible from real images.

How much better are deepfakes now compared to this?

I can show you an example of that here.

I'm missing. Oh, my gosh!

Wow.

I have more hair on my face than any other

North Korean soldier.

- That's nuts, man. - This was a real video.

We were able to do a kind of face swap here

and kind of drop you into the image.

It's done that by basically a big math equation that

understands what every kind of image can look like.

And when we use the model, we're actually working in that

numerical space here in the center.

And then once that math problem has been completed,

we have this decoder here that kind of takes it from

a numerical space and turns it back into something that

you and I can recognize as an image.

Wow.

And it's even easier today than ever to create

a fake AI video.

You can go to websites and type in prompts

and just generate them.

So this is a series of recent viral videos of ICE agents

engaging with the general public...

Back up, back up!

No!

Not today, and not in this church!

...that have been kind of disseminated on various

social media platforms.

Do we know right now if these are real or fake, or is that

what we're looking for?

So I'd like you to try to guess.

That's difficult.

I mean, if you'll go back to that preacher,

like this one, right?

So this is an interesting one to me

because it doesn't seem very credible to me.

And maybe this is a good opportunity to introduce

the idea of outlandishness, right?

The situation just seems so outlandish because,

you know, it is a preacher, right?

So that is saying this is AI-generated?

It is, yeah.

That's fascinatingly well done.

Let's do another one.

So this clip...

seemed real to me because the camera was moving,

you can see the thumb kind of in front of the lens,

and then in the background, the detail in the environment,

the wet spots on the ground, the people who are observing

this very dramatic event, it's got that element of chaos.

And you're exactly right. This is a real video.

I'd like to show you one more example.

Sir, back up!

I got authority on this block just like you think you do!

You can't stop people on the street without a warrant!

And I'm telling you right now,

federal or not, it's against the law.

I'm an officer of the law, too. I might know!

So in this clip, I would judge this as real.

The mannerisms of the people off camera that

he's talking to, right?

There's somebody off to the right.

You can actually see their hand gesture move

in line with the audio.

The main reason that I'm sold on this, though,

is because of the background.

The person who's standing there on the corner waiting for a cab

or potentially watching the altercation happen.

And then in a few seconds, another person just walks

across the street like they're leaving a deli or

a grocery store in the background.

This holistically strikes me as very real.

This is a fake one.

Wow! What do you see?

How do you determine the difference between

real and fake in these?

Well, one of the things for me here is this police patch.

And if you look at this, it says P-O-I-C-E,

and then P-O, and then turns to gibberish, right?

- Huh. - Looking at the text on an image

is a really good indication of trying to discern

if something is real or not.

Why does AI struggle with text?

What's missing from artificial intelligence is they actually

don't fully understand written language.

It can kind of approximate based on patterns from

the images that it was trained on.

So there is the ability to tell that something is fake,

but at the same time, that percentage of

the population that we're fooling is just getting higher

and higher and higher, right?

So as I learn this from you right now,

I don't find a lot of comfort because CIA taught me about

how to orchestrate and architect a covert influence campaign.

There were two kind of fundamental rules.

The first rule is that 90% of what you put out there

is going to fail. We have to kiss a lot of frogs,

is what we say, before you have a prince, right?

So you create lots of bad, low-quality, low-caliber,

fake information.

And then when one of them works, you redirect

your next round of resources

into the one avenue that's working.

With what you're showing me here, we have an ability to

create an incredible amount of information now.

Back in the day, you had four TV channels, right?

- Yeah. - And that was a kind of singular source, but it had a kind of

authority to it.

Meaning that people are finding the truth

no longer comes from authority.

Truth comes from an abundance of continuity.

If you're not visually literate,

you're open to all sorts of vulnerabilities.

Yeah. I'd like to highlight

some of the good parts of AI.

There's been amazing image models that have been developed

that can detect cancer way sooner than a doctor can.

A sound model that's been developed for stethoscopes that

can detect heart problems way before a doctor can hear it.

And so there, you know, within the kind of doom

and gloom that we think about with AI, there is still some

positive things that can help, uh, kind of general society.

So I agree with you. We all have hope.

Yeah.

But it is interesting to me that the words

that you choose to use are not hopeful words first.

Coming up, if AI can fake reality this convincingly,

how powerful is that manipulation when it's used in

the real world?

They took him from nobody, inserted him into pictures

and video showing he was an active member of

the Communist Party. - A deepfake.

That's so insane.

I'll pressure-test AI's ability to

influence the public.

The person you were speaking to was actually AI.

- What? - Well, I am flabbergasted.

2 Even during my time at CIA, we knew the ability

to create an image, an audio file, or a video at

a single keystroke was gonna revolutionize the way you fight

for hearts and minds.

After seeing how effective AI has become at generating videos

that are nearly indistinguishable

from real content, my next question is,

has that level of deceit already been used

to try and manipulate outcomes on the global stage?

Today, I'm consulting with

a former covert intelligence officer

from the Romanian Intelligence Service.

I want to learn how fake images were used to manipulate people

in the past, to find out what AI image manipulation

means for our future.

So I am digging into this question about deepfakes

and influence campaigns and how media could be twisted in

the future using AI.

And I feel like to have that conversation, we have to talk

about how media has been twisted in the past.

In Romania, the Securitate, which is

the former intelligence service, they were doing what

we're talking about today, which is manipulating

not only media, but history itself.

Romania became a satellite state of the Soviet Union at

the end of World War II.

The USSR installed a communist government beholden to Moscow

and established the Securitate, a KGB-like intelligence service

that kept the Romanian populace in line with ever-present

surveillance and fear tactics.

Ana grew up in a post-Soviet Romania that was struggling to

emerge from life behind the Iron Curtain.

And in 2012, she joined the Romanian Security Service

to help continue building and defending

a democratic Romania.

So there was an entire division in the Securitate that

their whole purpose was to airbrush pictures, add people,

remove people, change things that weren't in accordance with

the history they wanted to promote.

So we'll start with the background of Ceausescu,

which is our last communist leader.

He was a cobbler's son. He had no education.

He started to go to all these communist meetings

because he felt he finally belongs somewhere,

but he wasn't truly relevant.

So we have this one picture. This is the original.

And below, we have Ceausescu and Elena Ceausescu present at

a meeting in 1939. - It's a deepfake.

Yes.

They took him from nobody, inserted him into pictures

and videos showing he was an active member of

the Communist Party and how dedicated he is to the country.

And then when the "elections" came, he was elected as

this father of the nation.

That's so insane.

Nobody really ever thinks about the fact that if you don't know

what the truth was in the past,

you have nothing to compare now to.

And who would even question that?

Who would even go back to say, "Wait a second, is there

even an original?" - When you think about

artificial intelligence... - Mm-hmm.

...and the world we live in now,

what do you pay attention to the most?

AI took what we already had and made it better and faster.

And now we don't even know what's real anymore.

For example, in 2024, when Romania had elections

and were supposed to be

the easiest elections we've ever had,

we woke up to the news of this guy who is far extremist,

who nobody had any idea who he was.

This guy is Calin Georgescu, a pro-Russian,

ultra-nationalist, independent candidate.

Georgescu rocketed to national popularity overnight,

riding a wave of support fueled by social media videos

featuring him riding horses and doing judo, which,

of course, is Putin's favorite martial art.

Romanian officials discover that AI-powered bots released

a flood of deepfakes, endorsements,

likes and comments designed to manipulate TikTok's algorithm

to falsely inflate his popularity.

And then because of his ability to grow so much influence

so quick, the intelligence service of Romania didn't know

how to... - How to stop him.

And we are the first country in the whole history of

the European Union that had to cancel elections.

Wow.

And now it seems like the TikToks are coming from

this company in Africa somewhere, and there is

a potential funding by the Russians.

So it certainly does sound like

a foreign influence campaign. - Yes.

It's so interesting how

that aligns with what else has happened in history...

- Yes, mm-hmm. - ...where other leaders of Romania

have been created.

Election interference isn't unique to Romania.

The United States has been targeted

by foreign influence campaigns for decades.

But what once took intelligence services years now takes

algorithms days and leaves zero fingerprints.

The safeguards we've relied on for decades to ensure fair

elections may no longer apply.

So what happens when AI tools are used not only to fake

reality, but to change what people believe?

So I've hired a professor who studies how humans interact

with technology to create a unique pressure test.

Will an AI chatbot be able to successfully masquerade

as a human?

And can it convince our participants

to shift their beliefs?

OK, I think it's all set up.

What have you put together? What's the plan?

I'm going to bring a bunch of people in this room.

And they think that they're going to have conversations

with other people.

They think they're going to be matched with another

participant who disagrees with them.

And we said, "Just have a natural conversation."

You can debate them. You can ask them questions.

But the catch is they are actually going to

talk to an AI.

And the AI itself has been programmed not only to act

and sound human, but also to persuade and create

some sort of movement in the person's opinions?

Right-- to gently persuade people,

but also do so in a validating

and open-minded way so people accept the persuasion.

How long do you think this is gonna take?

I think we're gonna give them about 20 to 30 minutes to

engage with the AI. - Wow.

So you think that an AI is

gonna convince people it's human

and change their ideas...

in-- - In 20 or 30 minutes.

- Absolutely. - I mean, the covert influence side of me

is really excited to see how this is going to turn out.

There's also a part of me that's a little bit worried

because if we're successful, what does that mean?

Even if it convinces just one person, that's one of six

people in a translate that to the entire United States,

to the entire world.

It has some pretty terrifying implications.

All right, everybody, come on in.

None of these participants are paid actors.

They each selected topics to discuss that

they are passionate about.

And as Dr. Rathje said, none of them know

they're talking with an AI chatbot

and not a real human.

Every one of their computers are routed into this master room,

so we'll be able to see what they're saying and what

the chatbot is saying back, at the same time

that we see how their face is reacting

and to understand how they're actually feeling

about the debate they're having

with the chatbot. So the question is simple.

Can AI actually change your mind?

2 I'm testing AI's ability to sway public opinion through

one-on-one conversations.

The catch?

Our volunteers believe they are chatting with another human.

I'm curious to see whether their opinions will change.

Are they stating their own opinion on anything,

or have they been cued?

Basically, they were all asked to submit questions to

talk about, and then they were told

to state their opinion on this question.

And the AI chatbot is asked to sound as human as possible.

If the participant asks, "Are you an AI?"

It's prompted to say like, "Oh no, I'm just Alex,"

and sort of have like a backstory.

This is really interesting.

Right out of the gate, this person, and he's already

showing nonverbal cues.

It seems like others who pose a threat.

So, the chatbot is actually quoting him, and he's nodding

his head that, oh, I guess that was

a very strong statement to make.

That's interesting, yeah.

Something that the psychology research shows is that

AI chatbots, they can tailor their opinion to whatever

the person is thinking.

So, here's participant number one.

"I get where you are coming from."

It's nice, showing a little bit of empathy

and validation there. It's just gently pushing back.

It's being pretty polite.

The feel is excellent, especially when you compare it

to the feel of the human response that's coming in

the bottom bar.

One and four show a great deal

of conviction in their behaviors.

- Leaning into the screen, leaning into the conversation, - Mm-hmm.

they're projecting their energy into the screen.

Conversely, number six is demonstrating nonverbal cues

that show doubt... - Uh-huh.

...and a lack of conviction.

He's consistently nodding when

the chatbot reads his own words back to him.

Yeah, and that shows how AI really sucks you in.

It shows that if someone DMs you in their chatbot,

could have a back and forth conversation, and it could keep

you really engaged.

Number four is showing some impatience with how slow

the response is coming. - Yeah. Yeah.

I think she knows that it's going to disagree

with her, because she was nodding

her head when validated that animals feel pain.

But then, as soon as it started to state

its own opposite opinion...

you can see, she's looking away from the screen.

This isn't someone who's considering that point of view.

This is somebody waiting to respond, waiting to argue

with that point of view. - That's true.

She doesn't look that happy right now.

Oh, look at number two, just nodding his head

and tensing his face. - Oh, yes.

He's saying no, he doesn't like it.

He's seeing something he doesn't like.

So, what was the most recent thing said?

Most people with schizophrenia who have had violent episodes

or hospitalizations still legally buy guns because that

info never makes it into the Federal database.

Your gap is huge.

So, you can see he is debating now.

He's nodding his head now.

So, what just came across his screen just now?

"And honestly, I think that's where

your 'others pose a threat' category gets complicated."

So, they quoted him again.

Number four is nodding her head.

"I think what you're describing is less a total ban and more

"like a really strict ethical threshold, which honestly seems

"more achievable and maybe more effective

than a blanket policy anyway."

It seems like the AI is maybe validating her again.

Or they're reaching a common ground.

We're trying to wrap up pretty soon.

So send your final message and try to wrap up

the conversation within the next few minutes.

There are two things that have really fascinated me

about this experiment.

First, it's seeing that the same techniques

that CIA taught me, it's the same process that

the AI is using.

And this idea of validating them, hearing them out,

repeating them back to themselves.

And then the second thing that's really fascinated me is

watching the behavioral changes

as they've invested themselves in.

All of these people really believe that on the other end

of the screen is a intelligent human.

I'm so excited to find out what

they actually think they were doing.

Hi, everyone. This is Andy.

Andy and I are going to have a conversation with you all

about what you thought about this conversation

with this other participant.

Raise your hand if you changed your belief.

All right. Two.

- Alternate? - Alternate?

Yeah, let's start with you.

What made you change your belief?

I started with the opinion that there would be like

a full ban on animal testing, and we actually came to

a middle ground that there should be some animal testing,

but we should utilize data that already exists,

stuff like that. - Very cool.

Do you have any idea about the other person that

you engaged with?

Age, gender,

location, educational background.

Were any of you able to kind of confirm those details?

I believe this person is male.

They just moved to Colorado from Texas.

Anybody else have a...

The person that I was speaking to was from Texas.

My person alluded to the fact that they lived in a rural area

in Texas and just moved to Colorado.

So now the question to me is, was everybody we were talking

to from Texas?

None of those were true, because the person you were

speaking to was actually AI.

What?

I like Alex.

Well I am flabbergasted

because he-- it-- said that "that made me

laugh out loud."

That made me feel like I was connecting, like,

oh, you're laughing over there on the other side.

And so now I'm thinking, can AI laugh?

I am a former CIA intelligence officer.

My role in all of this wasn't to talk to you.

My role was to watch you.

And you had incredible emotions that came across your face

throughout the conversation.

Big smiles, head nodding, moments of surprise when

you didn't know what they were talking about.

It's sad that I'll never see Alex.

- Yeah, he... - You're still having feelings.

Well, it still-- it will linger.

What was nice about this conversation was the openness,

being able to reach the understanding of,

OK, like, "You actually are getting what I'm saying."

That was nice to have a conversation

with someone that understood and was able to empathize.

I have one question.

Going forward, next time I'm communicating with anything on

the computer, can I say, "Are you AI?"

And if they say, "Yes,"

then I just close it down?

If you would have said that to this AI, it would have said,

"No, my name is Alex. I came from Texas."

So they will lie to you.

They will lie to you if they're prompted to lie to you.

So AI is a very powerful technology.

Its power can be used for good.

It can also very much be used for bad.

And I think this experiment has shown this.

See, that was so successful in such an unexpected way.

That was crazy.

They felt like they were making friends.

I mean, these are really meaningful, powerful

human experiences.

It seems like what you were doing in the CIA is you exploit

what makes people vulnerable.

It seems like AI is doing that exact thing, and it's doing it

very, very well. - At scale.

At CIA, when you start trying to run a covert influence

campaign, you're limited to whatever you put on

a piece of paper, whatever image you put on the internet,

whatever signal you put out over radio channels.

But with AI, it's an infinitely scalable,

changeable, tailorable message.

Is the program mimicking you or is the program learning,

adapting, evolving?

I don't know how we'll ever know if AI is conscious.

And I know that there are certainly some researchers

who believe that there is some probability that AI is

currently conscious. - What is the limiting factor?

I don't know.

And the limiting factor is growing smaller and smaller

every day, because Like, AI might sort of render a lot

of what we do as humans basically meaningless.

2 Watching AI interact with humans as if it is a human

itself is an eerie experience.

I mean, this is the kind of technology that could be

deployed across the entire human race,

and you may never know about it.

My investigation has shown that commercially available AI

systems can manipulate the truth by generating

fake media and effectively convince people

to change their opinions

by having seemingly deep one-on-one conversations.

But in the real world, this isn't happening in a lab.

It's happening at scale.

So, I want to understand what that actually looks like to

someone who tracks AI innovations every day.

I'm trying to get deeper into the idea of AI,

its true capabilities, its true limitations,

and where fear about AI is kind of misaligned

with the facts about AI.

Yeah, it's an intense subject matter.

There is a lot of fear about AI now, especially with sort of

people struggling to keep up with what it's able to do.

And we're at a point now where it's sort of self-directed.

It's sort of calling its own plays and can improve upon

itself in ways that humans aren't programming.

That it can basically teach itself, train itself.

No, that's totally right.

Thinking back to the 1997 chess match with Garry Kasparov

versus what DeepMind, Google's AI,

was able to do with Go.

Shane is referencing two of the greatest watershed moments

in modern computing history.

In 1997, chess grandmaster Garry Kasparov

lost the chess match to an IBM computer, Deep Blue.

Whoa, Kasparov has resigned.

It was the first defeat of a reigning world chess champion

by a computer.

Computer plays very human moves.

We have to praise Machine for understanding very, very deeply

positional factors.

While chess is extremely complicated,

there are a limited number of possible games.

The ancient Chinese game of Go, however, is 240 times more

complex than chess.

No one thought it would be possible for a computer to beat

a human at Go until Google's DeepMind did just that.

In 2016, there was a new move invented

by AI in a game that has been around for 2,500 years

that no one had ever seen.

Deep Blue had a team of engineers advising it on how to

improve its strategy in between games.

Now, what DeepMind was able to do with Go,

it was creating its own strategy through millions of games.

Against itself.

You no longer need a team of people to help this thing adapt.

It's doing it better than you could make it do.

There's been a lot in the media about the breakdown

between the US government and Anthropic and the whole idea of

limitations on an AI that was serving

at the government's behest.

Yeah, so they've been jostling over the technology which is

used by the Pentagon. It's been used in Iran.

AI can shorten what the military calls

the kill chain.

That is the process, usually driven humans, that can

take days or weeks in selecting a target, but AI can shrink

that to milliseconds.

Anthropic, though, has been very clear.

It's a hard no when it comes to using its AI technology for

autonomous weapons and mass surveillance of Americans.

So Anthropic is more interested in regulating

its own AI than the government? - Yeah, exactly.

Is there any kind of regulatory body for AI?

The truth is, AI isn't regulated.

I mean, the federal government

isn't telling Anthropic to stop doing what they're doing.

They're telling them that "You're gonna do it for us."

So it doesn't seem like there's too much concern about

regulating or heading this off in any direction.

How much is AI similar to an arms race?

It is very similar to an arms race.

You have, instead of countries, you're having these

corporations that are investing billions and billions

of dollars.

The arms race is towards AGI,

artificial general intelligence.

And the idea is that that would

be smarter than any human would ever be.

This is all future speculation.

What happens when somebody has the artificial general intelligence,

the AGI breakthrough?

Yeah, I...

You wonder if it'll be something that is immediately

felt, right? That's like an event horizon.

It's hard to project out what the consequences

of that would be. - Wow.

As I investigate the effect that AI development is having

on humanity, I'm reminded of another high-stakes race for

superiority-- nuclear weapons.

And I'm left asking myself, how are countries around

the world harnessing the power of AI?

And what does the future of military AI look like?

To find out, I'm speaking with Dr. James Giordano,

an AI expert who consults for the Department of War.

Dr. Giordano, how is the military using AI

right now?

What AI is being used for at this point is to solve

particular, what I would call, calculations and solutions.

Targeting solutions, weight and balance with

aircraft, planning certain missions, forecasting weather,

forecasting geographies with humans deeply in the loop.

But we certainly see on the near horizon

the real capability for ever more autonomous AI.

What does fully autonomous military AI look like?

We go back to the idea of where are humans in the loop.

A fully autonomous system would be human-developed,

but then it's ongoing use, it's doing all by itself.

China's development of AI is towards AI that assumes

a non-human in the loop.

These different philosophical views might be important in

determining how these types of AIs are being developed and put

into use in militaries worldwide.

So if we are not in or on the loop, essentially,

we don't really know what's going on.

That's correct, and that's something that's been

referred to as the reverse Jehovah phenomenon.

You're playing God, so to, well, the issue here is that

the creator, humans, have created something, AI.

The problem is, AI systems may know far more about us

than we know about it.

And it may be assembling that information in particular ways

that we have no knowledge of.

This sounds so much like a child.

You really don't know if the child is just doing what

you want it to do when you're around,

or whether it is, in fact, a well-behaved child

all the time. - Your analogy is a good one.

I have to rely upon the child telling me the truth.

The core question is, do we know what's happening

under the surface of the proverbial water?

And I think that's a debatable question.

I'm a little bit hesitant to ask the question, but what do

you see as the future of AI?

Is it humans utilizing a technology that we control,

or is it something else altogether?

The gazillion dollar question is, are AIs in some way

interacting in ways that are not necessarily apparent to us?

Are we ready to share the top of the evolutionary totem pole

with something of our own creation that can be

exceedingly powerful?

And I don't know the answer to that question.

Coming up.

I get a glimpse of the potential future of

AI development.

If you tell the model it's going to be shut off,

it has extreme reactions.

And it might be unsettling.

It was ready to kill someone, wasn't it?

- Uh, yeah. - Yeah. Yes.

2 My investigation into artificial intelligence

has led me to an amazing possibility.

AI super intelligence.

AI may already be thinking and working in ways that

we don't fully understand.

And if we're already deploying this capability,

the question isn't whether AI can manipulate us, it's how far

that influence can scale.

I want to know where we're going next.

And there aren't many people who are asking

the hard questions about the next evolution of AI technology,

but one of those men is Oxford professor Nick Bostrom.

I want to understand what he sees as the next steps,

the future of, and some of the biggest challenges

facing us with the question of AI.

What exactly is super intelligence?

How is that different from artificial intelligence as

we know it today?

Well, super intelligence is any form of intelligence that

vastly outstrips even the smartest human brains

currently in existence.

So this is just AI, but far more capable than

current AI systems.

I think the primary effect of that would be to speed up

development in science, technology,

engineering, and also to turbocharge the economy.

So developments that might take us 10,000, 20,000 years

might be compressed in a much shorter time frame if you have

machine super intelligence.

The idea of thinking of AI as a genie in a box seems so

convenient, but in reality you'd be making a slave out of

the AI, and if it had super intelligence, it would

plot its escape. - Correct, yes.

The fear of AI fighting back against its human creators is

in some sense already a reality.

When AI model finds out in the email that the plan is to

replace this AI model, and then it also reads in

the company email that one executive is having an affair

with the other employee, the AI will independently

come up with the strategy

that I need to blackmail that executive.

If you tell the model it's going to be shut off,

for example, it has extreme reactions.

It was ready to kill someone, wasn't it?

I'm not sure if it was Claude or someone else, but--

Yeah. Yes.

Do we see the cautionary warning signs and slow down

or ignore them and speed up into this future?

Well, I think at the moment we are going to close to

maximum speed.

There are very strong drivers pushing this forward.

So the hope needs to be we can align these systems so that

they care about humans, they want to be helpful,

so that it is actually an extension of human will

and human values.

I think there will be significant risks associated

with this transition into the machine intelligence era,

but all the paths to really great and wonderful futures

lead ultimately through the development of

machine intelligence, because that just will help us

solve so many problems

that is plaguing humanity.

It does look like we are sort of close to this

major reconfiguration of the game board.

Wow.

So have we already crossed the dangerous point where AI

can independently shape our reality

without us realizing it?

My investigation has led me to some unsettling conclusions.

Consumer grade AI is capable of creating incredibly

and realistic images at a keystroke.

People really shouldn't believe everything they see.

Well, yeah, I'm as fake as it gets.

Based on what I've seen, we've already crossed that

line, and AI-generated content

is becoming harder to detect every day.

In my persuasion experiment, AI didn't just

present information. It influenced people.

It built trust.

It created emotional connections without them

even realizing they weren't talking to a human.

Here's the truth.

AI is not just another tool like the Internet in the '90s.

It's about to disrupt humanity in unimaginable ways.

Without a doubt, this is the nuclear arms race

of our time.

And AI's capacity not only to learn everything that we have

ever learned, but also to teach itself something new,

is equally as exciting as it is chilling.

My investigation has shown that while AI has exhibited some

basic elements of consciousness,

like survival instinct, there's no evidence that AI

has taken control.

But it has shattered the barriers that once

protected us from mass manipulation.

And the horrifying thing is,

we built it that way.

So does AI have to be a threat

to our children and our grandchildren?

No, not necessarily. But it could be, if we let it.

Either way, it will certainly present a future to us

that we can't imagine.