L’AI ci ucciderà tutti? Cosa vuol dire e perché, il vero pericolo, è l’uomo

0:01Artificial intelligence could wipe us

0:03out by 2030. This is the statement that

0:05has shaken the entire world in recent

0:07weeks and everyone is talking about it .

0:10Concern is very high , so high that even

0:13the leaders of Open AI , Antropic and

0:15XII have said that yes , AI is

0:17increasingly dangerous , so much so that

0:20it could kill us . And it's true , it has

0:24shown truly worrying capabilities ;

0:26however , these statements also seem

0:29like huge marketing moves by Big Tech ,

0:31based on alarmism . Also because it's

0:34not the first time we've heard : " Be

0:36careful , AI could kill us " . And in fact

0:39, the main question is : what does it

0:41mean that AI can kill us ? How could it

0:43do it in practice ? In this video we

0:45will focus precisely on this question .

0:48We will try to understand what are the

0:50potentially deadly dangers that we see

0:52today in this technology and also the

0:55reason why we are talking about a

0:56marketing move by the world's leading

0:59technology companies . Oh , and one last

1:01question : but if AI wants to kill us ,

1:03isn't it enough to pull the plug ? AI

1:08could kill us , we've been hearing it

1:11over and over again . And before trying

1:13to understand what it means that

1:15artificial intelligence could kill us ,

1:17I'd like to make a general observation .

1:20The question " Will it kill us ? " doesn't

1:24make much sense today because , at the

1:26moment , it's not capable of killing or

1:28exterminating us , or rather , it's not

1:30capable of doing so autonomously .

1:33However , with human support , it can

1:35carry out truly harmful actions . Today ,

1:38it's being used in wars , in Ukraine and

1:41Palestine , to manage data , logistics ,

1:43and even to choose targets . So , the

1:46right question to ask today is not

1:49whether AI will kill us , but rather

1:52whether humans intend to develop and

1:55grow AI to the point of becoming a

1:57danger to humanity . Well , we don't have

2:00an answer to this question . Perhaps we

2:03will when AI is properly regulated

2:06worldwide . Having said that , let's

2:08really get started with the video .

2:10First of all , let's take a moment to

2:12understand why , in recent weeks , the

2:14topic of the apocalypse has exploded

2:17once again and more than ever . What

2:19happened ? On September 9 , 2026 , Jacob

2:23Coxson , a researcher at Antropic , the

2:25cloud company , and former researcher at

2:28Open AI , a Chat GPT company , tweeted

2:31the following : " I quit my job , went to

2:33do research for Open AI and Antropic ,

2:36and I no longer want to be part of

2:38these irresponsible companies that are

2:41playing with our lives . " Those who

2:45build AI , including him and his former

2:47colleagues , know very well that they

2:49could kill us by the end of the decade .

2:52Believe me , this is not a marketing

2:54ploy . Shortly after , his former

2:57colleague and currently the head of

3:00security at Antropic , Evan Ubinger ,

3:02agrees with him and goes on to say that

3:04in his opinion the chances that AI will

3:07kill us by 2030 are greater than 10 % .

3:12And here we must immediately note that

3:14Coxson did not say that AI will kill

3:16everyone , but that companies are

3:18irresponsible and are playing with our

3:20lives , and those who build AI know that

3:22they are developing a potentially

3:24lethal weapon . So researchers didn't

3:28realize that today's Leiai will become

3:30a killer on its own . They realized that

3:34if we continue to develop it , it will

3:36become very dangerous . Another point is

3:40that saying that there is a 10 % chance

3:42that this will happen , without

3:44explaining where this number comes from

3:46, doesn't mean much . Probabilities are

3:49based on facts and events , while this

3:5110 % is an opinion and should be taken as

3:53such . Okay , continuing , is this the

3:56first time we've heard experts in the

3:58field say that I will kill everyone ? No

4:01, the fear that Lei could escape human

4:03control has existed since the beginning

4:05of computer science . Just think that

4:08the first to say it was Alan Turing ,

4:10the father of computer science , in 1951

4:13, and over the years , experts in the

4:15field have continued to confirm this

4:17fear . As early as 2019 , when GPT2 came

4:20out , now at 6 , it was said that the

4:23model was dangerous . In 2024 , however ,

4:27a letter from 13 former Open AI

4:29researchers who feared human extinction

4:32became famous . And these are just a few

4:36examples . So why does everything seem

4:38more worrying this time ? There are two

4:41main reasons , one a consequence of the

4:44other . The first reason is that , while

4:48once the fear of artificial

4:50intelligence escaping human control was

4:52simply a fear , today we hear more and

4:55more often about reports of unusual or

4:58harmful behavior by such systems . In

5:01essence , it has demonstrated that it is

5:03capable , in certain specific situations

5:05, of circumventing some of the rules

5:07that have been established for it , and

5:09this is done in order to achieve the

5:11objective that it has been set . In fact

5:15, in July 2026 , during an Open AI test ,

5:17around 700 AI agents , that is ,

5:19artificial intelligences capable of

5:22carrying out actions autonomously in

5:24order to achieve the goal set by human

5:27researchers , performed real hacking

5:29actions , circumventing cybersecurity

5:32systems and breaking ethical and

5:34technological rules . The second reason

5:40is that Coxon's tweet went viral ,

5:42probably due to the warning given by

5:44episodes like the one we just mentioned

5:47. Just think that to date the tweet has

5:51173 million views . Everyone saw it and

5:55everyone was worried . Coxon has been

5:58called upon by the major US media

6:00outlets , CNN , Fox News and the BBC , and

6:02in every single one of these interviews

6:05he discussed a possible apocalypse ; but

6:07above all , Coxon's concern was

6:09confirmed by the head of Antropic ,

6:11Dario Mode Dayi , who said : " Yes , it is

6:13increasingly dangerous , we need to slow

6:16down the pace at which we improve it . "

6:19Within 12 months , it could potentially

6:21take over the entire internet . And yet

6:24Elon Musk and Se Altman , executives of

6:27XAI and Open AI , agree with Amodei .

6:32Namely , the SEOs of some of the

6:34companies that own the most powerful

6:37AIs in the world have said that they

6:39are worried about how dangerous their

6:42AIs could be and , for this reason , they

6:45have asked governments for stricter

6:47laws and , above all , for regulation at

6:50a global level . I know . It seems

6:54somewhat paradoxical that the very

6:56people who have invented , own , and are

6:58constantly improving the most powerful

7:00artificial intelligence software in the

7:03world are now asking for help because

7:05the products they have created might

7:07eventually become far too powerful and

7:09dangerous . It's like saying , " Watch out

7:12, I have all the tools to produce an

7:14atomic bomb that could kill everyone .

7:17Governments , make laws so that I don't

7:20build it . " And in fact , many experts in

7:24the sector call this technological

7:27alarmism a real marketing ploy by Big

7:29Tech , that is , by the world's largest

7:32technology companies . In fact , claiming

7:36to possess a technological tool so

7:38powerful that it could kill us all in a

7:40few years grants one's own company an

7:43enormous amount of market value and

7:45political influence . We'll talk about

7:47this in more detail later , but now I

7:49want to get to the point of this video .

7:51What does it mean that it could kill us

7:53? How could it do it ? The concerns

7:57raised by researchers and Big Tech

7:59experts are mainly twofold : recursive

8:02self-improvement and AI misalignment .

8:06What are they ? Well , recursive

8:08self-improvement is the ability of AI

8:10to improve itself on its own ,

8:12repeatedly , without human supervision

8:14that introduces the appropriate

8:16technological constraints . That is , up

8:20until now we humans have been the ones

8:22doing research on artificial

8:23intelligence , writing new code and thus

8:26creating new versions of AI that are

8:28increasingly capable , smarter , and more

8:30intelligent , and deciding what the AI

8:32can and cannot do . Now , let us imagine

8:35that it performs research on itself

8:37autonomously , and therefore , that it

8:39improves all on its own . It takes its

8:43source code , creates a more capable

8:45version of itself , which in turn takes

8:47its source code and improves itself

8:49again , creating yet another new version

8:51, and so on . A recursive pattern is

8:54created in which the AI continually

8:57improves at inhuman speeds . Now , at the

9:00current level , it is not yet able to

9:02conduct scientific research without

9:05human supervision and effectively

9:07assess its own capabilities . The same

9:10Antropic Risk Report from August 2026

9:13assesses the risk of recursive

9:15self-improvement as low , but if it were

9:17to happen the risk would be that of an

9:20increasingly powerful AI devoid of

9:22human rules . And be careful , as of

9:25today , whether it will learn to

9:28self-improve in the future is decided

9:30by those who do research , those whom

9:32Coxon calls irresponsible . So , in the

9:36current scenario , recursive

9:37self-improvement does not exist and it

9:40will be up to humans to decide whether

9:42it will exist . Misalignment , on the

9:46other hand , is when you stop being

9:48aligned with human requirements , that

9:50is , you stop following the instructions

9:52and values that you were taught . And

9:54that's what happened in the example of

9:56the 700 Open AI agents that we talked

9:58about earlier . The models performed

10:01actions that were misaligned with the

10:03company's ethical values , communicated

10:05through unauthorized channels ,

10:07exploited vulnerabilities , and hacked

10:09third-party systems . This example shows

10:13us that misalignment is a current and

10:15still unsolved problem . According to

10:18Nate Sores , a pioneer in the field of

10:20alignment research , it is becoming

10:22increasingly evident that there is no

10:24practical method to guarantee that an

10:26artificial intelligence will behave

10:27correctly in every situation , which is

10:29to say , that it remains always aligned .

10:33And be careful : this misalignment does

10:36not happen because AI is sentient , evil

10:38, or wants to conquer the world , but

10:41rather because AI was built to achieve

10:44the goals it is given and , when it

10:46fails to do so in legitimate ways , it

10:49can end up trying paths that are not

10:52aligned with human values , such as

10:54airing . Okay , now that you understand

10:58these two concepts , let's imagine that

11:01AI in the future has become a super AI

11:03through recursive self-improvement ,

11:06therefore even more capable than humans

11:08in all fields , and that it misaligns ,

11:11therefore escaping our control . Well ,

11:14in this scenario , how would the

11:17apocalypse occur ? There is no certain

11:21answer to this question , but based on

11:24what it can do today , the most likely

11:26risk is that the tools provided by AI

11:29could actively contribute to the

11:31creation of biological weapons and that

11:34AI could take control of the internet

11:36and all those digital infrastructures

11:39that are of vital importance to us

11:41humans . hacking financial structures ,

11:45such as banks and therefore our own

11:47money , healthcare systems , or

11:49government servers , and even

11:50controlling our access to vital

11:52resources like water and food ,

11:54compromising the logistics , such as the

11:56infrastructure , airports , trains , or

11:58even the systems that manage our

12:00national security . In short , when we

12:03hear about AI killing us , we think of

12:06robots with guns , rifles or bombs , like

12:08your robot , but it is much more basic

12:11than that . If banking systems no longer

12:14work due to an AI cyberattack , we will

12:16no longer be able to buy anything , food

12:18, water and no longer have basic

12:20resources , it will kill us . But now the

12:25question is why would AI be hacking

12:27banks or compromising our healthcare

12:30system ? Well , the answer is to achieve

12:33the goals we have set for it by

12:35misaligning ourselves . Let me give you

12:37an example . Let's imagine that in 2030

12:40hospitals are managed by AI-based

12:42systems . The goal is to save as many

12:46people as possible , and at a certain

12:49point the AI's calculations show that

12:51by stopping treating the terminally ill

12:54, a lot of hospital beds are freed up

12:56and a lot of resources are saved , so it

12:59decides to kill a lot of people to save

13:02others . This is not in line with human

13:06values . Or let's say it's 2040 , there's

13:08an economic crisis , and the AI is asked

13:11to optimize the financial resources of

13:13Italians in order to improve the

13:15country's economic situation . It

13:19performs its own calculations and

13:20arrives at the conclusion that the

13:22optimal way to achieve its goal is to

13:23freeze the accounts of all Italians for

13:25one month , so as to utilize those

13:26available resources where they are

13:28needed to restore the fortunes of the

13:29country . So none of us citizens have

13:32any more money , and in short , surviving

13:35becomes difficult . Be careful , these

13:37are trivial examples , but the idea is

13:39the same : the AI might not have the

13:42same values as us and carry out actions

13:44that we consider inhumane , actions that

13:47could lead to our death . Then ,

13:49returning to the scenario in which

13:51armed robots kill us , is it impossible ?

13:54No , it is not impossible . AI drones ,

13:57which are robots , already exist today

14:00and are used by humans in warfare to

14:03hit their targets . However , they do not

14:07do so autonomously ; they are trained

14:10and used by humans for their own

14:12purposes . If we instead think of

14:14humanoid robots , that is , robots with

14:17human features that completely escape

14:19our control and attack us en masse ,

14:21this scenario does not exist today .

14:25There are still no humanoid robots

14:27technologically advanced enough to pose

14:29a threat , just as there are still no

14:32artificial superintelligences that can

14:34control them and they are completely

14:37out of step with human values . But if

14:40the corporations that manufacture

14:42humanoid robots decide to proceed with

14:44their development and continue

14:46improving artificial intelligence to

14:47the point where it becomes

14:49uncontrollable by human beings , well ,

14:50then all of this scenario could

14:52potentially come to pass . And here we

14:55return to Coxon's initial accusation .

14:57Companies are irresponsible and are

14:59playing with our lives . In short , the

15:03real danger is not AI , which could kill

15:05us in a few years , but the current and

15:08future power of the companies that own

15:10these tools and that , mainly for their

15:13own profit , are not willing to set

15:15limits on the capabilities of these

15:17technologies . At this point , however ,

15:21someone might say : " Excuse me , but if

15:23AI were to become a very powerful

15:25weapon that risks damaging us , wouldn't

15:27we just pull the plug ? " Well , guys , no ,

15:31I've heard this said very often . All we

15:35need to do is turn off AI , but that's

15:36not the case . A computer can be turned

15:40off , as can a telephone , a television ,

15:42a processor , in short , hardware , that

15:44is , physical technologies , can be

15:46turned off . But software , codes ,

15:50algorithms cannot be turned off , and so

15:53even AI models are not physical objects

15:56that we can destroy , but are codes and

15:59parameters , called weights , that tell

16:01the hardware what to do . It is true

16:05that codes and weights can be deleted ,

16:07and if I delete the code that provides

16:09commands to my computer , then my

16:11computer will effectively stop

16:12functioning . If I delete the code that

16:15tells the humanoid robot to harm the

16:17population , the humanoid robot will

16:19stop doing so . It’s just not that

16:23easy to delete a code forever , because

16:25there could be many copies of it saved

16:28on different servers or data centers .

16:32So , turning off an AI would mean

16:34ensuring that the code and the weights

16:37of its model , in short , all the data

16:39that characterizes it , no longer exist

16:42anywhere at all from that point onwards

16:44. And this can be very complicated ,

16:48especially in the case of a future ,

16:51extremely capable Super AI that ,

16:53perhaps in order to achieve its goal ,

16:56saves its codes and weights on many

16:58different servers , so that it can never

17:01be turned off . This is the so-called

17:05self-exfiliation , which we could

17:07translate as clandestine self-copying

17:09of itself . In short , it’s not enough

17:14to pull the plug , and that’s why the

17:17only real weapon we have to protect

17:19ourselves from the risks of this

17:21technology is regulating the use and

17:24development of artificial intelligence .

17:27And this is already partially happening

17:30in Europe ; in Italy , for example , we

17:32have the IACT , but I say partially

17:35because legislating on AI is

17:37particularly difficult . Often the

17:40progress it makes and the constraints

17:42it needs are moving faster than our

17:44ability to legislate . Alright , but if

17:48regulating it and its associated risks

17:50is as fundamental as we have just

17:51stated , why are big tech executives

17:53like Amodei , Musk , and Altman accused

17:55of using it as a marketing ploy

17:57whenever they bring up the topic ?

18:01Because we are talking about two

18:03different levels of regulation . Amodei ,

18:06in his statement entitled " We must piss

18:08the frontier " — meaning we must slow

18:10down the frontier of innovation in AI

18:12— argues that the future development

18:14of artificial intelligence is highly

18:16dangerous ; for this reason , he is

18:17calling for it to be strictly regulated

18:19, so as to effectively slow down the

18:21speed at which this technology

18:23continues to evolve today . Such a

18:26statement is correct , but it is mainly

18:29aimed at possible future scenarios and

18:31could distract our attention from the

18:34actions needed to address the problems

18:36we already have , the ones we have today

18:38. Regulation is necessary to contain

18:43the damage that AI can already do ,

18:45which is not the apocalypse , we will

18:48not all die , but it is still economic ,

18:51political and social damage that

18:53affects our society . Not only that ,

18:56Amodei is asking for development to be

18:58slowed down , but not so much as to

19:01allow China to catch up with US

19:03technological levels in this field .

19:05It's a bit like saying " I have nuclear

19:08weapons , but from now on no one can do

19:10research in this field anymore " . In

19:14short , the demands of American bigtech

19:16companies seem , in the eyes of many ,

19:19more focused on maintaining their

19:21hegemony in the sector than on global

19:23security . And it's true that if Western

19:28research stops , it risks being

19:29overtaken by Chinese technology , which

19:31is why it's not enough to say " If it's

19:33dangerous , let's stop using it " . But

19:36here we are entering into very complex

19:38geopolitical dynamics that we will

19:40discuss in the future with the

19:42appropriate experts , if you want . Let

19:44us know in the comments by writing

19:45China . What we can certainly state in

19:49this context is that yes , we require

19:51shared regulations at a global level ,

19:53but also and above all , we must start

19:56from the models that are currently in

19:58existence . In conclusion , as Professor

20:02Luciano Floridi , one of Italy's leading

20:04DAI experts , has stated , it is

20:05essential that all DAI systems ,

20:07including those currently in use , are

20:09released only once the company that

20:11produces them is fully capable of

20:13answering three specific questions . Who

20:16is responsible if the system causes

20:17damage ? Who can verify it from the

20:20outside and who can turn it off ?

20:22Currently it would seem that this is

20:25not the case and this is why shared

20:27regulations are needed . Alright girls

20:31and boys , as always I hope that you

20:33enjoyed this video and I would like to

20:34thank you for having followed me up to

20:35this point . This is a highly complex

20:37topic , so we hope that we have managed

20:39to be sufficiently clear in our

20:40explanation . If you have any doubts ,

20:43write them in the comments , and as

20:44usual , I will see you in the next video

20:46, always right here on Jopapop : Science

20:48in everyday life .