AI and the end of the age of St. Thomas, seeing is not enough to believe

by Alessio Jacona*

The image of former President Donald Trump marching at the head of a huge cheering crowd amidst flags with stars and stripes, and those of his daring arrest during which he is forcibly dragged away.

The shot that portrays French President Emmanuel Macron in the midst of street clashes in Paris against the pension reform, and those in which Russian President Vladimir Putin sees himself wounded on the ground, on the battlefield, or on his knees, in the act to kiss the hand of the President of the People's Republic of China Xi Jinping.

And then, again, the reportage from war-ravaged Ukraine, as beautiful as it is false, by Barbara Zanon, a Venetian photographer who created the images to revive the debate on the risks and opportunities associated with the use of AI.

The risks of generative AI

All of these “photos,” along with the myriad of Pope Francis dressed in fashion, clubbing or even parkouring, have at least three things in common: they are fake, but they are realistic enough to fool the public, especially if used on social networks, perhaps shared without comments by a trusted contact;

they were created with generative artificial intelligences such as DALL-E, Stable Diffusion and above all Midjourney, now in its very powerful version 5;

above all, they can potentially be used to manipulate, steer or even just confuse public opinion.

In fact, we have entered a new era of propaganda.

More fake news for everyone

Of course, the problem of fake news, of false news and contents artfully created to deceive people, is not something of today, and is indeed as old as communication itself.

But it is the first time in the history of humanity that to create them you have tools so powerful and at the same time so easy to use, accessible by practically anyone with an internet connection and the willingness to pay a small monthly subscription.

And, above all, usable without the need to have any IT skills: in most cases, just write a "prompt", i.e. a text that describes what characteristics the image must have, feed it to the AI, and that's it. .

You don't even need to know English, because these systems already recognize and understand dozens of languages, including Italian.

A political question

The fake photos of Trump's arrest - created for fun by Eliot Higgins, founder of the investigative journalism platform Bellingcat, and immediately went viral - warmed the hearts of many of his supporters in the days leading up to his indictment (this time real) for 34 counts, including illegal payments of $130,000 to buy porn star Stormy Daniels silence during her election campaign.

Those of the Pope "viveur" and lover of worldly life partly amused, partly scandalized those who initially mistook them for real.

Predictably, they have been revived by its detractors, with the intention of undermining its image and undermining its authority in front of an average public unprepared to recognize the deception.

These are just two examples,

There is still (some) time

The good news is that we still have some time to take measures, define rules, set up a vast cultural operation that makes all citizens more aware of the use of AI tools, because these technologies are still imperfect and can still be unmasked .

This happened, for example, with the false image of Trump marching at the head of his voters: shared by his son Eric on Twitter, accompanied by the phrase "One of a kind" (One of a Kind) without any other explanation, on 8 April was viewed nearly 10 million times, retweeted 23,000 times and commented on by nearly 26,000 people.

Reading the reactions, it turns out that many comments are critical or mocking, because for example the faces of the people in the crowd are deformed and therefore clearly the result of an error of the generative AI,

Constantly evolving

The question is what will happen with the next version of these AIs, or the next one, when they will be able to generate near-perfect images.

And it's a matter of months, not years: version 5 of ChatGPT - which according to some experts could be indistinguishable from humans - should be released next winter, while Midjourney 6 could arrive by the end of the year.

In spite of the appeals to curb the development of artificial intelligence for six months - signed by the protagonists of the sector and even by Elon Musk - and despite the intervention of the Italian Privacy Guarantor, which for the moment has blocked the access to ChatGPT, the Large Language Model of OpenAI, demanding greater guarantees for user data and also inspiring other countries such as Canada, Germany, France and Ireland.

Beyond the photos: the voice (and then the video)

And if what can be done with images is worrying, the question of audio deserves a separate discussion, or rather of the voice, and how it can already be counterfeited in a practically perfect way today.

In fact, there are already numerous online services based on artificial intelligence that are able to clone any voice, complete with tone, timbre and cadence, based on a few minutes of recording.

Born for more than commendable purposes, including restoring the voice to those who have lost it (like the actor Val Kilmer in "Top Gun: Maverick"), or allowing an actor to "read" aloud dozens of books in minutes, these services can easily be exploited for criminal purposes or to harm one's enemies.

This too has already happened: when, last January, the company IIElevenLabs opened its speech synthesis platform to the public in beta, someone immediately thought of using it to make Hillary Clinton read a transphobic text, make Bill Gates say that the vaccine against COV1D-19 causes AIDS and then, again, to make actress Emma Watson (Harry Potter's Hermione) read Adolf Hitler's Mein Kampf.

After all, an AI-created voice reads any text you throw at it, no questions asked.

Don't stop, but rule the AI

It's as impossible to stop the progress of artificial intelligence as it is to stop the wind with your hands.

Existing systems will evolve into even more powerful new ones, other players will come into play (like Google with its Bard), and with simple prompts it will soon be possible to create not only images, but also credible videos (even if this latest technology is still behind the others).

Our ability to use these extremely powerful technologies as tools for progress or destruction (of work, knowledge, trust) depends on the choices we will make now to govern their evolution and direct their use.

Choices that we will have to make quickly and with caution, before we find ourselves in a world in which seeing (and even hearing) is no longer enough to believe.

* Journalist, innovation expert and curator of the Artificial Intelligence Observatory