Real or AI? The Tech Giants Racing to Stop the Spread of Fake Images
27:38 Share

Amazon q business is the generate A I assistant from A W S, because business can be slow, like waiting through mud, but amazon q helps streamline work, so tasks like some mary zing monthly results can be done in no time. Learn what amazon q business can do for you at aw 点 com flash， learn more.

Hi charle. Hi alex. So I want you to take a look at what i'm showing you in my laptop screen here. Describe what you see.

It's the pope in the puffy coat. okay? So it's the pope wearing a gorgeous White, very fashionable puffy coat.

And he seems to be studying the streets of vatan city, a with a cross on a long chain. And he just looks really fly. He looks good, right? He looks really good.

Do you remember when you first saw this image?

Oh yeah, I do. I remember this whole thing.

There was a whole controversy around this. Yes, so many people got food by this end, if we're being honest here, I was one of them.

And a new image of pope Frances that is making the rounds on social media. Has some people confused?

This just appears to be pope Francis wearing a White puffer jacket with an iced out cross, was read last week, shared on twitter with the caption OK the people drip went viral. Honey, for a professor of computer science at the university of california, berkeley, says this is a perfect case of a manipulated image taking the world by storm.

I mean, the pope in the puffy coat, right? I think we will look back on that image as a very discrete moment in time where reporters realized they can no longer trust what they seek. A lot of reporters that I talk to in the interviewing weeks after that were like, I had no idea because I was just fantastic enough that you wanted IT to be true and so realistic tic that he wasn't obvious that I was fake.

AI has been able to generate text and video for a while, but for red says, what's changing is that this A I generated content is quickly becoming a lot more realistic.

stable. The fusion major ney di images are incredibly good, like the artifacts now are very, very minor.

There were ways to tell that the image of the pope and the puffy coat was fake. What ferid calls artifacts, 一个 lied internet users spotted some inconsistently in the image may have .

seen this viral image of the pope wearing a profit le。 This was also created using the A I tool mid journey. The picture contained one big clue that IT was fake, distorted about his fingers.

but soon freed, says these images won't be distinguish from the real thing. And that matters because some images move markets and change minds.

There is an incredible power of visual imagery that when you can change the photographs, you change history. And I think when we look at photographs, it's really impactful.

From the wall street journal, this is the future of everything.

I'm alexa la, and I am charlie garden burg. Today we're talking about A I generated images, how we're going to tell if they're real.

what's the stake if we can't, and the technological fixes that are gone to help us tell the difference.

Stick er out.

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through mud a help. Luckily, there's a faster, easier, less messy choice. Amazon q can security understand your business data and use that knowledge to streamline task? Now you can summarize quarterly results or do complex analysis in no time.

Q, got this. Learn what amazon q business can do for you at A W S dot com flash. Learn more.

Mini late images isn't new. It's nearly as old as photography itself. Even digital tools like photoshop have been around for more than thirty years. Dinner rw general council and chief trust officer at a doby, which makes photoshop says he's been thinking about how powerfully we react to things we see.

Once you're in a world where you're seeing deep fix, next time you see something you're going to add and over turn out, we're going to get desensitized to the information we're going to see because we have no way to know whether what we're seeing is to or not. And seeing is believing.

Rosas, A I image generators are changing the game to Better understand how they're up the stakes. We need to know more about how A I generated images are made. So if we wanted to make an image like the pope in the puffy coat, return to a tool like my journey.

Di or stable diffusion majority did not respond to request for comment for this story. OpenAI, which makes dali, and stability AI, which makes table diffusion, declined comment. But most of these AI image generators work pretty much the same way.

They're based on a technique called diffusion. Making an A I image generator takes a few steps. Step one, kicking an A I model, engineers can either create one or a pick one that already exists. But just having the model itself isn't all that useful. They need to train IT that takes us to step to gathering training images.

I think the reason you are seeing generated images is because over the last twenty years, we need you and everybody else has uploaded billions and billions of pieces of content that the machines are learning from .

that tony for read again, the computer science professor at U. C. berkeley. He says that algorithms trolled the internet and scraped billions of images, each of which had text captions.

That last part is really important. So now what has some five billion images and what IT does that takes an image with a caption, uh, five people sitting at a bar in nap of enjoying a nice carbery, badly. I have no idea what that captain came from IT.

Just, I love IT. okay? Engineers have got their training data now. It's time for step three training the algorithm they've designed. Exactly how this happens can vary a bit.

But generally using that training data, the AI system learns to associate certain visual features with certain words. Let's say, the engineers feed that system an image of people drinking carbery at a bar. It'll learn things like cabin is red and the bar is flat. The AI system adds noise to that image by changing the color of each pixel that makes the shapes in the image difficult to distinguish. The way this noise is making IT hard to make up the words i'm .

saying that takes the image and IT adds a little bit of noise to IT degrades the image and then learn how to go backwards, how to d noise. And then IT doesn't until it's IT right? Degrade clean.

Degrade clean. IT does this five billion times. And then what he has is IT knows how to go from a pure noise image with a caption to a clean image that depicts what the caption is.

Billions of images times millions of the noises. The result is that we can type and a prompt, like two people drinking beer on the moon and in the span of just a few seconds, the A, I can generate anything we want, even if it's never existed before. Okay, well, maybe not anything. Sometimes the system just doesn't understand the prompt. And many of the popular image generators have rules barring them from producing certain kinds of images, though people can work around those.

So you can put Gabriel on what s in. You can put guardias on the prompt. You're not allowed to ask for nudd or violence or blood.

Even with gardens in place, it's easy to simply type a prompt and get a realistic image in a matter of seconds. Clear liberte is head of the AI and media integrity program at the partnership on AI. That's a nonprofit focused on responsible use of artificial intelligence. Its funding comes from a combination of philanthropic organizations and corporate stakeholders, including big tech companies.

The ease with which people can create very easily and sophisticated and kind of photo realistic examples of content that could be confusing or deceptive or cast doubt on entire narratives or histories make IT really important for us to understand where content comes from in and how it's been manipulated.

It's not just that we can make them quickly. We can share them quickly. An image is seen by two eyeballs and then millions in a matter of minutes.

And liberal tz says this comes with consequences. This happened in a very real way. Back in may.

a few months ago, a image of the pentagon was on fire. IT seemed to be on fire.

a viral fake picture, followed by a real dip in the markets. Monday, the photo posted to social media seeming to show a fire explosion near the pentagon.

General A I. Experts said the image was probably made by A I, but that didn't stop the markets from reacting. Shortly after opening on may twenty second, the deal fell two hundred nineteen points, only to rebound later that morning. But some viewer said the image wasn't without its flaws.

See how the grass blends into the concrete here, part of that black pole disappears behind this barrier, and the fense itself looks off.

But libert says IT was a big deal that .

is deeply impacted and that IT moved the market and was really misleading and confusing.

And honey fried, the computer science professor, says that because of how the image was shared on social media, people weren't skeptical enough.

They got retweet and rehab value climates. And then, of course, IT rebounded. The power of visual imagery is incredible.

and the amount of AI generated imagery will only increase from here in twenty twenty two, europol, the E. S. Law enforcement agency, released a report. IT predicted that by twenty twenty six, ninety percent of online content may be synthetically generated. Okay, so AI generated images can cause real harm, especially if people think they're real. But how can we tell the difference between what's real and what's made by AI? Charlet walks through the technological solutions to the AI image problem after the break.

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through .

mod a help. Luckily.

there's a faster, easier, less messy choice. Amazon q can security understand your business data and use that knowledge to streamline task? Now you can summarize quarterly results or do complex analysis in no time. Q, got this. Learn what amazon q business can do for you at A W S dot com flash.

Learn more. So we've already been dealing with this problem. A I generated images are getting passed off as the real thing. And although we've been told not to trust everything we see online, some people call this media literacy. Most of us could stand to be a bit more skeptical, but clearly, bots from the partnership on A I says that's not enough these days.

I think that concept gets touted as the total panache for all of our chAllenges that if I just educate lay people about the capacity to IP late content, they will be totally powered to gauge of something real or true or not to be media literate is very complex today.

he says we need new solutions. There are a couple of different options here. One is eye detection or analysis after the fact.

Another approach is invisible water marking or fingerprinting. And finally, there's something called a content credential. Let's start with the most obvious approach. Test the image after it's been published when we find IT online, that's solution number one. And companies are working on IT.

Hi um my name is andrey. Don't on each and i'm the CEO of company called optic.

A number of companies make tools that can tell us if an image was generated by AI up has one called A I or not.

It's a dead simple service and it's all in this name, A I or not dot com. So what he does is that tells you whether or not the image that you are uploading to IT is A I generated or human generated. Is that simple? All you do is you have dragon drop, you are upload your file and then you get the result.

Optic has been around since march twenty twenty two, and dana chiv says that its tech can find things in an image that the human eye can't. Certain patterns backed into the code at the pixel level. These inconsistencies or artifacts indicate that an images made by A I neural nets are the algorithms A I generators are based .

on new or less. Every time they generate an image, they might try as much as they can to not leave any artifacts. But if you look at the vast number of image is generated by certain neural as a human, you will notice those patterns. But if you do IT with another AI that is extremely good at pet n recognition, IT will start noticing tiny, little pixel level artifacts that those neural ess leave behind.

In the free version of A I or not, a user can uploaded an image they find on the internet and get a binary result generated by AI or not, but it's not always that simple. The paid version of objects tool gives a lot more information, and deron ith v says some companies need a higher level of detail. He says he's heard from some of them, including insurance companies, that need to determine if an image of a house on fire is actually a house on fire.

Another vertical that is interesting is all source of people like banks and know your customer kind of flows where you upload you or driving license. So where you upload your photo, we're seeing a lot of interest from dating sites where people want to make sure that the human that they looking at as actual human.

The problem with services like A I or not is they're not perfect. As of August twenty forth, the company says its tool is between ninety seven point eight percent and ninety nine point nine percent accurate at detecting A I generated images from the most popular neural nets.

There's a bunch of services like this, and everyone can claim that, you know, they're good. Uh, the devil comes in in details. So there are pockets of types of images, uh, where, like some detectors works well and some failed, ourselves included.

Da chip says, getting stuff wrong just means that the tool can get Better. Engineers have to constantly retrain detection systems. Anyway, since A I image generation are always getting Better right now.

there is no way to stay ahead. There's only way to stay as little behind as possible.

How do we stay as little behind as possible? A dobie dino says that the only way to stay ahead is to start from the beginning.

So by the time you attach a label to rely, millions of people have already seen IT and believed IT and believed IT incorrectly. And then you come back and tell them as a deep fake. It's too late. You can't unring that bell.

What if, instead of trying to figure out if an images fake, the image told the user IT was fake, that has to be baked in from the moment the images made, that the idea behind solution number two, invisible water marking, clear lebowitz from the partnership on A I says, water marking is basically adding a little piece of code to in image that allows detectors to verify IT.

Very simplistic water marking, you can think of as inserting some marker into a piece of content to say that IT is that is so the generator can purposeful .

ly change some information embedded in the pixel in a way that's invisible to the human eye e, but visible to an AI detector. These pixel act like a stamp or watermark. Microsoft being and google deep mind have already added these to their generators. Last month, google deep mind launched a new type of watermark called since I D push me in a vice president of research, a google deep mind .

told us sinai is a new way of doing watermark. IT has two properties. One is invisible, the human eye, so IT does not affect the quality of the image, and to is persistent in the sense that even if the waterman image is later transformed in somewhere like IT is cropped a bit, or IT is rotated or transformed in some other way, you will still be able to detect that IT was watermelon. Its agenda content.

cole says for deep mind, a watermark was the right approach because its very hard to strip off, though not impossible for now. Since ID is in beta testing and IT could be integrated into other google products.

google and google demand, we believe that A I is a transformative technology. IT IT will have impact on many different areas, but at the same time, when we are thinking about this powerful technology, you have to approach IT over a lot of sort of care and a and sense of .

responsibility. In july, tech companies including google, meta, microsoft and amazon met with president biden. They agreed to voluntary safeguards around the use of A I, including water marking A I generated images.

The companies have a duty to earn the people's trust and a power users to make informed decisions, labelling content that has been altered or A I generated. And this week, eight more big tech companies, including adobe and IBM, made the same commitment. But some academics and computer scientists are skeptical that these Marks can stand up to tampering from bad actors. Clearly, libert says a fingerprint is similar to a watermark.

A fingerprint is kind of like my fingerprint is not added. It's something that helps people identify me. There will be some database to sancy. Is that the fingerprint that batches?

K, imagine I went to a police station and got my fingerprints taken. If I end up in the system again, the police can find me a database. That's exactly the kind of database that could exist for AI generated images.

Having a watermark baked into the code or a fingerprint that connects to a base is actually only half the battle. The other half is making sure users have this information when they're looking at the image. Enter solution number three, the content credential.

This takes us back to the folks at adobe. Adobe is one of the founding members of a group called the content authenticity initiative, is a global coalition with members from tech, media and policy that, among other things, is working to create an industry standard for online images. Dinner row says its solution is something called a content credential.

So what a content credential is it's like a nutrition label for an image or video or audio. And what IT will telling you is who made the image when IT was made, where IT was made, and what edits were made to IT along the way.

a nutrition label is a good analogy. A content credential lays out everything that went into the image we're seeing, who took the initial photograph and where, if any, adjustments were made to IT, whether those adjustments were made using A I or maybe some gentle cropping and finally, where IT was published and by whom. The overall idea is to have a record of everything that happened to the image from when I was made to when the user sees IT online.

And it's an indicator of trust. That's how we think about IT. It's a chain of trust built from from the very first place image is captured to work as publish the content .

authenticity ity initiative is made up of more than fifteen hundred members. The wall street journal is one of them. And we should note that honey for red, the computer science professor we've been hearing from, he works with the initiative as a paid adviser. He also works with linked in, which is owned by another member of the initiative, microsoft. Andrew jank, microsoft director of media experience, says it's being A I generator now includes information in the meta data to indicate if an image was .

generated by AI. What that actually means is that we're adding the ation that an image was created by the gender of a eye system. We believe it's important that we tell you that IT came from the being service when that happened and some identifying information that lets you know what model was.

But seeing that indication isn't easy. IT takes a couple of steps right now. Users have to go to a whole separate website. The content authenticity initiative makes one called verify. But down the line, the goal is to show this information with the image itself.

I hope, is that you can tell which ones were created and which ones were captured, so that you continue to have a relationship with reality.

There's another benefit to content credentials. Remember what we said about optics? A I or not tall. Most of the images we see online aren't either A I generated or not.

Say you take a family photo at niaga pales, and you use your phones object remover function to raise the tourists in the background. You and your family did go to niaga palls that parts real, but A, I filled in the gaps where the other tourists were removed. The content credential would keep a record of all that.

We have a whole range of tools that can use A I type functionality to change an authentic images somewhere. Is that thing purely authentic? no.

Is IT purely created? no. And so what I like to think of this is continuing, right? How do you express different points along that continuum .

for adobe dinner around the contact credential gives the user all the information they need to make .

their own assessment. The thesis of this approach is always that we are a power, the public, to decide what to trust.

There is a major obstacle. In order for content credentials to be effective, they have to be a consistent standard across all. Figuring out how to do that is part of clear lebow tz. His job at the partnership on A I SHE says, for IT to work, everyone or nearly everyone has to participate.

We kind like to differentiate three buckets. There's these folks building the technology, the code, the models that ultimately would allow a creator to create with IT, then their creators themselves, then there's distributors.

And that is a huge area of intervention, which is how should tiktok or facebook or twitter ex, excuse me, ultimately spread that content? How do they convey that I might be manipulated and label IT? So everyone's involved.

We reached out to tiktok, meta and x, formerly twitter. Tiktok declined to comment and referred as to its participation in the partnership on A S framework for responsible practices for synthetic media and its synthetic media policy. Meta and x didn't respond to a request for comment.

Getting everyone on board is one of the major goals of the content authenticity initiative. Assuming its successful, our experience of seeing images online will change. We will have a lot more information about every image from the moment I was snap by a camera or generated by AI to when it's published. We will have the tools to determine what to trust at our fingertips. But from a technological perspective, microsoft Andrew jank says no one solution will be enough.

There's really no silver bullet technology here. Content credentials is not a silver bullet. Detections not a silver bullet, what are markings is not a solar bullet. But when you take these things in combination, you start to make very robust, very well understood systems that can provide Better mitigation for people generally.

For you see berkeley's honey for red. This combination of technologies, along with user scepticism, is the way to go, and we Better get on IT because he says the future of the internet hangs in the baLance.

I think there's one of two scenarios, either we keep going down the road or going, which is this disobeying health. Cape of an internet is where everybody's living in their own echo chAmbers. I don't think that has to end up that way. I think we can start to write the ship.

The future of everything is a production of the wall street journal. Stephanie egan fitz is the editorial director of the future of everything. This episode was produced by me.

alex sola and me charlock garden burg, our fact checker is a partner of Nathan Michael level and just a fanton are our sound designers and wrote .

arthly music cater on nilsa is our supervising producer. I W A muslims is our development producer.

Scott sale si are the deputy editors, and philopater is the head of news audio for the world street journal. Like the show, tell your friends and leave us a five star review on your favor platform.

Thanks for listening.

Amazon q business is the new generative A I assistant from A W S, because many tasks can make business slow, as if waiting through mud a help. Luckily, there's a faster, easier, less messy choice. Amazon q can security understand your business data and use that knowledge to streamline task? Now you can summarize quarterly results or do complex analysis in no time.

Q, got this. Learn what amazon q business can do for you at AWS dot com. Slash, learn more.