The TED AI Show: Is AI destroying our sense of reality? with Sam Gregory

[00:00:00] TED Audio Collective

[00:00:08] Okay, picture this.

[00:00:10] We're in Miami.

[00:00:11] There's sun, sand, palm trees,

[00:00:14] and a giant open-air mall right on the water.

[00:00:17] Pretty nice, right?

[00:00:19] But then, late one Monday night in January of this year,

[00:00:22] things get weird.

[00:00:24] Cop cars swarm the mall, like dozens of them.

[00:00:28] I'm talking six city blocks shut down,

[00:00:30] lights flashing, people everywhere,

[00:00:32] and no one knows what's going on.

[00:00:34] The news footage hits the internet,

[00:00:36] and naturally, the hive mind goes into overdrive.

[00:00:40] Speculation, conspiracy theories are flying,

[00:00:43] and one idea takes hold.

[00:00:46] Aliens.

[00:00:48] Folks are dissecting grainy helicopter footage

[00:00:50] in the comments,

[00:00:52] zooming in, analyzing it frame by frame

[00:00:55] to find any evidence of aliens.

[00:00:58] So I thought, I'm a TikToker.

[00:01:01] What if I brought this online fever dream to life

[00:01:03] and shared it with the masses?

[00:01:05] Using the latest AI tools, I created a video.

[00:01:08] Towering shadowy figures, silently materializing

[00:01:12] amidst the flashing lights of the police cars.

[00:01:14] An alien invasion in the middle

[00:01:16] of Miami's bayside marketplace.

[00:01:19] Just a bit of fun, I thought.

[00:01:20] Some got the joke, a Miami twist on Stranger Things.

[00:01:25] But I watched as other people flocked

[00:01:26] into my comment section to declare this

[00:01:29] as bonafide evidence that aliens do in fact exist.

[00:01:34] Now you might be wondering, what actually happened?

[00:01:37] A bunch of teenagers got into a fight at the mall,

[00:01:40] the police showed up to break it up,

[00:01:42] and that's all it took to trigger this mass hysteria.

[00:01:46] It was too easy.

[00:01:48] Too easy to make people believe something happened

[00:01:51] that never actually happened.

[00:01:53] And that's kind of terrifying.

[00:01:58] I'm Bilal Al-Saddou, and this is the TED AI Show,

[00:02:01] where we figure out how to live and thrive in a world

[00:02:04] where AI has changed everything.

[00:02:13] This show is brought to you by Schwab.

[00:02:15] You're here because you like to keep a pulse

[00:02:18] on trends in technology.

[00:02:20] Well, now you can invest in what's trending,

[00:02:23] in artificial intelligence, big data, robotic revolution,

[00:02:27] and more with Schwab Investing Teams.

[00:02:30] It's an easy way to invest in ideas you believe in.

[00:02:33] Schwab's research process uncovers emerging trends,

[00:02:37] then their technology curates relevant stocks into themes.

[00:02:41] Choose from over 40 themes.

[00:02:44] Buy all the stocks in a theme as is,

[00:02:47] or customize to better fit your investing goals,

[00:02:50] all in a few clicks.

[00:02:52] Schwab Investing Teams is not intended

[00:02:55] to be investment advice or a recommendation

[00:02:58] of any stock or investment strategy.

[00:03:00] Learn more at schwab.com slash thematic investing.

[00:03:07] Support for TED Tech comes from Odoo.

[00:03:09] To put it simply, Odoo is built to save.

[00:03:12] Odoo saves time, Odoo saves money,

[00:03:15] but most importantly, Odoo saves businesses.

[00:03:17] That's right, Odoo's superhero software rescues companies

[00:03:21] from the perils of disconnected platforms,

[00:03:23] and Odoo's utility belt of user-friendly applications

[00:03:26] puts the power of total business management

[00:03:28] in the palm of your hand.

[00:03:30] Learn more at odoo.com slash TED Tech.

[00:03:32] That's O-D-O-O dot com slash TED Tech.

[00:03:36] Odoo, saving the world one business at a time.

[00:03:40] This episode is brought to you by Progressive.

[00:03:43] Most of you aren't just listening right now.

[00:03:45] You're driving, cleaning, and even exercising.

[00:03:48] But what if you could be saving money

[00:03:49] by switching to Progressive?

[00:03:51] Drivers who save by switching save nearly $750 on average,

[00:03:55] and auto customers qualify for an average of seven discounts.

[00:03:59] Multitask right now.

[00:04:00] Quote today at progressive.com.

[00:04:02] Progressive Casualty Insurance Company and Affiliates,

[00:04:05] national average 12-month savings of $744

[00:04:08] by new customers surveyed who saved with Progressive

[00:04:10] between June 2022 and May 2023.

[00:04:13] Potential savings will vary.

[00:04:15] Discounts not available in all states and situations.

[00:04:20] Hackers and cyber criminals

[00:04:22] have always held this kind of special fascination.

[00:04:25] Obviously I can't tell you too much about what I do.

[00:04:28] It's a game.

[00:04:29] Who's the best hacker?

[00:04:30] And I was like, well, this is child's play.

[00:04:33] I'm Dena Temple-Reston,

[00:04:35] and on the Click Here podcast, you'll meet them

[00:04:37] and the people trying to stop them.

[00:04:39] We're not afraid of the attack.

[00:04:41] We're afraid of the creativity

[00:04:42] and the intelligence of the human being behind it.

[00:04:45] Click here.

[00:04:46] Stories about the people making

[00:04:47] and breaking our digital world.

[00:04:50] AI machines.

[00:04:51] Satellite.

[00:04:52] Engine ignition.

[00:04:53] Click here.

[00:04:54] And lift off.

[00:04:55] Click here.

[00:04:56] Every Tuesday and Friday, wherever you get your podcasts.

[00:05:01] So I've been making visual effects,

[00:05:02] blending realities on my computer since I was a kid.

[00:05:06] I remember watching a show called Mega Movie Magic,

[00:05:09] which revealed the secrets behind movies' special effects.

[00:05:12] I learned about CGI and practical effects in movies

[00:05:15] like Star Wars, Godzilla, and Independence Day.

[00:05:19] I was already into computer graphics,

[00:05:21] but seeing how they could create visuals indistinguishable

[00:05:24] from reality was a game changer.

[00:05:27] It sparked a lifelong passion to blend the physical

[00:05:29] and digital worlds.

[00:05:32] Several years ago, I started my TikTok channel.

[00:05:35] I'd upload my own creations and share them with hundreds,

[00:05:37] then thousands, and now millions of viewers.

[00:05:40] I mean, just five years ago,

[00:05:41] if I wanted to make a video

[00:05:43] of giant aliens invading a mall in Miami,

[00:05:45] it would have taken me a week

[00:05:47] and at least five pieces of software.

[00:05:49] But this aliens video?

[00:05:51] It took me just a day to make using tools like Midjourney,

[00:05:55] RunwayML, and Adobe Premiere,

[00:05:57] tools that anyone with a laptop can access.

[00:06:02] Since ChatGPT came on the scene in late 2022,

[00:06:05] there's been a lot of talk about the Turing test,

[00:06:08] where a human evaluator tries to figure out

[00:06:10] if the person at the other end of a text chat

[00:06:13] is a machine or another human.

[00:06:15] But what about the visual Turing test,

[00:06:18] where machines can create images

[00:06:20] that are indistinguishable from reality?

[00:06:23] And now OpenAI has come out with Sora,

[00:06:25] a video generation tool

[00:06:26] that will create impressively lifelike video

[00:06:29] from a single text prompt.

[00:06:31] It's basically like ChatGPT or DALI,

[00:06:34] but instead of text or images, it generates video.

[00:06:37] And don't get me wrong,

[00:06:38] there are other video generation tools out there,

[00:06:41] but when I first saw Sora, the realism blew my socks off.

[00:06:45] I mean, with these other programs,

[00:06:46] you can make short videos, like a couple seconds long,

[00:06:50] but with Sora, we're talking minute long videos.

[00:06:53] The 3D consistency with those long dynamic camera moves

[00:06:57] definitely impressed me.

[00:06:58] There's so much high frequency detail

[00:07:00] and the scene is just brimming with life.

[00:07:03] And if we can just punch in a single text prompt into Sora

[00:07:07] and it'll give us full on video

[00:07:09] that's visually convincing to the point

[00:07:10] that some people could mistake it for something real,

[00:07:14] well, you could imagine some of the problems

[00:07:15] that might stem from that.

[00:07:18] So we're at a turning point.

[00:07:19] Not only have we shattered the visual turning test,

[00:07:22] we're reshattering it every day.

[00:07:24] Images, audio, video, 3D, the list goes on.

[00:07:28] I mean, you've probably seen the headlines.

[00:07:31] AI generated nude photographs

[00:07:33] of Taylor Swift circulating on Twitter.

[00:07:35] A generated video of Vladimir Zelinsky

[00:07:37] surrendering to the Russian army.

[00:07:39] A fraudster successfully impersonating a CFO

[00:07:42] on a video call to scam a Hong Kong company

[00:07:45] out of tens of millions of dollars.

[00:07:47] And as bad as the hoaxes and the fakes

[00:07:49] and the scams are, there's a more insidious danger.

[00:07:52] What if we stop believing anything we see?

[00:07:55] Think about that.

[00:07:57] Think about a future where you don't believe the news.

[00:08:00] You don't trust the video evidence you see in court.

[00:08:03] You're not even sure that the person

[00:08:05] on the other end of the Zoom call is real.

[00:08:08] This isn't some far flung future.

[00:08:11] In fact, I'd argue we're living in it now.

[00:08:14] So given that we're in this new world

[00:08:16] where we're constantly shattering

[00:08:18] and reshattering the visual Turing test,

[00:08:20] how do we protect our own sense of reality?

[00:08:24] I reached out to Sam Gregory

[00:08:26] to talk me through what we're up against.

[00:08:28] Sam is an expert on generative AI and misinformation

[00:08:31] and is the executive director

[00:08:33] of the Human Rights Network, Witness.

[00:08:35] His organization has been working with journalists,

[00:08:38] human rights advocates, and technologists

[00:08:41] to come up with solutions that help us separate

[00:08:43] the real from the fake.

[00:08:47] Sam, thank you for joining us.

[00:08:49] I have to ask you, as we're seeing these AI tools proliferate

[00:08:53] just over the last two years,

[00:08:55] are we correspondingly seeing a massive uptick

[00:08:58] of these visual hoaxes?

[00:09:00] The vast majority are still these shallow fakes

[00:09:02] because anyone can make a shallow fake.

[00:09:04] It's trivially easy, right, just to take an image,

[00:09:06] grab it out in Google search,

[00:09:08] and claim it's from another place.

[00:09:09] What we're seeing though is this uptick

[00:09:11] that's happening in a whole range of ways

[00:09:15] where people are using this generative media for deception.

[00:09:17] So you see images

[00:09:19] sometimes deliberately shared to deceive people, right?

[00:09:22] Someone will share an image claiming it's, you know,

[00:09:24] of an event that never happened.

[00:09:26] And then, you know, we're seeing a lot of audio

[00:09:28] because it's so trivially easy to make, right?

[00:09:30] A few seconds of your voice

[00:09:32] and you can churn out endless, endless cloned voice.

[00:09:37] We're not seeing so much video, right?

[00:09:39] And that's, you know, a reflection that, you know,

[00:09:42] really doing complex video recreation

[00:09:45] is still not quite there, right?

[00:09:47] Yeah, video is significantly harder,

[00:09:50] at least for the moment.

[00:09:52] And I personally hope that it would stay pretty hard

[00:09:54] for a while, though some of these generations

[00:09:57] are getting absolutely wild.

[00:09:59] I had a bit of an existential moment

[00:10:01] looking at this one video from Sora.

[00:10:04] It's the underwater diver video.

[00:10:06] For anyone who hasn't seen this,

[00:10:09] there's a diver swimming underwater, you know,

[00:10:12] investigating this historic,

[00:10:14] almost archaeological spaceship

[00:10:17] that's crashed into the waterbed.

[00:10:19] And it looked absolutely real.

[00:10:21] And I was thinking through what that would have taken

[00:10:23] for me to do the old-fashioned way

[00:10:25] and I was just gasping

[00:10:27] that this was just a simple prompt

[00:10:29] that produced this immaculate one-minute video.

[00:10:33] I'm kind of curious, have you had such a moment yourself?

[00:10:36] It's funny because I was literally showing

[00:10:38] that video to my colleagues

[00:10:40] and I didn't cue them up that it was made with Sora

[00:10:44] because I wanted to see whether they clicked

[00:10:46] that it was an AI-generated video

[00:10:48] because I think it's a fascinating one.

[00:10:50] It's kind of on the edge of possibility.

[00:10:52] There's definitely a kind of a moment

[00:10:54] that's happening now for me

[00:10:55] and it's really interesting

[00:10:56] because, you know, we first started working on this

[00:10:58] like five or six years ago

[00:10:59] and we were just doing what we described as

[00:11:01] prepare, don't panic

[00:11:03] and really trying to puncture people's hype,

[00:11:04] particularly around video deepfakes

[00:11:06] because people kept implying

[00:11:08] that they were really easy to do

[00:11:10] and that we were surrounded by them

[00:11:11] and the reality was

[00:11:12] it wasn't easy to fake, you know, convincing video

[00:11:16] and to do that at scale.

[00:11:17] So it's certainly for me,

[00:11:18] Sora has been a click moment

[00:11:20] in terms of the possibility here

[00:11:21] even though it feels like a black box

[00:11:23] and I'm not quite sure how they've done it

[00:11:25] and how accessible this is actually going to be

[00:11:26] and how quickly.

[00:11:27] So related to this,

[00:11:29] a lot of these visual hoaxes

[00:11:31] tend to be whimsical, even innocuous, right?

[00:11:33] In other words, they don't cause serious harm

[00:11:36] in the real world

[00:11:37] and are almost akin to pranks

[00:11:39] but some of these visual hoaxes

[00:11:41] can be a lot more serious.

[00:11:44] Can you tell me a little bit about

[00:11:45] what you're seeing out there?

[00:11:47] The most interesting examples right now

[00:11:49] are happening in election context globally

[00:11:51] and they're typically

[00:11:54] people having words put in their mouths.

[00:11:56] In the recent elections in Pakistan,

[00:11:58] in Bangladesh,

[00:11:59] you had candidates saying

[00:12:01] boycott the vote

[00:12:02] or vote for the other party, right?

[00:12:04] And they're quite compelling

[00:12:06] at a first glance,

[00:12:07] particularly if you're not very familiar

[00:12:08] with how AI can be used

[00:12:10] and they're often deployed

[00:12:11] right before an election.

[00:12:12] So those are clearly

[00:12:14] in most cases malicious,

[00:12:15] they're designed to deceive

[00:12:17] and then you're also seeing ones

[00:12:18] that are kind of these leaked conversation ones

[00:12:20] so they're not visual hoaxes

[00:12:22] and so you've got really,

[00:12:23] you know,

[00:12:24] quite deceptive uses happening there

[00:12:26] either directly just with audio

[00:12:28] or at the intersection of

[00:12:29] audio with animated faces

[00:12:31] or audio with the ability to make a lip sync

[00:12:33] with a video.

[00:12:35] If I wanted to ask you to zoom in on

[00:12:38] one single example

[00:12:40] that's disturbed you the most,

[00:12:41] something that exemplifies

[00:12:42] what you are the most worried about,

[00:12:45] what would it be?

[00:12:46] I'm going to pick one that is

[00:12:48] it's actually a whole genre

[00:12:50] and I'm going to describe this genre

[00:12:51] because I think it's the one

[00:12:52] that people are familiar with

[00:12:53] but once you start to think about it

[00:12:54] you realize how easy it is to do this

[00:12:56] and that is pretty much everyone has seen

[00:12:58] Elon Musk selling a crypto scam,

[00:13:00] right?

[00:13:01] Often paired up with a newscaster,

[00:13:03] your favorite newscaster

[00:13:04] or your favorite political figure

[00:13:06] in every country in which

[00:13:08] I work,

[00:13:08] people have experienced that.

[00:13:09] They've seen that video where it's like

[00:13:11] the newscaster says,

[00:13:12] Hey Elon, come on and explain

[00:13:14] how you follow this new crypto scam

[00:13:16] or come on political candidate

[00:13:17] and explain why you're investing

[00:13:19] in this crypto scam.

[00:13:21] For anyone who hasn't seen it,

[00:13:23] these are just videos

[00:13:24] with a deepfake Elon Musk

[00:13:25] trying to guilt you into buying crypto

[00:13:27] as a part of their

[00:13:28] Bitcoin giveaway program.

[00:13:31] And so the reason I point to that is

[00:13:33] not because it has massive human rights impacts

[00:13:35] or massive news impacts

[00:13:36] but it's just

[00:13:38] this is so commodified

[00:13:39] but we have this sort of bigger question

[00:13:41] of how it plays into our

[00:13:43] overarching understanding

[00:13:44] of what we trust,

[00:13:45] right?

[00:13:45] Does this undermine people's confidence

[00:13:47] in almost any way

[00:13:49] in which they experience

[00:13:50] audio or video or photos

[00:13:52] that they encounter online?

[00:13:53] Does it just reinforce

[00:13:54] what they want to believe

[00:13:56] and for other people just let them believe

[00:13:57] that nothing can be trusted?

[00:14:02] We're going to take a quick break.

[00:14:04] When we come back,

[00:14:04] we're going to talk with Sam about

[00:14:06] how we can train ourselves

[00:14:07] to better distinguish the real

[00:14:08] from the unreal

[00:14:10] using a little system he calls SIFT.

[00:14:12] More on that in just a minute.

[00:14:46] We're back with Sam Gregory of Witness.

[00:14:55] Before the break,

[00:14:56] we were talking about how these fake videos

[00:14:58] are starting to erode our trust

[00:15:00] in everything we see.

[00:15:02] And yeah,

[00:15:03] maybe you can find flaws

[00:15:04] in a lot of these videos,

[00:15:05] but some of them are really,

[00:15:07] really good.

[00:15:08] And nobody's zooming in at 300%

[00:15:10] looking for those minor imperfections,

[00:15:13] especially when they're scrolling

[00:15:14] through a feed,

[00:15:15] right?

[00:15:15] Like before their morning commute

[00:15:17] or something.

[00:15:18] Yeah, and you're hitting on the thing

[00:15:19] that I think,

[00:15:20] you know,

[00:15:21] the news media has often done a disservice

[00:15:23] to people about how to think about

[00:15:24] spotting AI, right?

[00:15:26] We put such an emphasis on kind of like,

[00:15:28] you know, you should have spotted

[00:15:29] the Pope, you know,

[00:15:30] had his ring finger on the wrong hand

[00:15:32] in that puffer jacket image, right?

[00:15:34] Or didn't you see that his hair

[00:15:36] didn't look quite right on the hairline?

[00:15:38] Or didn't you see he didn't blink

[00:15:40] at the regular rate?

[00:15:41] And it's just so cruel almost to us

[00:15:43] as consumers to expect us to spot those

[00:15:46] things. We don't do it.

[00:15:46] I don't look at every TikTok video

[00:15:48] in my, you know,

[00:15:49] and in my for you page and go like,

[00:15:51] let me just look at this really carefully

[00:15:53] and make sure if someone's trying to

[00:15:54] deceive me.

[00:15:54] And so we've done a disservice often

[00:15:56] because people point out these glitches

[00:15:58] and then they expect people to spot them

[00:16:00] and it's it's it creates this whole

[00:16:02] culture where we distrust everything

[00:16:03] we look at

[00:16:05] and we try and apply this sort of

[00:16:06] personal forensic skepticism

[00:16:08] and it doesn't lead us to great places.

[00:16:11] All right.

[00:16:12] I want to talk about mitigation.

[00:16:13] How do we prepare

[00:16:14] and what can we do right now?

[00:16:16] When we first started saying prepare

[00:16:17] don't panic.

[00:16:18] It was five or six years ago

[00:16:19] and it was in the first deepfakes hype cycle,

[00:16:22] which was like the 2018 elections

[00:16:24] when everyone was like deepfakes are going

[00:16:25] to destroy the elections

[00:16:27] and I don't think there was a single deepfake

[00:16:29] in the 2018 US elections of any note.

[00:16:32] Now, let's fast forward to now,

[00:16:34] right 2024 when we look around the world

[00:16:37] the threat is clear and present now

[00:16:39] and it's escalating.

[00:16:40] So prepare is about acting listening

[00:16:42] to the right voices

[00:16:43] and thinking about how we balance out

[00:16:45] creativity expression human rights

[00:16:47] and do that from a global perspective

[00:16:49] because so much this conversation often

[00:16:51] is also very US or Europe centric.

[00:16:55] So what can we do now?

[00:16:56] You know, the first part of it is

[00:16:58] who are we listening to about this

[00:16:59] and I often get frustrated in AI conversations

[00:17:02] could get this very abstract discussion

[00:17:04] around AI harms and AI safety

[00:17:06] and it feels very different

[00:17:08] from the conversation.

[00:17:09] I'm having with journalists

[00:17:10] and human rights defenders on the ground

[00:17:11] who are saying I got targeted

[00:17:13] with a non-consensual sexual deepfake.

[00:17:15] I got my footage dismissed

[00:17:17] as faked by a politician

[00:17:18] because he said it could have been made

[00:17:19] by AI.

[00:17:20] So as we prepare the first thing is

[00:17:22] how do we listen to right

[00:17:23] and we should listen to the people

[00:17:24] who actually are experiencing this

[00:17:26] and then we need to think

[00:17:27] what is it that we need to help people

[00:17:30] understand how AI is being used

[00:17:32] this kind of question of the recipe

[00:17:34] and I use the recipe analogy

[00:17:36] because I think we're not in a world

[00:17:38] where it's AI or not.

[00:17:39] It's even in the photos.

[00:17:40] We take on our iPhones.

[00:17:42] We're already combining AI and human right

[00:17:44] the human input

[00:17:45] then the AI modifications

[00:17:46] that make our photos look better.

[00:17:48] So we need to think

[00:17:49] how do we communicate

[00:17:50] that AI was used in the media we make

[00:17:54] we need to show people

[00:17:55] how AI and human were involved

[00:17:58] in the creation of a piece of media

[00:17:59] how it was edited

[00:18:00] and how it's distributed

[00:18:01] the second part of it

[00:18:02] is around access to detection

[00:18:04] can the thing that we've seen

[00:18:05] is there's a huge gap

[00:18:07] in access to the detection tools

[00:18:08] for the people who need it most

[00:18:10] like journalists and election officials

[00:18:12] and human rights defenders globally

[00:18:13] and so they're kind of stuck

[00:18:14] they get this piece of video or an image

[00:18:16] and they are doing the same things

[00:18:18] that we're encouraging ordinary people to do

[00:18:20] look for the glitches

[00:18:22] you know take a guess

[00:18:22] drop it in an online detector

[00:18:24] and all of those things

[00:18:26] are as likely to give a false positive

[00:18:29] or a false negative

[00:18:29] as they are to give a reliable result

[00:18:31] that you can explain

[00:18:32] so you've got those two things

[00:18:33] you've got an absence of transparency

[00:18:35] explaining the recipe

[00:18:36] you've got gaps in access to detection

[00:18:39] and neither of those will work well

[00:18:40] unless the whole of the AI pipeline

[00:18:44] plays its part in making sure

[00:18:45] the signals of that authenticity

[00:18:47] and the ability to detect is retained

[00:18:49] all the way through

[00:18:51] so those are the three key things

[00:18:52] that we point to

[00:18:52] is transparency done right

[00:18:54] detection available to those who need it most

[00:18:57] and the importance of having an AI pipeline

[00:18:59] where the responsibility is shared

[00:19:01] across the whole AI industry

[00:19:03] I think you covered like

[00:19:05] three questions beautifully right here

[00:19:07] so a key challenge is telling

[00:19:09] what content is generated by humans

[00:19:11] versus synthetically generated by machines

[00:19:14] and one of the efforts you're involved in

[00:19:16] is the appropriately named

[00:19:18] content authenticity initiative

[00:19:20] could you talk a bit about

[00:19:21] how does that play into a world

[00:19:22] where we will have

[00:19:23] fake content purporting to be real

[00:19:25] yes so

[00:19:26] about five years ago

[00:19:27] there was

[00:19:28] there were a couple of initiatives founded

[00:19:30] by a mix of companies

[00:19:31] and media entities

[00:19:32] and Witness joined those early on

[00:19:34] to see how we could bring a human rights voice to them

[00:19:36] and one of them was something called

[00:19:37] the content authenticity initiative

[00:19:38] that Adobe kicked off

[00:19:40] and another was something called

[00:19:41] the coalition for content provenance

[00:19:42] and authenticity

[00:19:43] the shorthand for that is C2PA

[00:19:46] so let me explain a little more about

[00:19:48] what C2PA is

[00:19:49] it's basically a technical standard

[00:19:51] for showing what

[00:19:52] we might describe as the provenance

[00:19:54] of an image or a video

[00:19:56] or another piece of media

[00:19:57] and provenance is

[00:19:58] basically the trail of how it was created right

[00:20:01] this is a standard

[00:20:01] that's being increasingly adopted by platforms

[00:20:04] in the last couple of months

[00:20:05] you've seen Google and Meta

[00:20:06] adopted as a way they're going to show to people

[00:20:08] how the media they encounter online

[00:20:11] particularly AI generated

[00:20:12] or edited media was made

[00:20:14] it's also a direction that governments are moving in

[00:20:16] some key things that

[00:20:17] that we point to

[00:20:18] around standards like the C2PA is

[00:20:21] you know the first thing is

[00:20:22] they are not a foolproof way of

[00:20:24] showing whether something was made with AI

[00:20:26] made with a human

[00:20:27] what I mean by that is

[00:20:28] they tell you information

[00:20:30] but you know we know that people can remove

[00:20:31] that metadata for example

[00:20:33] they can strip out the metadata

[00:20:34] and we also know that some people may not

[00:20:36] add this in

[00:20:37] for a range of reasons

[00:20:38] so we're creating a system

[00:20:40] that allows additional signals of

[00:20:43] of trust or additional pieces of information

[00:20:45] but no

[00:20:46] one confirmation of authenticity

[00:20:49] or reality

[00:20:51] I think that's really important

[00:20:52] that we be clear

[00:20:53] that this is in some sense

[00:20:54] a harm reduction approach

[00:20:56] it's a way to get people more information

[00:20:58] but it's not going to be conclusive

[00:20:59] in a kind of sort of silver bullet like way

[00:21:03] and then the second sort of thing

[00:21:04] that we need to think about these is

[00:21:05] you know we need to really make sure

[00:21:07] that this is about the how of how media was made

[00:21:09] not the who of who made it

[00:21:11] otherwise we open a back door to surveillance

[00:21:14] we open a back door to the ways

[00:21:15] this will be used

[00:21:16] to target and criminalize journalists

[00:21:19] and people who speak out against governments globally

[00:21:21] beautifully said

[00:21:22] especially in the last point

[00:21:23] I noticed Tim Sweeney had some interesting remarks

[00:21:25] about all of the content authenticity

[00:21:27] initiatives happening

[00:21:28] as kind of described it as sort of

[00:21:31] surveillance DRM

[00:21:32] where you cannot upload a piece of content,

[00:21:35] right?

[00:21:35] Like if people like you aren't pushing on this direction

[00:21:38] we may well end up in a world

[00:21:39] where you cannot upload imagery

[00:21:41] onto the internet without

[00:21:42] having your identity tied to it

[00:21:44] and I think that would be a scary world indeed

[00:21:46] the thing that we have consistently pushed back on

[00:21:49] in systems like c2pa

[00:21:51] and is on the idea

[00:21:53] that identity should be the center

[00:21:54] of how you're trusted online

[00:21:56] it's helpful right

[00:21:57] and in many times

[00:21:57] I want people to know who I am

[00:21:59] but if we start to premise

[00:22:01] trust online

[00:22:02] in individual identity as the center

[00:22:04] and require people to do that

[00:22:06] that brings all kinds of risks

[00:22:08] that we already have a history of understanding

[00:22:10] from social media,

[00:22:10] right?

[00:22:11] That's not to say we shouldn't think about things like

[00:22:13] proof of personhood

[00:22:14] right?

[00:22:14] Like how do we understand

[00:22:15] that someone who created media was a human

[00:22:18] may be important

[00:22:19] right?

[00:22:19] As we enter an AI generated world

[00:22:21] that's not the same as knowing that it was Sam

[00:22:23] who made it

[00:22:23] not a generic human who made it

[00:22:25] right?

[00:22:26] So I think that's really important.

[00:22:28] It's a slippery slope indeed

[00:22:29] and really good point on sort of the distinction

[00:22:31] between validating your

[00:22:33] a human being

[00:22:34] versus

[00:22:35] you know validating you are Sam Gregory.

[00:22:37] That's a there.

[00:22:38] It's a very subtle but you know,

[00:22:40] crucial distinction.

[00:22:42] Let's move over to fears and hopes,

[00:22:44] you know back in 2017.

[00:22:46] You felt the the fear around deepfakes were overblown

[00:22:48] clearly now it is is far more of a clear

[00:22:51] and present danger.

[00:22:52] Where do you stand now?

[00:22:53] What are your hopes and fears at the moment?

[00:22:56] So we've gone from a scenario in 2017

[00:22:59] where the the primary harm

[00:23:02] was the one that people didn't discuss

[00:23:04] that was gender-based violence

[00:23:07] and the harm everyone discussed political

[00:23:10] usage was non-existent

[00:23:12] to a scenario now where the gender-based violence

[00:23:15] has got far worse

[00:23:16] right? And targets everyone from public figures

[00:23:19] to teenagers in schools all around the world

[00:23:22] and the political usage is now very real.

[00:23:24] And the third thing is you have people realizing

[00:23:26] there's this incredibly good excuse

[00:23:28] for a piece of compromising media,

[00:23:30] which is just to say hey that was faked

[00:23:32] or hey plausibly I can deny

[00:23:35] that piece of media by saying that it was

[00:23:37] faked. And so those three are the sort of the core

[00:23:39] fears that I experienced now that have translated

[00:23:42] into reality.

[00:23:43] Now in terms of hopes,

[00:23:45] I don't think we've acted yet on those three

[00:23:48] core problems sufficiently right?

[00:23:50] We need to address those and we need to make

[00:23:52] sure that you know, we criminalize

[00:23:55] the ways in which people target primarily women

[00:23:57] with non-consensual sexual deepfakes

[00:24:00] which are escalating. In the second area of

[00:24:02] fears, which is the fears around their misuse

[00:24:04] to in politics and to undermine

[00:24:07] news footage and human rights content.

[00:24:09] I think that's where we need to lean into a lot of

[00:24:12] the approaches like the authenticity

[00:24:14] and provenance infrastructures like the C2PA

[00:24:17] the access to detection tools for

[00:24:19] the journalists who need it most and then

[00:24:21] smart laws that can help us rule

[00:24:24] out some usages right and make sure that it is

[00:24:26] clear that some uses are unacceptable.

[00:24:28] And then the third area, that's the hardest one

[00:24:30] because we just don't have the research yet

[00:24:33] about what is the impact of this constant

[00:24:35] sort of drip, drip, drip if you can't

[00:24:38] believe what you see in here.

[00:24:40] We can only reach an 84% probability

[00:24:42] that it's real or false which is not great

[00:24:44] for public confidence.

[00:24:46] But we also don't know how this plays into this

[00:24:49] broader societal trust crisis

[00:24:51] we have where already people want to

[00:24:54] lean into kind of almost plausible believability

[00:24:56] on stuff they care about or just plausibly ignoring

[00:24:59] anything that you know challenges those beliefs.

[00:25:01] I think you brought up a really good point about

[00:25:03] it's almost like the world is fracturing into

[00:25:05] the multiverse of madness.

[00:25:06] I like to call it where people are looking for

[00:25:08] whatever validation to sort of confirm

[00:25:11] their beliefs at the same time it can

[00:25:13] it can result in people being jaded right where

[00:25:15] they're just going to be detached.

[00:25:17] Well, I don't trust anything.

[00:25:19] And so I'm curious how do you see consumers

[00:25:21] behaviors changing in this world

[00:25:24] where the visual Turing test gets shattered

[00:25:27] over and over again for all sorts of different

[00:25:29] more complex domains?

[00:25:31] Are people going to get savvier?

[00:25:33] What do you think is going to happen to society

[00:25:35] in such a world?

[00:25:37] So we have to hope that we walk a fine line.

[00:25:40] We're going to need to be more skeptical

[00:25:42] of audio and images and video that we encounter

[00:25:44] online, but we're going to have to do that

[00:25:47] with a skepticism that supported by signals

[00:25:49] that help us what I mean by that is if we

[00:25:51] enter a world where we just like hey everyone

[00:25:53] everything could be faked.

[00:25:55] It's getting better every day.

[00:25:56] Hey look out for the glitch then we enter

[00:25:59] a world where people skepticism quite rightly

[00:26:02] will accelerate because all of us will experience

[00:26:04] like on a daily basis being deceived, right?

[00:26:06] And I think that's very legitimate for us to

[00:26:09] then feel like we can't trust anything right

[00:26:12] in the ideal world.

[00:26:13] Everyone's labeling what's real or fake,

[00:26:16] but when that's not happening,

[00:26:18] what do people do?

[00:26:20] I always go back to you know, basic media literacy.

[00:26:22] I use an acronym called sift that was invented

[00:26:25] by an academic called Mike Caulfield and sift

[00:26:28] is s I F T s stands for stop right because it's

[00:26:33] basically stop before you're emotionally triggered

[00:26:35] right whenever you see something that's too good

[00:26:37] to be true.

[00:26:38] I stands for investigate the source which is

[00:26:41] like who shared this is is it someone I should

[00:26:45] trust the F stands for find alternative coverage

[00:26:48] right did someone already write about this and

[00:26:50] say wait, that's not the pope in a puffer jacket

[00:26:52] in reality.

[00:26:52] That's an AI image and then the fourth part of

[00:26:55] that which is getting complicated is T for trace

[00:26:58] the original which used to always be a great way

[00:27:01] of doing it in the shallow fake era because

[00:27:03] you'd find that an image had been recycled but

[00:27:05] is getting harder now.

[00:27:07] So what I look at the sort of the knife edge.

[00:27:08] We've got to walk.

[00:27:09] It's to help people do sift in an environment

[00:27:12] that is structured to give them better signals

[00:27:13] of how I was used and where the law has set

[00:27:17] parameters about what is definitely not acceptable

[00:27:20] and where all the companies all the players in

[00:27:22] that AI pipeline are playing their part to make

[00:27:25] sure that we can see how the recipe of how our

[00:27:27] AI and human was used and that it's as easy as

[00:27:30] possible to detect when AI was used to manipulate

[00:27:33] or create a piece of imagery audio or video.

[00:27:36] I really like sift.

[00:27:37] I think that's also very good advice for people

[00:27:39] when they come across something that is indeed

[00:27:41] too good to be true.

[00:27:42] Very often we will be like, oh, well, that's

[00:27:45] interesting and go about our day.

[00:27:47] The devices we use every day aren't foolproof,

[00:27:50] right?

[00:27:50] They've got vulnerabilities.

[00:27:51] There is this game of whack-a-mole that happens

[00:27:54] with patching those vulnerabilities and now

[00:27:56] we've got these cognitive vulnerabilities almost

[00:27:58] and you know on the detection side the tools

[00:28:01] are going to need to keep improving because

[00:28:03] people are going to find ways to use the detectors

[00:28:05] to create new generators that evade them, right?

[00:28:08] And so that game of whack-a-mole will continue

[00:28:10] but that isn't to say that all hope is lost.

[00:28:13] We can adapt and we can still have an information

[00:28:17] landscape where we can all thrive together.

[00:28:20] That's the future.

[00:28:21] I want the way we describe it a witness.

[00:28:23] We talk about fortifying the truth, which is

[00:28:26] that we need to find ways to defend that there

[00:28:28] is a reality out there.

[00:28:30] Thank you so much, Sam.

[00:28:31] I will certainly sleep easier at night knowing

[00:28:33] there are people like you out there making sure

[00:28:35] we can tell the difference between the real and

[00:28:37] unreal.

[00:28:38] Thank you so much for joining us.

[00:28:42] Sam Gregory and I had this conversation in mid-March

[00:28:45] and a few days later there was another development.

[00:28:48] YouTube came out with a new rule.

[00:28:50] If you have AI generated content in your video

[00:28:53] and it's not obvious, you have to disclose its

[00:28:55] AI. This move from YouTube is an important one.

[00:28:59] The kind Sam and his colleagues at Witness have

[00:29:01] been advocating for.

[00:29:02] It shifts the onus onto creators and platforms

[00:29:06] and away from everyday viewers because ultimately

[00:29:09] it's unfair to make all of us become AI detectives

[00:29:13] scrutinizing every video for that missing shadow

[00:29:15] or impossible physics, especially in a world where

[00:29:18] the visual Turing test is continually being shattered.

[00:29:22] And look, I'm not going to sugarcoat this.

[00:29:24] This is a huge problem and it's going to be difficult

[00:29:27] for everyone. Folks like Sam Gregory have their

[00:29:30] work cut out for them and massive organizations

[00:29:33] like TikTok, Google and MetaDoo too.

[00:29:35] But listen, I'm going to be back here this week

[00:29:38] and the week after that and the week after that

[00:29:41] helping you figure out how to navigate this new

[00:29:43] world order. How to live with AI and yes, thrive

[00:29:46] with it too. We'll be talking to researchers, artists,

[00:29:50] journalists, academics who can help us demystify

[00:29:53] the technology as it evolves. Together, we're going

[00:29:57] to figure out how to navigate AI before it navigates

[00:30:00] us. This is the TED-AI show. I hope you'll join

[00:30:04] us. The TED-AI show is a part of the TED Audio

[00:30:09] Collective and is produced by TED with Cosmic

[00:30:12] Standard. Our producers are Ella Fetter and

[00:30:15] Sarah McCray. Our editors are Ben Bensheng and

[00:30:18] Alejandra Salazar. Our showrunner is Ivana Tucker

[00:30:22] and our associate producer is Ben Montoya. Our

[00:30:26] engineer is Asia Pilar Simpson. Our technical

[00:30:29] director is Jacob Winnick and our executive

[00:30:32] producer is Eliza Smith. Our fact-checker is

[00:30:35] Christian Apartha and I'm your host Bilal El-Siddiou.

[00:30:39] See y'all in the next one.