‘We Can’t Be Certain’: Blumenthal Warns Biden Deepfake Has Made Unknown ‘Impacts’

8 months ago

Earlier this month, Sen. Richard Blumenthal (D-CT) questioned experts on the effects of recent deep fakes of President Biden during a Senate Judiciary Committee hearing.

Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:

https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript

Stay Connected
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com

Transcript

00:00 Thank you very much. Thanks to all of our witnesses for your really excellent

00:05 testimony. Mr. Scanlon, you hypothesize, as we can't know for sure, that the Biden

00:17 deepfake had minimal impact, but we can't be certain what the vote would have been

00:26 but for those calls. And I understand there is an investigation ongoing. The

00:36 Attorney General is conducting it. It's under New Hampshire law. I assume it's

00:42 criminal law as well as civil, but there are no federal remedies. In your view,

00:49 would it be helpful to have criminal penalties under federal law specifically

00:56 aimed at this kind of deception? And I think it was Mr. Coleman suggested that

01:03 criminal penalties could be an effective deterrent, but they have to be really

01:08 more specific and stringent than they are now. Mr. Chairman, I have to agree

01:14 that we truly don't know what the impact was on the New Hampshire presidential

01:20 primary. We only know that we had a good turnout and the results were what

01:26 they were. And we still have an active prosecution going on. The AG in New

01:34 Hampshire has identified a company or companies that participated and an

01:41 individual that is a suspect and and they're moving forward with that. At

01:49 some point, I believe that there is a federal component to this because it's

01:56 going to be a national problem. And I'd like to give a shout out to Kate Conley

02:02 who works with Jen Easterly at CISA. Kate was in New Hampshire on the day of the

02:09 presidential primary and she traveled around to polling places with me to try

02:14 and get a handle on, you know, how big this thing actually was, even though that

02:18 was difficult to determine. But yes, I think that, you know, these things in a

02:26 national election are going to be generated nationally, whether it's

02:32 foreign actors or some other malicious circumstance. And I think we need

02:36 uniformity and the power of federal government to help put the brakes on

02:41 that. Instances that happen locally, certainly government, federal government

02:46 assistance would be helpful, but I think that should remain the prerogative of

02:50 state law enforcement and the Attorney General. With assistance from federal

02:55 authorities where it's appropriate, let me ask you and the other witnesses,

02:59 Senator Hawley and I have proposed a framework which includes an independent

03:06 oversight entity, a set of standards that would be imposed by that entity, a

03:14 requirement for some licensing before models were deployed, testing to assure

03:21 that they were safe and effective, just as the FDA reviews drugs to make sure

03:27 they are safe and effective, and potentially penalties such as we've been

03:33 discussing, as well as export controls to assure that our national security is

03:38 protected. I'm assuming, just for the sake of speed, I'm assuming that all of you

03:47 would agree that some kind of framework like that one makes sense. I actually

03:54 have specific thoughts on that framework. I think it's a good start, but I really

03:57 think it's important that whatever framework we set adopts a, what's called

04:02 a defense in depth approach, right? So we need metadata, watermarking, cryptographic

04:07 hashing, which is a little complicated, but it's invisible watermarks and a hash

04:11 database, kind of like NCMEC, AI detection and AI poisoning. It also needs to cover

04:17 both the generative AI platforms and the online platforms. We need both of those

04:22 folks. We can't just say licensed generative AI companies and leave it at

04:25 that. Honestly, we need government buy-in, generative AI buy-in, platform buy-in,

04:30 journalist buy-in, and then detection companies. And all of those points are

04:34 encompassed by our framework, particularly the watermarking. Yeah, I

04:37 really think watermarking is getting a lot of attention here, and it really

04:41 doesn't solve that much of a problem. You need cryptographic hashing, invisible

04:46 watermarking. That's really important. Mr. Coleman? Yeah, just to add on to that, and

04:50 I think just to unpack two things here. We're talking about watermarking and

04:54 cryptographic hashing, effectively what's called provenance. It's either there or

04:59 it's not. The challenge with that is it presupposes that everybody's gonna

05:02 follow those same rules. All the bad actors will follow the same rules, and

05:06 we've seen time and time again a lot of the applications, whether they're on your

05:09 on your phone, in the App Store, or online, or they're open-sourced, they just

05:14 aren't gonna follow the rules. So we can't expect everyone to say, "Hey, we're

05:17 gonna play nice within this walled garden," when the bad actors, by definition,

05:20 are not playing by the rules at all. And so, you know, with Reality Defender, we

05:24 focused on inference. We don't touch any watermarking. We don't touch any personal

05:28 data. We actually assume we'll never see the ground truth. We'll never even know

05:32 if it is real or not, which means, instead of saying yes or no, we're taking a more

05:36 measured probabilistic approach. A probability saying maybe we're 95%

05:40 confident, maybe we're 62% confident. We build that into a larger framework of

05:44 just one signal among many to make a better insight, to have a platform or a

05:49 team decide to block or flag a piece of media, or a person, or an action. We're

05:55 gonna adhere to five-minute rounds on the first round. I hope to come back to

06:00 this line of questioning, and I apologize that others of you, Mr. Ahmed, you may

06:04 have some comments as well, but in deference to my colleagues who have

06:07 other commitments, I'm gonna turn to the ranking member. Thank you again, Mr.

06:11 Chairman, and thanks to everybody for being here. Mr. Coleman, you raised in

06:15 your opening statement what is, I think,

‘We Can’t Be Certain’: Blumenthal Warns Biden Deepfake Has Made Unknown ‘Impacts’

Category

Transcript

Recommended