Google working on tech to discern AI-made content: Company executive

Google is developing systems to identify AI-generated synthetic media and combat political misinformation.

Google is working on systems to identify AI-generated synthetic media that will also be part of its effort to combat political misinformation, a top executive of the company said on Thursday, adding that these will be implemented across its products and potentially offered for wider use in the industry.

The comments come against the backdrop of rising concerns over the use of artificial intelligence-generated text and deepfakes for misinformation.(AP)

The comments come against the backdrop of rising concerns over the use of artificial intelligence-generated text and deepfakes for misinformation, especially in light of a specific case when athletes protesting against the government were targeted with realistic-looking morphed images.

These systems, Google’s vice president for Trust and Safety, Laurie Richardson, said during an interaction with a closed group of journalists, include yet-to-be launched watermarking technologies for its own products, new transparency requirements, and machine-learning classifiers that can detect synthetic media generated by other companies’ tools.

“We shared two big announcements at I/O [Google’s developer conference held in May]. One is that we are committed to watermarking all of our products so that people can see whether it is Google itself creating [the content]… We haven’t given specifics on that yet, but that is underway. And the other is that we are providing metadata to images and text so that you can actually trace back and understand in a simple way what their origin is on the internet,” she said, without giving a timeline of when these can be expected.

Other steps, she added, will be based on a more layered approach “from us and I hope others” to try to help users understand what they’re seeing. “There will be some transparency requirements that we can enforce around labelling and making clear when content is generated or not. And we will be partnering with the broader ecosystem because we know that harms don’t start and stop on our own platforms,” she said.

One of these, for instance, she said would be technical solutions similar to how the company has developed a machine learning algorithm to detect child sexual abuse material, or CSAM, that social media service TikTok uses.

“There’s a lot of really interesting work happening on this detection side. One piece of this that we’ve externalized is a classifier that can detect synthetic audio with, I think, 99% accuracy right now. So I do have some optimism about our ability to bring at least the best sort of technical solutions to bear,” she added.

Detecting synthetic media and deepfake has become an increasingly difficult challenge. Recent versions of AI art programmes Midjourney, Dall-E and Stable Diffusion have yielded viral, realistic but fake images, such as those of the Pope disc-jockeying, while new photo editing features by Adobe and Google make what used to be challenging photoshopping techniques — such as removing backgrounds or objects — easy.

In April, Tamil Nadu minister P Thiagarajan alleged he was targeted via a deepfake audio clip that purported to suggest he attacked his own party — allegations that could neither be confirmed or borne out by technical analysis by independent experts, according to reports.

The following month, wrestlers on protest against the Union government were targeted with manipulated images that showed them to be smiling while being taken to prison when the original photograph had no such expression. This was easier to rebut since the original image was quickly traced and people could replicate the results with publicly available apps.

“For me it is top priority that we help users understand what they’re seeing when they encounter information on our platforms since you know the barrier of entry to generating images, audio and text is going to be lower than before,” she said.

Richardson added that Google’s specific interventions around AI-related misinformation are “still evolving”, and how it shares these — “whether that will take a database or a hashing approach or some similar approach” is still to come.

‘Encryption debate is challenging’

Hashing and databases are at the heart of the present efforts to fight CSAM. Google and its prominent industry rivals include techniques where they collaborate with NGOs to track known child sexual imagery and create a fingerprinting algorithm. If an image or a file on a user’s device has the same fingerprint – or the hash – that is deemed as being a positive match for a child sexual image.

But these – for now – work only when an image is uploaded to the cloud. They have been criticised for both, being inadequate in combating the problem and often throwing up false positives.

The UK, as an example of some countries considering legal provisions that want tougher steps on CSAM, has proposed that tech companies scan for an prohibit such content even when they are stored on devices, a process that will likely break what is known as end-to-end encryption (E2EE) that prohibits such companies from being able to determine what is on someone’s personal device.

Google, unlike Apple which last month took a public position against the UK’s provisions, has not yet indicated a specific position on it.

“I think the encryption debate is a really tough one. I don’t want to speak specifically to specific forms of regulation that might be pending but trying to thread the path to being as aggressive as we all want to be on CSAM while at the same time not creating vulnerabilities in the way that our products are built is really important. This is an industry wide problem,” Richardson said on a question about encryption.

The government India, as per its IT Rules, 2021 that have now been stayed by courts, too legally mandated to trace the origin of a message in an effort to combat misinformation, but such an approach too is impossible without breaking E2EE, experts have said.