Fighting deepfakes with content ‘nutrition labels’: Opinion

There may come a time when “nutrition labels” for online content become mainstream, just as they are for packaged food. 

It seems a bold vision and a tall order. 

But for Mr Andy Parsons, senior director of content authenticity at Adobe, it is the need of the hour amid the urgency to protect online users from scams, misinformation and artificial intelligence (AI) deepfakes.

“The reason that there are now laws coming to bear is because everyone is concerned about cryptocurrency scams and untrue news,” said Mr Parsons in an interview with The Straits Times on May 27. 

The European Union’s AI Act – which requires users to visibly disclose that they had used AI to manipulate or generate content – came into force in 2024, a world-first legislation to rein in AI harms. 

A second, the California Digital Content Provenance Standards Act, which requires AI-generated images, video and audio to be identifiable as such through metadata, is expected to go into effect in 2026.

How bad is the deepfake problem?

While creating deepfakes isn’t new, governments and think-tanks are alarmed about the levels of realism they can achieve, and the speed at which they can be churned out, with generative AI tools that have become widely available since 2022. 

The potential for harm increases exponentially when deepfakes are shared via social media that amplify reach. 

The most recent example involves Pope Leo XIV. He has been head of the Catholic Church for only a few weeks, and social media is already rife with AI-generated deepfake videos and images of the new pontiff. 

In a video shared on TikTok and Instagram, the Pope was depicted criticising US President Donald Trump and opposing everything Mr Trump stands for. The AI-manipulated video had the Pope saying: “Trump, the immigration policies you’ve implemented are a blatant trampling on both the teachings of the church and the promises of the American dream.” It was a deepfake.

Singapore politicians have their own encounters with deepfakes. Spoofed videos of Senior Minister Lee Hsien Loong promoting cryptocurrency schemes with guaranteed returns circulated online in 2024. Scammers appear to have manipulated his video interview in March 2023 with Chinese news network CGTN. 

The Singapore police frequently send advisories to the public detailing the latest ruse. In many instances, scammers have convincingly replicated the logos and layouts of news publications such as ST and CNA, complete with the bylines of known journalists. 

Sometimes the articles quote prominent figures like Prime Minister Lawrence Wong or a local bank’s chief singing praises about obscure cryptocurrency platforms. Many carry links to fraudulent websites, prompting unsuspecting victims to enter bank card details. 

ST has received letters from readers asking if the investments were genuine, and had published a story to debunk these fake articles.

Little surprise then that the World Economic Forum’s 2025 Global Risk Report named misinformation and disinformation as the top global threat over the next two years.

Mr Parsons is helping to counter this. He was in Singapore in May to meet officials from the Centre for Advanced Technologies in Online Safety, Singapore’s $50 million effort to build tools to spot misinformation and deepfakes. He also participated in a panel on misinformation at the Infocomm Media Development Authority-organised ATxSummit, which took place on May 28 and 29. 

Countering fakes with nutrition labels

Conmen ride on the trust that people have in personalities and brands, and will keep going. To counter their schemes, there must be an easier way for consumers to tell the genuine from the fabricated.

Thankfully, lawmakers are getting involved. The success of the EU and California laws could inspire global new norms in online content tagging.

A crack team at Adobe under Mr Parsons had a head-start five years ago in developing the tools to enable this reform. Its technology received a timely update in the past year. 

One tool, called Content Credentials, provides metadata about who created an image, when it was created and how (whether it came from a camera or was generated using AI) and what edits have been made. 

An accompanying tool, called Trust Mark, embeds invisible watermarks in images or videos, ensuring that a piece of the authentic content’s metadata is not lost when someone takes a screenshot. The invisible watermarks embedded in the copied content allows its metadata to be recovered. 

There are several competing tools in the market, but many are not free and are based on proprietary standards. This means that there may be compatibility issues. 

Conversely, Trust Mark and Content Credentials are designed to be open and free for all, including non-Adobe customers, as part of the firm’s broader push to make the tech an industry standard. 

To be precise, Adobe is the key funder of the Content Authenticity Initiative (CAI), a global non-profit coalition it co-founded with Twitter and the New York Times in 2020 to combat misinformation. 

Twitter, now called X, is no longer involved. But the CAI has expanded to include more than 4,500 members from across industry, civil society and journalism – including the Associated Press, The New York Times, BBC, Getty Images, The Wall Street Journal, Nvidia, Nikon, Panasonic, Leica and Universal Music Group. 

Mr Parsons, who leads the charge at CAI, is working on tools to allow nutrition labels to be tagged to webpages. Genuine content creators can attach these labels to indicate that the content has not been manipulated.

News websites and brands vexed by criminals impersonating them online should watch this space.

Spotting the labels

Attaching labels is only one part of the equation. The other is that these content labels should be as easy to spot and inspect as food labels on cereal boxes. They are. That is one checkbox ticked.

Content Credentials-marked images and videos carry a “CR” icon at the top left corner. Mousing over the CR icon will display a metadata box detailing origins and edit history.

But there is one snag. Users need to download and install the free Adobe Content Authenticity add-on for Google Chrome to view the metadata. Those without the add-on will not be able to see or interact with the CR icon on laptops and mobile phones. Because Chrome accounts for more than 60 per cent of global users, this extra step could limit the technology’s reach.

The Apple users’ camp is even worse off. Safari does not have the Adobe Content Authenticity add-on. Neither does the browser natively support the viewing of Content Credentials metadata.

For content nutrition labelling to take off, more checkboxes must be ticked.

First, the app makers could play a bigger role. They are starting to display Content Credentials, with LinkedIn and TikTok leading the way. But the approach is not consistent across the industry. 

YouTube, for instance, does it differently. It provides a “Captured with a camera” disclosure in the “How this content was made” section in the expanded description of some videos. It also lets users disclose AI-generated content with the “altered or synthetic content” label through an honour system that does not attach the labels by default.

Second, more work can be done by hardware makers so that nutrition labelling is done by default, say, when smartphone users snap a picture or shoot a video. 

Although the majority of the traditional camera market represented by firms such as Sony, Canon and Nikon are already using Content Credentials, only one smartphone player does this today. And smartphones are what most people use to produce multimedia content these days.

So kudos to Samsung, which is trailblazing this space. Its latest Galaxy S25 series, which started shipping earlier in 2025, allows users to automatically generate Content Credentials nutrition labels when users snap a picture or shoot a video with the devices. It is unclear if the Chrome browsers in the Samsung devices have also been updated to display the labels.

Third, phone and laptop operating systems can be updated to allow the viewing of content nutrition labels by default – so users need not download any browser add-ons. 

Standards are still evolving. But judging from the size of the CAI, with its more than 4,500 members, organisations seem to have coalesced around the open standards that Content Credentials support.

To be sure, Adobe’s approach is not meant to identify bad actors but rather to empower good ones by allowing them to prove their authenticity. It is a pragmatic strategy that deserves more support.

Irene Tham is an assistant news editor and oversees tech coverage as The Straits Times’ technology editor. She writes regular columns that look at how technology is shaping the world.

Source: The Straits Times © SPH Media Limited. Permission required for reproduction.

ADV: <Newly Published> Mental Capacity Act in Singapore: Law and Practice Woman wins civil suit for battery against former Mindef colleague who sexually assaulted her