Inside the race to detect AI’s most convincing fakes

With additional reporting by Michelle Anindya

On April 1, an image of two people kneeling on a beach with their dogs started the latest AI firestorm.

The reason? The image recreated the aesthetic of the renowned Japanese animation house Studio Ghibli, and it did so thanks to the image generation tool launched on OpenAI’s ChatGPT.

Image credit: Made by Tech in Asia using AI

Within 24 hours, a cease-and-desist letter, reportedly from the studio’s lawyers, appeared online.

The letter, too, turned out to be an AI-generated fake, adding another level of confusion about the authenticity of, well, anything and everything online.

The surge in AI fakes comes at a time when traditional fact-checking support is in retreat. Meta, for example, cut back its funding for independent verification efforts in January.

“I think the issue is that there is this giant fact-checking vacuum in social media now,” says Dennis Yap, CEO of Singapore-based AI Seer, which developed the fact-checking platform Facticity.

Fighting fire with fire

Facticity is part of a new wave of startups that are building tools to help users verify what’s real. The kicker: They’re using AI to do it.

For example, users can submit a YouTube link to Facticity’s website, and it will analyze each statement from the video, labeling them as either true, false, or unverifiable. For each verdict, Facticity will provide explanations and give sources.

“We came up with ‘unverifiable’… so that [it’s clear that] some things are just not simple yes and no answers,” Yap tells Tech in Asia.

World ID’s Orb verifies a person once in the real world, then allow them to reuse that proof of identity across platforms / Photo credit: World ID

This approach is powered by a scaffolded AI architecture. “Scaffolding” is a technique that allows an existing large language model (LLM) to interact with other tools or models. It’s similar to assigning specific roles to different AI agents – such as extracting claims, searching for evidence, and cross-verification – combining them to form a single tool, explains the company’s CTO Shahruj Rashid.

Trust in the age of deepfakes

Web3 makes headway in identity verification

Tagging authentic content