Now we are being overrun by them — “deep fake“photos that are virtually indistinguishable from real ones (except for extra fingers), AI-generated articles and essays that sound realistic (though still awkward), AI-generated reviews, and more. Additionally, AI systems may be collecting copyrighted or intellectual property material from websites as training data, leaving users open to infringement.
Also: Most people worry about deepfakes — and overestimate their ability to detect them
The problem, of course, is that AI content is getting better all the time. Will there ever be a perfect way to identify AI-generated material? And what should AI creators and their companies understand about emerging techniques?
“The original use of generative AI was for entertainment and educational purposes, but now we see a lot of bad guys using AI for bad purposes,” Andy Thuraivice president and principal analyst at Constellation Research, told ZDNET.
Media content — images, videos, audio files — is particularly vulnerable to “misattribution, plagiarism, theft, or lack of attribution,” Thurai added. This means “creators will not receive proper credit or revenue.” Another danger, he added, is that “the spread of misinformation can influence decisions.”
From a textual perspective, a major problem is that many prompts and repetitions compared to language models tend to blur the watermark or provide only minimal information, according to a recent study paper compiled by researchers at the University of Chicago, led by Aloni Cohenassistant professor at the university. They call for a new approach – multi-user watermarking – that “allows model-generated text to be traced to individual users or groups of colluding users, even in the face of adaptive prompts.”
Also: Photoshop vs. Midjourney vs. DALL-E 3: Only One AI Image Generator Passes My 5 Tests
The challenge for both text and media is to digitally stamp language models and AI output, you have to implant detectable signals that cannot be modified or removed.
Industry-wide initiatives are underway to develop infallible AI watermarks. For example, Content Origin and Authenticity Alliance (C2PA) – a joint effort formed through an alliance between Adobe, Arm, Intel, Microsoft and Truepic – is developing an open technical standard that aims to give publishers, creators and consumers “the ability to trace the origin of different types of media.”
Also: AI Scientist: ‘We Need to Think Beyond the Big Language Model’
C2PA unifies Adobe’s efforts by Content Authentication Initiative (CAI)focuses on systems that provide context and history to digital media and Project Originan initiative led by Microsoft and the BBC to tackle misinformation in the digital news ecosystem.
“Without standardized access to detection tools, checking whether content is AI-generated becomes a costly, inefficient, and arbitrary process,” Shutterstock’s Alessandra Sala wrote in a post. report published by International Telecommunication Union (ITU) — the United Nations’ digital technology agency. “In practice, it involves trying all available AI detection tools one by one and still being uncertain whether some content is AI-generated or not.”
The development of AI generated The platform “needs to have a public registry of watermarked models, along with common detection tools,” Sala urged. “Until then, ethical AI users have to query each company’s watermarking service at random to check whether a piece of content is watermarked.”
Also: Today’s Challenge: Addressing AI’s Unclear Returns and Questionable Accuracy
The C2PA initiative promotes “widespread adoption of content authentication, tamper-proof metadata that can be attached to digital content,” Thurai explained. He likened content authentication to a “nutrition label” that creators can attach to their digital content, which can be used to track the origin of the content. With this open standard, publishers, creators, and consumers will be able to “track the origin and evolution of a media work, including images, videos, audio, and documents,” he added.
Thurai says the way it works is that content creators can “get recognition for their work online by attaching information like their name or social media handles directly to the content they create.” This can be as simple as clicking on a pin attached to a piece of content or visiting a website to verify the source. Such tools “validate relevant information, as well as provide a detailed history of changes over time.”
