E-Paper
Was That Amazing Video in Your Feed Real or AI? Tech Platforms Are Struggling to Let You Know
Meta, YouTube and TikTok are grasping for ways to protect users’ trust as their platforms fill with AI-generated photos and videos of events that never happened
Meta, YouTube and TikTok are grasping for ways to protect users’ trust as their platforms fill with AI-generated photos and videos of events that never happened.
But their patchwork of imperfect tools and voluntary policies sometimes inadvertently punishes “real” content and leaves plenty of AI work unlabeled.
TikTok star Nikola Savic, for example, says he has had “AI-generated” labels misapplied to many of his popular videos of surprising transitions between disparate scenes, provoking dismissive comments from some of his 4.8 million followers.
Savic said he occasionally taps artificial intelligence to smooth out rough patches but otherwise uses standard editing tools in a painstaking process that can take several days for each clip.
“People already have in their minds that I’m only using AI,” Savic said. “This hurt my reputation so much. And at TikTok, nobody really seems to care.”
The largest tech platforms have pushed to label AI-generated content since soon after OpenAI’s ChatGPT opened the floodgates in late 2022, but the effort is becoming more urgent every time a tool like Midjourney or Google’s new Veo 3 improves the quality of AI photo and video simulations. The fear is that consumers who discover they have been fooled by an AI ad, news video or creator post will become skeptical of the other content these platforms carry.
Their initiatives leave room for mistakes or omissions. Meta, the owner of Facebook and Instagram, this year introduced an option for some advertisers to disclose when they have used generative AI tools from other companies like OpenAI to create or edit their ads.
“We will continue to evolve our approach to labeling AI-generated content in partnership with experts, advertisers, policy stakeholders and industry partners as people’s expectations and AI technology evolve,” Meta said as it announced the policy.
AI work without disclosures meanwhile continues to circulate freely.
Kalshi, the prediction-market startup that lets people bet on current events and pop culture, ran an ad on YouTube TV during last month’s NBA Finals that was generated entirely with Veo 3, according to a Kalshi spokesman.
The ad, which depicts a series of gamblers enthusing about their bets, cost $2,000 to make and has accumulated 19 million views, the spokesman said.
By the time discerning viewers see an alien chugging beer at a college party, they may realize that nothing in the Kalshi ad is real. But it wasn’t labeled as AI.
A Veo 3 video promoting conservative credit-card brand Coign doesn’t include such an obvious tell, as all the featured players appear to hail from planet Earth. It ran on Facebook, YouTube and LinkedIn without any disclosures.
And a Veo-generated ad for the multivitamin brand Ritual displays a small, brief watermark identifying it as AI-generated when it appears on YouTube, but not on LinkedIn or Instagram.
The video, which stars a series of new mothers who immediately reveal that they are actually AI-generated characters, was designed to position Ritual’s products as an alternative to misleading claims that crowd today’s wellness space, according to founder and CEO Katerina Schneider.
Ritual doesn’t have any other AI campaigns planned, but would gladly add disclosures in the interest of transparency, she said.
The labeling game
TikTok was one of the first companies to try to address the matter in 2023 by developing a tool meant to automatically detect and label AI-generated content as well as a button that allows creators to disclose their own use of third-party AI tools.
Others have followed with their own formulas, tweaking them as events unfold.
Meta last year updated its “AI info” tags to identify and distinguish between content created with AI and content that’s only modified by other companies’ AI tools, in part to address claims from some creators that their work had been inaccurately labeled. Meta also automatically tags anything created with its own AI tools.
Google says it automatically attaches small visual watermarks to nearly all content created with its Gemini app, which includes Veo; this content also includes digital metadata identifiers that can be used to automatically label it as AI. The ads from Kalshi and Coign appear to have run without any visual marks, however, and videos created with a Gemini tool aimed at filmmakers are exempt.
The mothers in a recent ad for multivitamin company Ritual want to make sure viewers know that they’re not real.
Policy at Google’s YouTube unit since early last year has required creators to disclose the use of AI in any content that viewers may confuse for real life. YouTube only requires political advertisers to disclose their use of AI, since general AI-generated ads don’t inherently violate ad policies designed to prevent harm to consumers, a YouTube spokesman said.
Most brands’ creator marketing contracts prohibit the use of generative AI to avoid losing fans’ trust, said Dana Neujahr, managing director at social-media and influencer firm We Are Social.
TikTok videos by one We Are Social client, however, have been slapped with automatic AI labels after they used an AI tool to add subtitles, Neujahr added. “To a novice or someone not in this industry, they could think that this entire piece of content was AI-generated,” she said. “It feels like a bit of a slippery slope.”
Social-media companies may have developed these suites of AI labeling tools, at least in part, to stay ahead of any controversies or regulatory concerns regarding AI content, she said.
The tech industry has attempted to address the need for transparency collectively and comprehensively.
A consortium called the Coalition for Content Provenance and Authenticity, or C2PA, offers standardized credentials that individuals and companies can use to indicate whether their content is AI-generated.
LinkedIn, for example, says it displays the coalition’s logo, which users can click for more information, on any images and video uploaded with those credentials.
OpenAI supports the coalition’s effort, though it also warns that its metadata “is not a silver bullet.” News Corp, owner of The Wall Street Journal, has a content-licensing partnership with OpenAI.
“It can easily be removed either accidentally or intentionally,” OpenAI says on its site, noting that social-media platforms frequently strip metadata from uploaded images and that screenshots don’t capture it.
The credentials offered by C2PA are a well-meaning idea that will struggle to be effective against the swaths of AI content, from sources both reliable and deceptive, that are rising in social-media feeds, according to Matthew Stamm, professor and head of the Multimedia and Information Security Lab at Drexel University.
“If this paradigm is going to work, it has to be universally adopted,” Stamm said. “To me, that doesn’t seem like a realistic solution.”
Write to Patrick Coffee at patrick.coffee@wsj.com
