YouTube, Instagram, and TikTok: The Failure of AI Filters

Social media users on platforms like YouTube, Instagram, and TikTok are increasingly confronted with content generated by artificial intelligence (AI). Although these platforms have implemented labeling systems to identify this type of content, they do not offer effective solutions for those who wish to avoid it entirely.

It is nearly impossible to avoid encountering AI-generated content online. YouTube, Instagram, TikTok, and other platforms have ramped up their content authentication efforts over the past year, automatically applying labels to distinguish AI-generated images, videos, and music from those created by real humans.

Currently, the labeling efforts on these platforms have not significantly altered the presentation of content online. For instance, Meta has introduced "AI info" labels on Facebook and Instagram, but this does not allow users to actively filter out this content. Filtering attempts on DeviantArt and Pinterest also show limitations, as these options are often hidden and underperforming. DeviantArt offers an AI filter, but it is difficult to access and not very effective. Pinterest provides a similar system, but the filters are hard to find and do not completely eliminate AI content.

AI labeling systems, such as C2PA and SynthID, incorporate metadata or invisible watermarks to authenticate content. However, these methods are not foolproof, especially against open-source AI models that can bypass these measures. Additionally, labeling initiatives can sometimes incorrectly flag authentic content, as experienced by Meta and YouTube. Platforms are wary of mistakenly flagging genuine content, which has already been an issue for Meta and YouTube.

Companies like OpenAI promote these solutions as a means to combat deepfakes and other digital deceptions. However, if regulators become aware of their ineffectiveness, platforms may be forced to develop more robust solutions.

Finally, one alternative could be to verify human creators, as Instagram proposes for its users. This would help reduce exposure to low-quality content from unverified sources. However, platforms that benefit from AI-generated content may be reluctant to encourage such filters. They argue that they risk incorrectly flagging authentic content if they push labeling initiatives too far, which is a major concern.

Provenance-based systems like C2PA and SynthID work by embedding metadata or invisible watermarks into content at the time of creation. But there are many open-source AI models that do not do this (especially if designed for malicious purposes), and even then, metadata can be too easily removed for it to be reliable. There are also detection-based methods that analyze patterns in digital content and then assess the likelihood that AI was used to create it, but these can yield false positives. None of this currently works effectively at scale.

Nevertheless, companies, including AI providers like OpenAI, are currently touting these AI labeling solutions as something that will help prevent people from being duped by deepfakes and other deceptions. If regulators become aware of their ineffectiveness, online platforms and AI providers may need to find a solution that actually works, rather than what currently appears to be a smokescreen.

Platforms will argue that they risk incorrectly flagging authentic content if they push labeling initiatives too far. Meta and YouTube learned this the hard way after applying AI labels to images and videos that creators claimed to have produced without the aid of such tools. If this is such a concern for current labeling systems, then find a better solution. Improving the user experience for your millions of users is surely a worthwhile investment to fend off competition?

And while I’m asking, why can’t I report all the unlabeled AI-generated content I see every day? Given the scale of the problem — with a study by Kapwing last year revealing that over 20% of YouTube videos shown to new users are low-quality generated content — I imagine a large number of human moderators would be needed to effectively review each report.

And perhaps that’s the problem. At a time when big tech companies are replacing workers with AIs that can supposedly outperform them, can they afford to backtrack on their carefully constructed narrative by rehiring them to solve AI issues? Humans tend to have annoying requirements like salaries and benefits, compared to automated moderation systems that lack nuanced investigative skills.

An alternative to labeling AI-generated content would be to start labeling verified human creators. This wouldn’t necessarily identify synthetic content published by these creators, but it could help us see less content from unverified content farms that produce low quality. This is the future that Mosseri of Instagram has proposed for Meta's image-sharing platform and something that Spotify is already doing with verified artists.

Of course, Meta, Spotify, and Google are not just hosting AI-generated images, ads, and music; they are also responsible for creating the tools that produce them. That’s why they insist that not all AI-generated content is low quality and that it’s more a matter of quality — if it becomes convincing enough, they hope you won’t notice and will continue to enjoy that content. Allowing users to filter it out would go against all the efforts these platforms have made to profit from AI: they want you to embrace this low-quality content factory.

I would be delighted to be proven wrong. I am actually pleading with online platforms to show that AI labeling efforts have not been a waste of time. But for now, they hold all the cards, and we can only hope that their AI moderation efforts are up to par. So, give us a basic "no AI" or "verified human creator" filter, and we will judge the effectiveness of all this.