The Sloppening Begins: Identifying AI slop just got easier
For months now, AI nonsense has been foisted upon poor audiences. All at the expense of artists, writers, and other content producers, with their hard work and craft siphoned into LLMs. It might all be coming to an end now. Pushed by regulators in the EU, Anthropic has announced a series of measures to help identify content that’s been produced at the hands of AI. Article 50 of the EU’s AI Act insists upon it, so others will inevitably follow.
The measures, according to Anthropic, will leave a watermark embedded within content that’s hard to remove. Even heavy editing of AI-generated content isn’t a guarantee users can pass outputs off as authentically human. Slop-fatigued audiences everywhere can heave a sigh of relief. So too can analog content producers (like the Luddites here at Shortfuse). But it opens up a big can of worms for everyone else.
The AI cat is well and truly out of the bag in most companies. Convincing workers to stop using their digital intern for shortcuts won’t be easy. Nor will the legions of newly-minted thought leaders and pundits taking over LinkedIn. For both, there’s an inevitable naming and shaming on the way, this time with the full credibility of Anthropic behind it, rather than spurious AI content identifiers that have critiqued everyone from Mary Shelley to segments of the Bible for blatant generative AI use.
The authors of those publications are unlikely to call a lawyer to refute those claims. But with a multi-billion-pound company now behind the accusations, living authors just might. This is especially true given Anthropic mentioned model-level statistical text watermarking can also trigger on human-written copy that Claude simply proofread, translated, or edited.
Aside from risks of litigation, this change doesn’t bode well for AI companies in other ways. Content creation is a big use case for Anthropic and its peers. Analysis of OpenAI usage data shows 28% of queries ultimately involve ChatGPT editing or writing content. But in reality, the change may be of benefit. A lot of content generation, such as writing blogs and LinkedIn missives, is ultimately very cheap—at least at the standard AI alone is capable of delivering. Expensive AI subscriptions won’t last long with enterprises or consumers if pumping out drab and repetitive content is the peak of its capabilities.
The other side of the coin is AI slop is actively damaging the reputation of AI giants. It’s annoying audiences, driving rebellions in schools and offices, and, in more extreme cases, is the source of crimes such as fraud and incitement to riot. Clamping down by making AI slop identifiable is a good solution to mitigate these challenges in the long-term. Slop-merchants will eventually be pushed to stop, or will be sidelined by search engines and social media platforms looking to keep the crud out of their audiences’ feeds. Just last week, LinkedIn launched a slop-flagging system with protecting audiences in mind. It’s likely others will follow.
AI giants wouldn’t be the first companies to regulate how their product is used to protect brand image. Toyota’s vehicles have been synonymous with a series of wars, including the aptly named Toyota Wars in the ’80s. But the more recent conversion of their vehicles for military use by militias and terrorist organizations pushed the company to enforce strict policies preventing their sale to untoward buyers. Ultimately, while the rugged ability of their vehicles was a testament to quality engineering, there was little brand value in being the chosen vehicle supplier for war criminals.
In this respect, it’s possible that making AI-produced content easier to identify will be a boon for an industry that, on the surface, is currently just very annoying. It remains to be seen how these measures will impact AI content production; the first KPI we’ll be monitoring here is if LinkedIn content becomes less monotonously painful.