AI slop is a feature, not a bug

Over the years, lots of organizations bought into “rapid AI adoption,” imposing LLMs, image generators, and other GenAI technology into workflows as quickly as possible. Unfortunately, many businesses did so without taking time to really understand the technology—and why, as folks like me argued, it requires human oversight. So we probably shouldn’t be surprised when companies find out that their AI models didn’t perform as expected.

Ernst & Young Global Limited, for instance, conducted a survey of senior AI executives, who revealed that AI adoption has outpaced AI governance, leading to cybersecurity incidents, legal risks, financial losses, and reputational damage. To mitigate these problems, workers must spend substantial time fixing AI’s erroneous and low-quality output—sometimes known as AI slop.

But error-prone and generic-sounding content is, in general, what we should expect from GenAI, at least to some extent. Take LLMs as an example:

• LLMs work by finding statistical patterns in information, based on data correlations, some which may be spurious. For this reason, hallucinations, or erroneously fabricated answers, will remain an issue with LLMs.
• Since LLMs generate content from statistical patterns, the resulting text is based on, and likely biased toward, these common patterns. Hence, the AI models will often favor frequently occurring words and phrases, standard sentence structures, and widely held viewpoints. That’s why responses from LLMs can sound so formulaic or generic.

In other words, generic-sounding content that’s prone to error is a feature, not a bug, of GenAI technology like LLMs.

But sometimes generic content is fine

Of course, with sufficient human oversight to mitigate the errors and oversee basic quality, some generic-sounding content can be useful for certain tasks. For example, LLMs may be just fine for:

• Automating templates, where standardized or commonly used words and phrases are exactly what’s appropriate.
• Giving routine responses to common questions or queries, such as FAQs.
• Conducting enhanced searches online, by letting users enter much more than just a single common keyword.

And sometimes even AI slop has its uses

Additionally, even some of the most ridiculous examples of AI slop may have a use, even if unconventional. One of the funniest online trends that I’ve seen lately is the ironic use of GenAI to create purposely bad video parodies of movies. They’re usually labeled: “[Movie title] but it’s ruined by AI.” Here was one such video that really had me in stitches.

I suppose what also has me laughing from this satirical content is that people have become well aware that AI “ruined” something. So we might as well make fun of the absurdity. Perhaps that lampooning is a sign of AI hype coming to an end. If so, it’s about time.