What is an AI recommendation engine? Algorithms vs. AI

What's the difference between recommendation and manipulation?

Recommendation engines have existed for nearly as long as the internet. From search engines to e-commerce portals, billions of dollars have been invested in ensuring that your internet experience is tailored to you.

But it has come at a cost. The more customized our experience of the internet is, the more we diverge from each other. My internet is not the same as your internet. Your internet is not the same as the internet of your family members or your friends.

To understand the role that AI recommendation engines play in today's society, we need to go back. Way back.

Do you remember webrings?

In the very early days of the internet, there weren't search engines. There were directory sites. And those directory sites provided a list of categorized links. To get on a directory site, you had to manually submit your pages. Somewhere, a human reviewed it.

And yet, the web felt like a place of discovery. To facilitate that discovery, users created things like "webrings." Sites that were similar to each other or had similar audiences would link to each other. It was a personal digital recommendation.

Then came search engines. The first search engines were very simple. You typed "plumbers" and it returned pages that mentioned "plumbers." The next search engines were algorithmic. They didn't just look for keywords - like plumbers - but also algorithmic trust signals, such as the number of other sites that linked to a site.

And here we are. Today, search engines are highly complex. They tailor their results based on who we are - based on what we've looked for before.

Similarly, e-commerce sites recommend products to us based on our past purchases. But still, algorithms didn't entirely rule our lives until the advent of algorithmic social media.

The advent of the algorithm

TikTok, of course, wasn't the first to create an algorithmic feed. By the time TikTok emerged into global consciousness, Facebook and others had already been sorting their feeds algorithmically for some time.

Before algorithms, social media timelines were simply chronological. You saw what your friends had posted ordered by how new those posts were. But seemingly all at once, social media companies seemed to realize they could increase engagement with two major changes:

• Showing you more than just the people you follow, but also other accounts you could be interested in.
• Sorting your feed not chronologically, but based on what you were most likely to interact with.

And that wasn't all good. Famously, Facebook tested whether showing people sadder content on their feed could make them sadder as people. Surprisingly, the answer was "yes, of course."

But TikTok got people talking about "the algorithm." Because more than many other social media platforms, TikTok seemed to have a very powerful algorithm. Without fail, TikTok learned quickly what people wanted. And it kept them scrolling.

Algorithm, meet AI

AI systems are, of course, comprised of algorithms. We draw these boundaries loosely. In some form, any recommendation engine could be considered a rudimentary AI. A recommendation engine is, after all, not any further from generative AI than generative AI is from general AI.

But today, when we talk about AI, we are usually talking about generative AI: Large Language Models like ChatGPT and deepseek. These systems fundamentally work differently from classic recommendation engines.

Asking for a great book from an algorithmic recommendation engine

• You go to an online book store to review "recommended books."
• The online book store looks at the titles that you've read before.
• Each title you've read before has a classification, such as "Drama," and "Romance."
• • This may even become more granular, such as "Western Drama," or other factors, like length.
• The online book store sees that you tend to like long alternative history novels with at least two authors.
• The online book store looks at its inventory for long alternative history novels with at least two authors.
• The online book store returns one of the five books in existence that fit this category.

Your recommendation has been designed for you. It has used information that you provided, one way or another, and compared that to hard-coded, compiled information regarding a library of books.

Asking for a great book from an LLM

• You go to ChatGPT and ask ChatGPT to recommend a book.
• ChatGPT asks you what type of book that you like - or accesses information that it has previously stored on you.
• ChatGPT sees, with the information provided, that you are likely looking for a short romance novel.
• ChatGPT then references the most likely response from wide samples of responses online.
• ChatGPT generates a response, listing the most likely short romance novels.

Your recommendation has been tailored by either your request or previously retained information. But your recommendation is predictive. It's not based on anything quantifiable, it's just the most common response to this type of question.

The hidden dangers of AI recommendation engines

Most highly addictive recommendation engines are more algorithmic than generative AI. They are highly attuned to the user and that's what makes them highly addictive.

The advantage of an AI recommendation engine is that it's fast and easy. The book recommendation engine described above doesn't need a library of books or collected user information; it's already been trained on user information. It could be fine-tuned, but it doesn't have to be to work good enough.

The advantage of a more algorthmic AI is that it can really drill down to what will keep you watching, reading, or scrolling. And that is when AI can become quite dangerous - when the AI's goal is not the same as our own. Because while the AI might want to keep us consuming, we might have other things we need to do with our lives.

Want to learn more?

Check out recommendations with AI to play around with television show recommendations generated via LLM.