For twenty years, recommendation systems have worked by watching what you do and inferring what you like. The Netflix Prize proved how hard that gets: a polarizing film like Napoleon Dynamite broke the math because viewers with nearly identical taste reacted in opposite ways. Now the platforms are trying something simpler. Spotify’s Taste Profile, YouTube’s Custom Feeds, Google Discover and Threads’ Dear Algo all let you describe what you want in plain language instead of training the algorithm through weeks of clicks. The more interesting shift is still ahead, when the system starts asking you questions back. Behavioural signals aren’t going away, but conversation gives recommenders a source they have largely ignored, which is the person they are trying to understand.
For most of the last two decades, recommendation systems have worked the same way: watch what we do and try to figure us out.
I’ve been fascinated by this since the Netflix Prize. In 2006, Netflix offered $1 million to anyone who could improve its recommendation algorithm by 10%. Thousands of teams spent years trying to squeeze better predictions from Netflix’s data.
Along the way they encountered the “Napoleon Dynamite problem.” Napoleon Dynamite was unusually difficult to predict. People tended to love it or hate it, and even viewers with otherwise similar tastes could react completely differently. Other polarizing films created similar problems.
It exposed something fundamental about recommendation systems: sometimes human taste is just difficult to infer.
Twenty years later, we’re trying something surprisingly different. We’re talking to the algorithm.
Telling the algorithm what you want
Traditionally, if YouTube misunderstands my interests, I have to retrain it through behaviour.
Watch something. Ignore something else. Click “Not interested.” Subscribe. Unsubscribe. Eventually it figures me out.
Generative AI creates another option: I can just explain what I want. Several major platforms are now experimenting with this.
Spotify’s Taste Profile lets users describe their interests in natural language and influence subsequent recommendations. Instead of Spotify only inferring my musical tastes, I can correct or refine its interpretation.
YouTube Custom Feeds takes a slightly different approach. I can describe a feed such as “long documentaries about engineering and infrastructure” and YouTube creates a continuously updating recommendation feed around that intent.
That recognizes something recommendation systems often struggle with: we don’t have one set of interests. I might want AI research while working, cottage projects Saturday morning, comedy at night and NFL analysis Sunday afternoon.
Google Discover is moving in the same direction, allowing people to describe what they want more or less of in natural language.
And Threads’ Dear Algo makes the idea wonderfully literal:
“Dear Algo, show me more posts about podcasts.”
Instead of spending weeks training the algorithm through behaviour, you simply tell it.
The next step is an actual conversation
So far, most of these are one-way natural-language instructions. The more interesting possibility is when the algorithm starts asking questions.
Imagine telling Spotify:
“I’m getting too much classic rock.”
Spotify could respond:
“Less classic rock generally, or mostly less ’70s arena rock?”
I might answer:
“Mostly arena rock. I still like Talking Heads and post-punk.”
Now the recommendation engine is trying to understand me. This field is known as Conversational Recommender Systems (CRS), and large language models make the idea much more practical.
A future recommender could combine:
- Implicit intent: what I watch, read, listen to and skip.
- Explicit intent: what I follow, subscribe to, like or dislike.
- Conversational intent: what I actually tell it.
The algorithm could even recognize uncertainty and decide when it’s worth asking. YouTube might notice I’ve suddenly watched 20 plumbing videos and ask:
“Is plumbing a new interest or are you fixing something?”
One answer could prevent weeks of terrible recommendations.
Maybe Napoleon Dynamite just needed a conversation
This brings us back to Netflix. Instead of finding increasingly subtle statistical correlations to predict whether I’ll like Napoleon Dynamite, a recommendation system could ask:
“How do you feel about awkward, deadpan humour?”
Behavioural recommendation isn’t going away. It’s incredibly powerful. But conversational AI gives recommendation systems a source of information they’ve largely ignored, the person they’re trying to understand.
Twenty years ago, Netflix offered $1 million to build an algorithm that could better predict what we wanted. We have spent the years since getting extraordinarily good at it.
Sometimes it’s easier to ask.
Frequently Asked Questions
What is a Conversational Recommender System?
A recommender that gathers what it needs by talking with you rather than only inferring it from your behaviour. It can ask clarifying questions, accept corrections and refine its understanding across a few turns. The idea has been studied for years, but large language models are what finally made the conversation feel natural enough to ship.
Does this replace behavioural recommendation?
No, and it shouldn’t. Behavioural signals are enormously powerful and they cost you nothing to produce. What you actually do is often a better guide than what you say you like. Conversation adds a layer the system never had, which is the ability to ask when the behavioural signal is ambiguous.
Which platforms let me talk to the algorithm today?
Spotify’s Taste Profile, which launched in the US in September 2026 after a New Zealand pilot. YouTube Custom Feeds, announced at Made on YouTube in September 2026 and powered by Gemini. Google Discover, which rolled out natural-language preferences in the Google app. And Threads’ Dear Algo, which arrived in February 2026 after Meta noticed people were already posting requests to the algorithm as a joke.
Is anything actually asking questions back yet?
Google Discover is the closest. It can ask follow-up questions to narrow down what you meant and it remembers the request afterward. Most of the others are still one-way instructions, which is why the two-way version is the part worth watching.
What was the Napoleon Dynamite problem?
During the Netflix Prize, competitors found that a handful of polarizing films wrecked their accuracy. Napoleon Dynamite was the worst of them. Ratings clustered at one star and five stars, so viewers who looked almost identical in the data landed in opposite camps. One competitor found that single film accounted for roughly 15 percent of his remaining error.
