How Recommendation Algorithms Train Us Like Puppies

A guide to recommender systems, feedback loops and the small choices that shape what appears in a feed.

The short version

A recommendation engine is an AI system that predicts what a user is likely to prefer based on data. It analyses patterns in user behaviour such as viewing history, clicks, ratings, and purchases. Using machine learning, it compares these patterns with other users and generates personalised suggestions.

Recommendation engines are widely used in streaming platforms, social media, online shopping, and news feeds. But, Rosie says “they are not mind readers”. They do not understand you in the human sense. They look for signals: what you watched, skipped, searched for, followed, shared or returned to. They then estimate which other items might hold your attention.

That creates a loop:

1. You do something.
2. The system records a signal.
3. The ranking changes.
4. You see a different selection.
5. You respond again, creating another signal.

Rosie learns that jumping through the hoop earns a treat. A recommendation engine learns that a particular kind of video earns a long watch, a replay or a tap. The comparison is useful, as long as we remember that a platform is not a puppy and a person is not a training target.

The long version

Recommendation engines rely on algorithms. The word algorithm can make the whole thing sound like one mysterious machine. In practice, platforms usually combine many algorithms, business rules, safety systems, user settings and data sources.

The Australian eSafety Commissioner describes recommendation engines as systems that select, filter and personalise content and other items across online platforms and applications. They may help people discover artists, products, activities and ideas, but they can also shape what people encounter and how often they encounter it.

The system may assign a score to each possible item. The score can be influenced by signals about the content, signals about the user and predictions about what might be relevant. The item with the highest score is not necessarily the best, truest or healthiest item. It is simply the item the system currently predicts will meet its chosen objective.

Those objectives differ. A music service may want to help you find something you will enjoy. A shopping site may want to sell a product. A social platform may balance relevance, safety, variety and time spent. The same person can receive very different recommendations from different services.

What does the system notice?

Different platforms use different signals, and they do not all disclose their methods in the same detail. TikTok’s explanation of its For You feed says that recommendations can use interactions such as likes, shares, follows, comments, content creation and whether someone watches a video through to the end. It also describes video information such as captions, sounds and hashtags, along with device and account settings.

Some signals are obvious. Others are easy to miss:

• watching a clip twice;
• pausing on a post;
• opening a comment thread;
• searching for a topic;
• following an account;
• scrolling past quickly;
• selecting “not interested”;
• watching several related videos in a row;
• changing language, location or other account settings.

The system may treat these actions as clues about relevance. A clue is not the same as a considered preference. You might watch a disturbing video because you are checking whether it is real. You might pause because you are confused. You might replay something because you missed the important detail. The system may not know the difference.

The feedback loop

Suppose you watch one video about basketball. The platform shows you a few more. Maybe a video about how to practice your dribbling skills (dribbling, not drooling Rosie). You watch it because it plays automatically. The system records the viewing as a signal of interest and supplies more related material. Before the next video, an advertisement appears for basketball sneakers. The platform has not read your mind. It has followed a chain of signals and found another opportunity to hold your attention or sell something.

Researchers describe this as a feedback loop. Recommendations influence what people see and choose; those choices become new data; the new data influence later recommendations. Research on simulated recommender systems has found that these loops can amplify popularity bias, reduce the variety of items shown and make users’ experiences more similar over time.

The loop can be pleasant. A person looking for beginner guitar lessons may quickly find useful demonstrations. It can also become narrow or unhelpful. A person researching a health worry may receive increasingly alarming material.

The important point is not that every recommendation is harmful. It is that a feed is an environment shaped by repeated signals, not a neutral window onto the world.

Why does the feed sometimes feel like a tunnel?

Personalisation is useful because no one can inspect every possible video, product or article. The trade-off is that relevance can crowd out surprise.

Platforms often try to add variety, remove duplicates or limit content that fails safety rules. TikTok, for example, says that its For You system may deliberately introduce content outside a person’s immediate interests and avoid showing the same creator or sound repeatedly.

Those safeguards matter, but they do not make the feed neutral. A platform still chooses what counts as relevant, what counts as safe, how much variety is enough and which signals receive the most weight. eSafety’s 2026 position paper also warns that engagement-based systems can optimise for attention in ways that do not always align with a person’s long-term interests.

This is where the puppy analogy needs a boundary. Rosie does not choose her own reward schedule, and a recommender system does not have a personality or intention. The analogy is about repeated feedback: an action is followed by an outcome, and the outcome makes some future actions more likely.

How to give yourself more room

No individual can personally fix a platform’s design. User controls are useful, but they are not a substitute for safer systems and accountable companies. eSafety explicitly cautions that the burden of safety should not fall solely on users.

Still, a few habits can make a feed less automatic:

• Search directly instead of relying only on the home feed.
• Follow several reliable sources rather than one recommendation stream.
• Use “not interested”, hide, mute or reset controls when available.
• Pause before replaying or sharing something designed to provoke a strong reaction.
• Talk about why a post appeared, not only whether it was entertaining.
• Take breaks from infinite feeds and autoplay when they make stopping difficult.
• Keep a record of repeated harmful recommendations and use BARK if an official report may be appropriate.

The aim is not to remove all fun or surprise. It is to remember that the feed is a designed selection, and to keep some choice about what trains your attention.

Rosie’s rule of paw

A recommendation is a suggestion, not a verdict about what matters.

If a feed keeps narrowing, change one input. Search outside the loop. Follow a source directly. Ask what the system may be learning from your behaviour. Then decide whether the recommendation deserves your attention.