Using Friction As A Feature In Machine Learning Algorithms — Smashing Magazine

Why Slowing Users Down Can Speed Up Machine Learning

Friction in UX is usually framed as a problem to eliminate. The smoother the path from intent to action, the better. But that assumption, while valid for many happy paths, misses a crucial point: resistance can be engineered deliberately to improve the quality of machine learning algorithms.

In the physical world, friction is the force that resists motion between two surfaces. In digital design, it is any element that makes an interaction less than instantaneous—a confirmation dialog, a multi-step form, or a well-placed pause. While these are often seen as annoyances, their most valuable function is preventing unintended consequences. A misclick on a high-stakes action, for instance, is far less likely when a modal forces a brief moment of reflection.

The Cost-Benefit Calculus of Resistance

Intentional friction is already standard in error prevention, a cornerstone of usability heuristics. In scenarios where mistakes carry serious consequences—like a permanent deletion—designers often add cognitive hurdles. Typing “DELETE” to confirm might seem cumbersome, but it acts as a powerful safeguard against the kind of careless click that Steve Krug famously warned about in Don’t Make Me Think: users don't read pages; they scan them.

This approach is not about making things difficult for its own sake. It is a cost-benefit analysis. If the rewards of a streamlined flow are outweighed by the risks of accidental catastrophe, friction is the correct trade-off. The same principle applies to subtler, non-catastrophic problems. Social media platforms, for example, have used friction to nudge users away from impulsive behavior. Prompting users to read an article before retweeting it, or warning them before they post harmful language, adds a step but works to improve the overall quality and trustworthiness of the platform. In these cases, the friction is meant to change behavior, not just confirm intent.

Personalization Requires Input

When applied to machine learning, friction takes on a more profound role. Recommendation algorithms are only as good as the data they receive. An app that understands a new user's preferences does not do so by magic; it gathers that information, often through an onboarding flow. These flows, which ask users to select interests or rank content, are a form of friction that feels justified because they solve a fundamental problem: no one wants the fastest path to something irrelevant.

Major platforms rely heavily on this mechanism, with a significant percentage of purchases and viewership stemming from algorithmic recommendations. A user who churns is often one who didn’t understand the product or get value from it. Onboarding directly addresses both issues by educating the user and immediately personalizing the experience, even if it comes at the cost of a few extra clicks before the user can explore.

The design tension here is inherent. An algorithm improves with more data touchpoints, but each new question risks alienating a new user. Teams must carefully balance the need for rich data against the need for a smooth first impression. Some companies find that adding steps yields surprisingly little dropoff, especially when the resulting personalization improves retention enough to offset any initial friction. The right answer is never universal; it depends entirely on the specific algorithm’s needs and the product’s audience.

Algorithm-Friendly Design as a Product Strategy

TikTok’s recommendation engine feels almost prescient, which stands in contrast to its onboarding flow: sign up and browse, no preference selection needed. Instagram, by comparison, guides new users through multiple setup steps yet lacks the same reputation for personalization. The difference comes down to interface design that turns everyday behavior into training data for recommendation models — an approach some describe as giving an algorithm “vision” or, in Spotify’s words, “glasses.”

Signals Hidden in Plain Sight

Algorithm-friendly interfaces prioritize bidirectional feedback: the system responds to user actions while also collecting data on whether its outputs meet expectations. Everything from button taps to the absence of action becomes a signal. Apple’s Machine Learning Research Department has documented this effect, using engagement signals as input for Reinforcement Learning from Human Feedback. Their findings describe “significant accuracy gains in a production deep learning system,” suggesting that naturally occurring behavior patterns are sufficient to build accurate personalization models without explicit preference capture.

Instagram applies similar tactics, though less cohesively, as its approach remains in flux. The principle holds regardless: interaction design that captures implicit intent can replace upfront preference collection.

One of TikTok's key design decisions was showing only a single video per viewport. This reduction in information density masks friction but clarifies intent — a swipe or pause can be attributed to specific content with confidence. Spotify CTO Gustav Söderström calls this “giving the algorithm glasses,” arguing these focused interfaces outperform the traditional feed of mixed content for signal extraction.

Friction-as-a-tool helps separate meaningful engagement from noise. Although every interaction can in theory provide feedback, the quality of that signal depends heavily on the interface framing it. A single-content viewport keeps user attention (or lack thereof) aligned with one item, making every action — or inaction — interpretable.

Where the Tradeoff Works

Evaluating this design choice requires looking at the broader user hierarchy of needs. Finding relevant content quickly typically outweighs minor UI friction, since each interaction, while individually slower, potentially accelerates discovery more than a faster but less informative interaction would. Over longer timescales, compounding personalization gains can make an interface feel more frictionless over quality: as recommendations improve accuracy, each subsequent session demands less user effort. The same algorithms run behind both paradigms — one just has a more feed-forward interface for signal collection.

“The secret of why some of these products are so good at recommendations is not actually that they have better algorithms. It’s the same algorithms with a more efficient user interface.” — Gustav Söderström

This interface template wasn't invented by social video: dating apps relied on single-card swiping and directional gestures as engagement signals since Tinder's debut in 2012. That paradigm limited viewport congestion while allowing actions to modify recommendations. TikTok pushed the model mainstream, but it wasn't first — just more broadly applicable, since dating app usage tends to churn once a user finds a match.

Instagram and Spotify have since pursued similar redesigns into TikTok-like layouts, often facing pushback and appearing on “worst updates” lists. Spotify's leadership frames this as inevitable, betting that signal clarity over time outweighs interim resistance. App makers risk losing competitive positioning to superior insight-generation if their interfaces stay inefficient at data collection, even when product iterations meet transient criticism.

Compounding Data as a Feature

Strong engagement signals create data compounding effects. An individual’s generated signals improve the output quality not just for their personal feed but for the entire system. A single user's noise amplifies slightly; bad recommendations spread less if other users’ interactions take away their visibility, making the platform partially antifragile — able to improve after poor outcomes instead of merely resisting cumulative damage.

The approach dates back at least to Amazon's item-to-item collaborative filtering in 2003 (now synonymous with “customers who viewed this also viewed this”), which derives its power from inter-item relationships. The algorithm learns even from sparse user activity, extrapolating preferences from a handful of actions, making pairwise connections across a dataset perform better with collective input rather than individual history. Two or three clean data points can reliably plot a user’s preference space.

Personalized feeds like TikTok therefore need surprisingly little to get started: a few swipes effectively chart a new user’s tastes. The result is that apps deliberately generate good, low-volume feedback instead of relying on abundant but undeveloped click streams feeding offscreen model weights.

Where Friction Is Heading

Friction has moved from a UX afterthought to a core design tool, and the rise of machine learning is accelerating that shift. Social platforms like TikTok already demonstrate how implicit engagement signals can replace explicit onboarding. The next phase is broader: as large language models make it possible to process unstructured data at scale, algorithms can start recommending more than just media.

Feature recommendation is one promising direction. Instead of assuming users will discover buried functionality, an interface could suggest relevant tools based on patterns from other users. Microsoft’s Copilot push is an early example, with the company noting that the average person uses less than 10% of what PowerPoint can do.

This approach could also challenge the widely cited Pareto Principle — the idea that 80% of users only ever need 20% of features. As Joel Spolsky has pointed out, that assumption tends to break down in practice:

“A lot of software developers are seduced by the old “80/20” rule. It seems to make a lot of sense: 80% of the people use 20% of the features… Unfortunately, it’s never the same 20%. Everybody uses a different set of features.”

If the value of features is distributed unevenly across individuals rather than uniformly small, then the problem shifts from simplification to routing. No one would suggest that YouTube should delete most of its videos; the goal is to surface the right content to the right person. Feature discovery could work the same way, and algorithmic recommendations would need the same kind of usable, low-friction interfaces to be effective.

Hallucination as a Design Question

Jakob Nielsen has called LLM-driven interaction “the first new UI paradigm in 60 years,” describing it as intent-based outcome specification. Rather than instructing a computer on what to do, users describe the outcome they want. ChatGPT’s popularity shows the appeal of this model, but it also exposes a serious issue: the system sometimes fabricates responses.

In one widely reported case, a lawyer relied on ChatGPT for legal research only to discover in court that the cited cases did not exist. The user’s defense — that he was unaware the content could be false — underscores how poorly interfaces communicate the technology’s limits. An empty-state disclaimer is not enough when the consequences of blind trust are high.

Extra friction, such as confirmations or mode selection, could better educate users about the risk of “hallucinations” — outputs that confidently stray from the training data. Some observers, including Marc Andreessen, argue that this behavior is better described as creativity, and there is value in a system that can brainstorm or improvise. But a one-size-fits-all chatbot makes that nuance invisible. Users who want a deterministic, verifiable calculation and users who want an imaginative exploration have different needs. A creativity slider, a persona selector, or other explicit controls could align expectations without sacrificing the magic of the technology.

Putting Friction into Practice

Adding deliberate resistance is easier said than done. A few principles can help guide the implementation.

Design for the Peak and the End

Research by Daniel Kahneman and Amos Tversky suggests that people remember experiences not by their total duration but by their most intense moment and their final moment. This peak-end rule is useful for friction design. A long or complex flow can still feel successful if it contains positive peaks and ends on a rewarding note. That shifts the conversation away from raw step counts toward emotional payoff.

Treat Metrics as Approximations

Friction is experienced emotionally, but it is often analyzed numerically — a map that does not fully capture the territory. Analytics may not register a user who deletes the wrong file and curses at the screen, nor the session spent recreating lost work. Similarly, a process that adds steps but yields better training data might not score well on a simple funnel chart, even though the downstream algorithmic gains compound across the model. Measuring friction properly requires a holistic view of the product and its second-order effects.

Keep Iterating

As AI becomes more central to software, the applications that survive will be the ones that adapt their interfaces to a new interaction model. Friction, used thoughtfully, is a way to make those systems safer, more transparent, and more aligned with what users actually need. The instinct to smooth every path may be exactly the wrong instinct in the machine learning era.