Notifications as behavior design

Source: The Ridiculous Engineering Of Notifications, Enrico Tartarotti, 13:15, uploaded 2026-07-08.

A notification is the rare product surface that starts the interaction itself. It can arrive during an idle moment, supply a reason to act, and repeat that prompt until the action becomes routine. The video’s argument moves through three layers: the engineering of one prompt, the conversion of prompts into habits, and the new AI intermediary that can rewrite the prompt before it reaches the person.

Three layers

1. Message design: The product chooses a detail, an open question, a possible loss, and a delivery context that make the prompt hard to ignore.
2. Habit design: Repeated prompts move an action from an outside cue toward a routine that can continue without the cue.
3. AI mediation: The operating system can now rewrite, shorten, group, reorder, or suppress the message before the person sees it.

Specificity beats shallow personalization

The first example begins with a generic Google Photos message: “view some memories.” Adding a first name would technically personalize it and still leave it as spam. The proposed reconstruction uses the person’s archive to find a genuinely particular event.

• Rank matching calendar days from previous years by photo count, since trips and parties often produce more photographs than ordinary days.
• Read EXIF location data to distinguish travel from everyday pictures near home.
• Use the visual-analysis tags Google Photos already creates for search to separate a skiing trip from a folder of photographed documents.
• Feed the useful tags through a language model to produce a playful label and pair it with a preview image.

The result feels simple, but it depends on historical analysis, visual recognition, location data, and language generation operating at the scale of millions of archives. This is the video’s first important distinction: putting a name into generic copy is personalization; selecting a recognisable episode from someone’s life is specificity.

Specificity is not always the goal. Sometimes the message should construct a precise unanswered question.

• Loom withholds the viewer’s name. The missing detail creates a question that the app can answer only after it opens.
• Hinge can reveal a name while withholding the rest. This changes the question from “who liked me?” to “what is this person like?”

Specificity resolves uncertainty, while a curiosity gap preserves selected uncertainty. Neither alone guarantees action, so the Duolingo example adds loss aversion: the notification does not offer another Spanish lesson; it warns that an existing streak is about to disappear.

Timing then becomes context rather than a generic hour on a marketer’s schedule. The video says Duolingo waits roughly 23.5 hours after the previous session, because that is a moment when the person has already proved they can practice, but caps delivery before 10 p.m. so action remains possible. It attributes a roughly 60% lift in taps to timing alone, and opens with larger reported effects: a 40% reduction in subscription cancellations and 4.5-times growth in daily active users. These figures make the engineering sound decisive, but the video does not provide links that would let me verify them.

Notifications as the push channel

The history matters because notifications did not begin as a general-purpose retention surface. A BlackBerry’s blinking LED indicated something waiting now, strongly enough to help produce the “CrackBerry” nickname. Apple’s first iOS notification design was similarly blunt: a full-screen interruption with no Notification Center. The message occupied the immediate-action end of a time spectrum.

Email sits at the opposite end. It also arrives immediately, but an inbox makes it durable: deal with it later, perhaps tomorrow, perhaps never. The video cites an analysis of more than 200 million notifications to add a third window. Message notifications are often tapped within about 15 seconds, while only 26% of the notifications shown at the top of the phone led to an immediate tap. Many prompts instead operate in the next-hour zone: not urgent enough to interrupt, but persistent enough to shape what happens soon.

Repetition matters more than one tap. The stated Duolingo model is that a seven-day streak makes a person 2.4 times more likely to return the next day. The prompt first produces the visit; the maintained streak later supplies its own pressure. This fits gamification that changes behavior and variable reward and habit loops in social apps.

This leads to the video’s push-versus-pull distinction. Scrolling still requires boredom or some other internal cue to make a person pick up the phone. The app waits to be pulled. A notification reaches into the person’s day and creates the moment on the product’s schedule.

The cited experiment in which participants disabled all phone notifications complicates the obvious remedy. They completed more work, but also felt more anxious, lonely, afraid of missing out, and worried about letting other people down. They continued checking their phones without prompts, and every participant eventually restored notifications. Removing the push channel did not remove the learned pull.

The ethical result depends on the behavior and the business model. A prompt can support language practice, exercise, or contact with friends. The same habit can raise retention, sell storage, or make cancellation feel like a loss. Google Photos can produce a genuinely lovely exchange between old friends while also making its storage subscription harder to leave. “Useful” and “commercial” can be true at the same time.

AI becomes an editor

Notification design used to assume a direct path from product to person. OS-level AI adds an editor between them. The carefully specific Google Photos message can emerge from Apple’s summary as the same generic message the product team worked to avoid.

• A summary can remove the exact detail that made the original message useful.
• A model can attach a false claim to a trusted app identity. The video uses Apple’s incorrect summary of a BBC notification, which falsely stated that a man had shot himself and placed the invention next to the BBC logo, as the main failure case.
• Reordering and suppression can change which sender gets attention, even when the original text remains accurate.
• Email already demonstrates the strange endpoint: one model writes a long message and another summarizes it back down for the recipient.
• Product teams may start to add machine-readable intent, urgency, and context for agents, much as llms.txt addresses machine readers on the web.
• The design target then shifts: the message must first convince an agent that it deserves to reach a person.

That shift creates a control problem. The sender knows why it wants attention. The operating system knows more about the person’s current context. Neither should quietly decide that product growth and user value are the same thing. This belongs near human-centered AI and explainability and marketing as context.

Caveats

The video reports large effects from notification timing and design, a dataset of more than 200 million notifications, and research on disabling notifications. It does not name or link the studies in its description. Treat the figures and the claim that every participant restored notifications as leads, not verified evidence.

The argument also moves quickly from current AI summaries to future notification agents. Rewriting and grouping already exist; agent-specific notification metadata is the author’s forecast.

Further reading / references

• Hooked: How to Build Habit-Forming Products, named in the video as the habit-design model.
• Figma Design Agent, shown in the sponsored segment.
• Related: dopamine and product use, revolt against technology, perception, attention, and memory in interface design.