UX Psychology for AI Products

There's a good breakdown of UX psychology doing the rounds — uxpeak's "The UX Psychology Behind Apps People Can't Stop Using" — covering six principles with a before-and-after screen for each: smart defaults, endowed progress, reciprocity, the IKEA effect, loss aversion, and anchoring.

The principles are sound and the research behind them is real. But every example in it is a conventional app: a booking form, a loyalty card, a storage upgrade, a protection plan at checkout.

We build AI products, and these principles don't transfer cleanly. Some get dramatically stronger. Some become actively hostile. And AI products have a trust problem that none of the six principles address at all. Here's what we've found actually applies.

1. The blank prompt box is the worst form ever designed

The original point: empty forms cause decision fatigue. The research is the Columbia jam study — 24 flavours on display, 3% of shoppers bought; six flavours, 30% bought. And 70–90% of users never change a default, because they read a default as a recommendation.

Now consider what almost every AI product opens with: a single text box and the words "What can I help with?"

That's not a simplified form. It's a form with one field and infinite valid values, which is the worst case for decision fatigue. The user has to invent the task, phrase it well enough for a model, and guess the system's capabilities, all before pressing anything. "Blank canvas paralysis" is the single biggest drop-off point we see in AI onboarding.

Suggestion chips are the smart default here. The engineering part matters more than the design part: they have to come from real usage data, not from a brainstorm. Pull the top queries that actually succeeded — ones where the user didn't immediately rephrase, and where the answer got a thumbs up — and offer those. Suggestions invented by the team tend to showcase what the team finds impressive rather than what users actually want, and they set expectations the model can't meet on the second question.

2. Endowed progress is honest in AI products — because the wait is real

The original point: the car-wash study, where a loyalty card with ten stamps and two pre-filled beat an eight-stamp empty card at nearly double the completion rate, for the same eight washes. Never start the user at zero.

Here's where the video undercuts itself. It warns against faking progress in its own description, then says "progress, even a fake one, creates real momentum." Both things can't be true for long. A progress bar that doesn't track real work is a lie the user discovers the first time it sits at 80% for a minute, and the trust you lose is worth more than the completion you bought.

The good news is that AI products rarely need to fake it. We have genuinely slow, genuinely multi-stage work to show:

Indexing a 200-page PDF takes real time. Naming the stage you're in converts dead waiting into visible progress, and it's all true. You get the motivational benefit of endowed progress and a debuggable system — when a user says "it got stuck", they can tell you which stage.

The honest version of "never start at zero" is to count work the user has genuinely already done: the document they uploaded, the integration they connected, the first question they asked.

3. Reciprocity is the strongest lever an AI product has, and most waste it

The original point: give before you ask. Cialdini rates reciprocity the most powerful driver of behaviour; free samples lift purchases by up to 2,000%. The video's example is a site-audit tool that blurs the report behind "create an account to see your results" versus one that shows the real score and top issues first.

This matters more for AI than for anything else, because one real inference is the most persuasive artifact you own. Nobody believes a feature list about AI any more — the category has burned too much credibility. One actual answer, over the user's own data, does what a landing page cannot.

Most AI products still gate it. Upload your document, sign up to see what we found. That's a restaurant asking for your card before showing the menu, except worse, because the user also can't tell whether the output would have been any good.

The reason teams gate it is real, though, and it's an engineering problem rather than a design one: inference costs money, and an ungated demo is an invoice someone can run up for you. So meter it properly — rate-limit by session and IP, cap tokens per anonymous request, use a smaller model for the free first answer, and cache aggressively on common inputs. We built Marian around showing a real answer with citations back to the source passage; the citation is doing double duty here, because it's also what makes the free answer verifiable.

Give one complete, genuinely useful result. Then ask.

4. The IKEA effect is really about context, and context is the moat

The original point: people value what they built. Duolingo gets you ten minutes in — language picked, goal set, first lesson done — before it ever shows a signup screen.

For AI products the thing users build isn't a profile or a colour palette. It's context: the documents they uploaded, the prompts they saved, the corrections they made, the conventions the system learned about their work.

This is worth being clear-eyed about, because it's also the honest answer to "what's your moat when the model is a commodity?" It isn't the model. It's that a user who has spent three weeks building a corpus and correcting the output has something no competitor can hand them on day one. Docsmith is built on exactly this premise — a team's runbooks, conventions and decisions become shared memory that both the people and the agents work from.

The design consequence: let users build context before they register. One document, one question, one correction. And when they do register, carry it over — nothing destroys this faster than making someone sign up and landing them in an empty workspace.

5. Loss aversion is where this gets dangerous

The original point: Kahneman showed losing hurts roughly twice as much as gaining the equivalent feels good. The video's example swaps a storage app's "Upgrade now / Maybe later" for a screen naming the user's actual files with a countdown, where the dismiss button reads "I'll risk it".

We'd push back on that one. Threatening a user with the deletion of data they created is not persuasion, it's hostage-taking, and it's the clearest dark pattern in the video. Users remember it. They tell other people about it.

The framing does work, but point it at capability rather than at their data:

The second is still loss-framed, still more motivating than a feature list, and still true. For AI products the natural version of this is usage: credits, context-window limits, how many documents you can keep indexed. Those are real constraints with real costs behind them, so you can state them plainly without manufacturing anything.

6. Anchoring works, but AI pricing is already incomprehensible

The original point: $50/month in isolation reads as $600/year and gets declined; the same $50 beneath a $1,900 laptop, labelled "just 2.6%", barely registers. Never show a cost in isolation.

AI pricing has a harder problem than framing: most users genuinely cannot tell what they're buying. Tokens, credits, seats, requests per month, and a context window measured in a unit nobody thinks in. The anchor can't fix incomprehension.

So anchor against something the user can verify from their own experience — the hours of manual work the task replaces, or the volume they've already processed in the trial. That's checkable, which means it survives scrutiny. An invented comparison ("normally $499") doesn't, and the moment a user catches it, every other number on the page becomes suspect too.

The principle none of the six covers: AI output has to be verifiable

Every one of these principles assumes the product works. For conventional software that's mostly a safe assumption — a booking form either books or it doesn't. Model output is probabilistic, occasionally confidently wrong, and the user often can't tell which case they're in.

That changes the job. The most important UX work in an AI product isn't reducing friction, it's making the output checkable:

• Citations that point at the exact passage an answer came from, not a document-level "source".
• Saying plainly when nothing relevant was found, instead of generating something from irrelevant context. A system that admits it doesn't know is worth more than one that's right slightly more often.
• Showing which tools ran and what they returned, so a wrong answer is diagnosable rather than mysterious.

No amount of endowed progress or smart defaults saves a product the user has quietly stopped trusting. We've written about this from the engineering side — retrieval quality treated as a measurable, defended metric rather than a vibe — and it's the same concern wearing a different hat.

Where the line is

All six principles are legitimate when they do one of two things: remove friction from something the user already wants, or show real value earlier. Smart defaults, honest staged progress, a real free answer, carried-over context, true constraints, verifiable comparisons — all fine.

They become dark patterns the moment the progress is fake, the scarcity invented, the loss manufactured, or the anchor made up. The mechanics are identical; the difference is whether the thing you're telling the user is true.

The practical test we use: if the user found out exactly how this screen was designed, would they feel helped or handled? It's a low bar, and it rules out most of what gets called growth hacking.

Framing and research citations from uxpeak's UX psychology breakdown. The AI-specific application, and the disagreement in §5, are ours.