The Best Onboarding Trick for AI Products Is Showing Users Less

Based on the research ofJiang, Lakhiwal, Liu and Duan, "Seeing Less, Engaging More: Rethinking Early User Experience on GenAI Co-Creation Platforms, Findings from a Field Experiment," Information Systems Research, 2026

That is the finding at the center of a new *Information Systems Research* paper by Shenyang Jiang, Akshat Lakhiwal, Che-Wei Liu, and Jiang Duan. Most generative AI content-generation (GCG) tools are built on the opposite assumption: that speed and completeness are the product. Type a prompt, get a finished image, a finished paragraph, a finished avatar, instantly, because instant gratification is supposed to be the whole pitch of the technology. The paper's authors call the design variable "fulfillment," the extent to which co-created content is revealed before a user is asked to register, and they test it across a spectrum from nothing shown to everything shown. Partial fulfillment beats both ends.

The Field Experiment, and the Online Replication

The researchers ran a randomized field experiment on a live GCG platform, then followed it with an online experiment to isolate mechanism. Users were assigned to see no generated output, a partial version of it, or the full output, before being asked to create an account. Partial fulfillment produced the highest registration rate of the three. The result held up when the authors layered in message framing: loss-framed prompts, the kind that tell a user what they stand to lose by not registering, raised registration on average, but that lift shrank under full fulfillment. When a user already has the whole finished asset in hand, there is nothing left to lose by walking away, so the urgency message has less to work with. Framing and fulfillment substitute for each other rather than stacking.

The mechanism analysis is the more useful part for anyone building a product roadmap. Both partial and full fulfillment increase what the authors term value-in-use, the user's recognition that their own input meaningfully shaped what came out. But only partial fulfillment sustains curiosity, the anticipatory pull of not yet knowing the rest. Full reveal answers the question before the user has a reason to keep going; zero reveal never lets them form the question in the first place. Partial reveal does both jobs at once, and the paper reports that the effect does not stop at the registration wall: it carries into later return visits and continued engagement, but only when the user actually co-produced the output rather than passively receiving it, and the effect strengthens as output quality improves.

The Instant-Gratification Default Most AI Products Still Ship

Set the paper's finding against how consumer generative AI products are actually built, and the mismatch is obvious. A large share of them optimize for the opposite of partial reveal: either everything is free until a usage counter runs out, or nothing is available until an account exists. OpenAI's ChatGPT free tier spent years capping message counts within a rolling window before loosening those limits for text prompts in 2026, while continuing to meter image generation, file uploads, and voice mode separately, according to OpenAI's own help center. Character.AI requires a Google or email sign-up before a user can chat with a bot at all, and has since layered on daily caps for regenerating replies. Midjourney went the furthest in the other direction: after a wave of viral deepfakes and, by its own account, abuse of multiple free accounts, it ended free trials entirely in 2023 and now requires a paid subscription before a user generates a single image.

None of these is the design the paper recommends. Rate-limited full access shows everything up to a wall, which flattens curiosity the same way full fulfillment does, one output at a time until the counter hits zero. No-trial-at-all, Midjourney's approach, is the paper's "no fulfillment" condition taken to its logical extreme: a user cannot experience value-in-use because there is nothing to experience before the paywall.

Where the Middle Path Already Exists, Quietly

The partial-reveal pattern the paper validates is not hypothetical; it already runs quietly under some of the most-used AI-adjacent tools on the web, just not usually framed as a generative-content product. remove.bg, the AI background-removal tool, processes every image for free but caps the free preview at roughly 0.25 megapixels, about 625 by 400 pixels, according to the company's own help documentation; the user can see exactly that the tool worked, cleanly and to their own image, but the full-resolution file sits behind a paid credit. Canva applies a version of the same logic to its AI-assisted design tools: content built from Pro-tier elements or AI features carries a visible watermark for free users, per Canva's own help center, and is removed only by buying a license for that design or upgrading to Canva Pro. In both cases, the user is not told to imagine the output or to trust a demo reel. They see their own result, partially, which is precisely the value-in-use and curiosity combination the paper isolates, even though neither product was likely designed with this literature in mind.

The teaser is not a withheld feature. It is the feature that makes the rest of the product worth wanting.

Adobe Firefly and Notion AI sit closer to the rate-limited camp: Firefly requires a signed-in Adobe account before generating anything and then allocates a monthly pool of free generative credits, per Adobe's help documentation, while Notion AI grants a one-time trial of a fixed number of free AI responses per workspace before it asks a user to upgrade, per Notion's own pricing page. Both are freemium-by-volume, not freemium-by-reveal. They ration how many times a user can get the full thing rather than how much of any single output they get to see, which is a different lever entirely, and one the paper's evidence does not speak to directly.

What This Means for a Roadmap

The practical rule is not "add friction." Friction for its own sake is exactly what most growth teams have spent a decade removing, for good reason. The rule is narrower: for a generative co-creation product, the moment right before registration is not the moment to prove completeness. It is the moment to prove that the user's own input mattered and to leave one visible reason to come back and see the rest. A locked variant, a blurred second half, a preview resolution, a single held-back edit, calibrated rather than arbitrary, will very likely out-convert both a hard paywall and a fully finished free sample. And once a product commits to full fulfillment for competitive or trust reasons, growth teams should expect their loss-framed urgency copy to do less work than it does anywhere else in the funnel, because a user staring at the finished asset has nothing left to lose by leaving.