The Half-Finished Draft Converts Better Than the Finished One

Based on the research ofShenyang Jiang, Akshat Lakhiwal, Che-Wei Liu, and Jiang Duan, "Seeing Less, Engaging More: Rethinking Early User Experience on Generative AI Cocreation Platforms, Findings from a Field Experiment," Information Systems Research, 2026

The Co-Creation Illusion That Backfires

Generative content-generation platforms let a stranger type a prompt and watch a finished image, song, avatar, or paragraph appear in seconds, before any account exists. The obvious product instinct is that speed delights and delight converts. Jiang, Lakhiwal, Liu, and Duan test that instinct directly, running a randomized field experiment on a live generative AI cocreation platform, backed by a follow-up online experiment, and find the opposite of what most roadmaps assume.

Their design lever is what they call fulfillment: how much of the co-created output a visitor sees before being asked to register. Full fulfillment shows the finished piece, complete, on screen. No fulfillment shows nothing, just a promise that something was made. Partial fulfillment shows some of it and withholds the rest. The authors argue two motivational pathways are in tension. Value-in-use is the user's sense that their own input meaningfully shaped what they are looking at. Curiosity is the pull of an experience that still feels open, not yet resolved. Full fulfillment starves curiosity, because there is nothing left to wonder about, and it can even dull value-in-use, because a complete, polished artifact reads as the model's achievement rather than the user's. No fulfillment starves value-in-use outright, since there is no evidence the user's prompt did anything at all. Partial fulfillment is the only condition that can feed both at once, and in their experiment it outperformed both extremes at driving registration.

That is a genuinely counterintuitive result for anyone optimizing an AI product the way most teams default to optimizing one, which is to make the output as fast, as complete, and as impressive as possible, as early as possible. The paper's finding says that completeness is not free. It is a lever that can work against you at the exact moment you most need a visitor to decide the product is worth an account.

Most Generative AI Products Default to the Wrong Extreme

In practice, pre-registration AI experiences tend to swing between the two losing conditions the paper identifies, rarely landing on the partial middle. Midjourney's original Discord trial gave anyone a batch of free, complete, full-resolution images with no account required beyond a free Discord login, pure full fulfillment. In March 2023 the company suspended free trials entirely, citing extraordinary demand and trial abuse, and as of mid-2026 generating an image requires at minimum a paid Basic plan. That is a jump from full fulfillment straight to no fulfillment, with nothing in between ever tested at scale.

Downstream of registration, the same binary shows up again. OpenAI has reportedly been testing a watermark on images generated through ChatGPT's free tier, with the option to save a clean, unwatermarked version reserved for paying subscribers, a detail multiple outlets picked up from beta code in 2026. Functionally, every free-tier image is still complete; only the mark on top signals that something remains locked. It is a cheap, partial-looking cue bolted onto a fully fulfilled artifact, rather than genuine partial fulfillment of the generation itself, and it sits at a different gate (free to paid) than the one this paper studies (anonymous to registered). But it shows the same instinct product teams reach for once they notice that "finished and free" has a cost: mark the output as unclaimed rather than actually withholding part of it.

Suno's free tier, by contrast, still leans toward full fulfillment at the cheapest possible gate. Multiple guides describe free accounts generating up to ten complete songs a day with no credit card required, once a user signs in with a Google or Microsoft account. The generation itself, prompt to finished track, is complete every time. The open question this paper raises for a product built that way is whether handing over a finished, listenable song the instant someone clears the lowest-friction login is quietly undercutting the harder conversion that follows, deeper subscription, commercial rights, higher-fidelity models, because the first song already felt like a finished transaction rather than an open one.

The Loss-Frame Substitution, and Why It Matters for Your Copy

The paper's second finding is easy to miss and arguably more useful operationally than the headline result. Loss-framed registration prompts, the "don't lose this" style of message, increase registration on average. But that lift shrinks under full fulfillment, a substitution relationship between the design of the experience and the design of the message. The likely reason is straightforward once you see it: a message that threatens loss needs something left to lose. When the artifact on screen already looks finished and fully possessed, telling the user they will lose it rings hollow, because nothing about the experience still feels unresolved. Under partial fulfillment, where part of the output is visibly still withheld, the same loss-framed line has real teeth, because there is a concrete, specific thing, the rest of the image, the rest of the song, that registering would actually unlock.

Content that arrives finished in one second asks nothing of the user, and leaves nothing for them to come back for.

Most growth teams A/B test subject lines, button copy, and urgency language in isolation from the product experience that copy sits on top of. This paper's result says that is backwards for a generative AI onboarding flow specifically: the fulfillment level set upstream determines whether a given message frame has anything left to work with. A loss-framed CTA bolted onto a fully revealed, finished piece of content is optimizing the wrong variable.

The Partial Reveal Pattern Already Works, Just Not Yet in AI

Publishers solved a version of this problem decades before generative AI existed, and their data back the paper's logic. Digiday's reporting on registration walls, citing Piano's subscription benchmark data, found that registered users are roughly 45 times more likely to convert to a paying subscriber than non-registered visitors, and that registered readers view roughly five times more pages than anonymous ones. The mechanism publishers lean on is exactly partial fulfillment: show the headline and the opening paragraphs, withhold the rest, ask for an email before the article closes. Nobody runs a news site on full fulfillment, giving away the entire piece free with no registration ask, or on no fulfillment, hiding the article behind a wall with no preview at all. The entire registration-wall industry is a decades-long experiment in finding the partial middle this paper finds for generative AI.

Subscription benchmarking from RevenueCat's 2026 State of Subscription Apps report shows a parallel pattern on the monetization side: across more than 115,000 mobile apps, offering any kind of trial period before the paywall roughly doubled median conversion compared with a hard paywall shown immediately, evidence that some structured partial access beats an all-or-nothing gate even downstream of registration. Generative AI products have mostly not imported the registration-wall playbook to the point of co-creation itself. They gate by quantity, a daily cap on generations, or by quality, a watermark or resolution ceiling, rather than by revealing part of any single generated artifact the way a metered article reveals part of a single story. A small, preliminary literature on the "effort heuristic," including a 2026 preprint on what its authors call a struggle premium, offers a cognitive reason why that specific move, revealing part of one artifact rather than capping access to many, should work: people price perceived effort into how much they value a creative output, and content that appears complete in zero visible time offers no effort cue at all, human or otherwise.

The Rule for Any Generation-First Product

If your product's core move is producing something complete in seconds, the completeness itself is working against your registration funnel, not for it, unless you deliberately interrupt it. The instinct to make generation faster, higher-fidelity, and more finished before anyone signs up is optimizing for an emotion, the wow of watching content appear, rather than the two things that actually predict registration: did the user feel they shaped it, and is there still something open enough to pull them forward. A generator that nails both by showing some of what it made and holding the rest back converted better in this paper's field experiment than one that either gives everything away or gives nothing at all. Before your next release ships a faster, more complete generation pipeline, ask whether you are also removing the one thing that made the old, slower version worth registering for.