Skip to content

Eight AI Services, One Design Question: Who Holds the Permission and the Memory

The service field of the 2026 International AI Creator Award collected eight AI products from Japan, Vietnam, Korea and the US. Read as design specs instead of demos, they point at the same two problems: what a synthetic persona is allowed to say, and who keeps the context when the model changes.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
Three translucent glass panels standing in a row on a dark surface, each etched with concentric rings and lit from behind

Awards for AI work usually collect pictures. The service field of the 2026 International AI Creator Award, run by the Japanese outlet AICU, asked for something harder to photograph: products where AI takes over someone's labor or someone's worry. The organisers put it plainly. These entries are not things to look at, they are things to use, and the descriptions are written around who gets helped.

As of November 5, the public gallery holds eight entries. Four from Japan, two from Vietnam, one from Korea, one from the United States. That distribution is the first interesting data point: in a field where ChatGPT or Claude could have entered, half the work arrived from outside Japan.

I read the eight as design specs rather than demos. Four questions keep coming back: who gets credited for the work, what a synthetic persona is allowed to say, who holds the record of what it said, and what happens to all of it when the model underneath gets replaced.

Radi.fm: say which layer is the machine

Radi.fm by NANMOKU Toworu is a classical radio station that starts playing the moment you open the page, in six language editions: Classical Global in English, 中文古典音乐 in Chinese, クラシック Nippon in Japanese, Clásica Global in Spanish, and two more. AI builds the schedule, picks the music, writes the scripts, answers listener messages and reads them aloud.

The music and the listener messages are all made by people. The author's framing is that AI runs the venue.

That division is the whole product. A fully generated radio station would be cheaper and worse, and listeners would feel it without being able to name it. By drawing the line in public, and by keeping the part that carries emotional weight human, the AI stops being the author and becomes operations. The same rule applies to creative tools: automation is easier to trust when the surface says what it touched.

A Vietnamese entrant, Dr. Yulia B., asked the sharpest question in the gallery comments: does the AI host keep the same personality and delivery across six languages? That is not a translation question. A host persona is voice, pacing, editorial judgment about what is worth saying, and how much silence to leave. Most multilingual products treat that as a locale file and end up with six different people reading the same script. If personality has to survive language, it needs to be specified as behaviour, then checked by listening to the same segment in every language side by side.

OpenClips: in a multi-model market, the product is the router and the receipt

OpenClips, from Nina Rasovic in the US, makes 30-second clips and lets you choose between Seedance, Veo, Kling and Hailuo from a single screen. You attach reference images, pick the model that suits each shot, and pay per render out of one credit pool. The site shows a public feed of clips, and every clip carries its prompt and the steps that produced it. Anyone can remix one in a click, with the original author's name kept on it.

Two decisions there are worth copying. First, per-shot model choice: video models differ enough that a single-shot workflow wastes money on the easy shots and under-delivers on the hard ones. We made a similar argument about cost routing in Wan 3.0; the difference here is that routing is the product rather than an internal optimisation.

Second, prompts as public receipts. Most platforms hide the prompt because it feels like the secret sauce. OpenClips publishes it and makes remixing attributed. That trades a small moat for a large distribution surface, and it turns the feed into a place where the technique, not just the output, is the content. If your clips are already public, withholding the prompt protects nothing and costs you the remix traffic.

Permission tiers beat a privacy switch

Dr. Yulia B. entered two connected pieces from Vietnam. The first is PRIVACY-FIRST AI IDENTITY: an AI avatar named Ava speaks to visitors in multiple languages on the owner's behalf, and the owner decides what is speakable in three tiers. PUBLIC, PROTECTED, FORBIDDEN. The avatar relays only approved work information. Location, personal details and the owner's own appearance stay out of it.

Most tools ship a binary: the assistant is public or it is private. That is the wrong shape, because real professional disclosure is graded. Your job title is public. Your rate card is protected, available on request. Your home address is forbidden. A three-tier model maps to how people actually decide what to reveal, and it gives the owner a place to put the awkward middle category that a switch forces them to round up or down.

The second piece is an AI AVATAR IDENTITY PASSPORT: a designed document for an AI-born character that bundles appearance, a structured identity record, certificate-style presentation and a multilingual video into one artifact. Put the two side by side and you get a coherent pair, one governing what a persona says, the other establishing who it is when it says it. As synthetic characters accumulate clients, credits and history, some portable proof of identity stops being a novelty.

STAY: memory should outlive the model

The entry I keep thinking about is the least visual. STAY Platform / STAY OS, by Changwoo Lee in Korea, grew out of a book of dialogues called The Discovery and turns its questions about memory, relationships, continuity, correction, permission and human agency into a working system. The design goal is that when the underlying model is replaced, the memory and the relational context stay with the person. The line on the cover is the argument: models change, memory should stay.

For anyone building on top of model providers, this is the least optional idea in the eight. Every vendor swap, deprecation or price change currently costs users their accumulated context, because that context lives in the vendor's thread history. Creative tools have the same exposure in a milder form: a project's approved references, style notes, rejected directions and shot decisions are the actual asset, and they usually live wherever the last generation happened.

Keep that outside the model. Store the brief, the references, the approvals and the reasons, in a form you can replay against a different engine. The model is a component that will be replaced; the project memory is the thing the client is paying to keep.

What the field itself got wrong

Two honest caveats. The categories are self-declared, and one entry, TIME STEP, is a Taisho-era romance film about tap dancers that you watch on TapNow. Beautiful premise, but it is a narrative work filed under services, which tells you the field boundaries are still soft. The organisers also entered their own magazine issues, including a GPT-Image-2.5 special, so the award is partly a publication platform.

And all of this is design intent, not traction. Eight prototypes with a public vote say nothing about retention or willingness to pay. The vote runs until November 13, 2026 (Anywhere on Earth), which is November 14 at 21:00 Japan time, one vote per field and free.

Still, the pattern across four countries is consistent enough to borrow. Label which layer is machine and which is human. Grade disclosure instead of toggling it. Keep the prompt and the memory portable, and keep the author's name attached when someone builds on the work. None of that is generation quality, and all of it decides whether the thing survives its second model version.

Sources

Share this guide
X LinkedIn