Originally published in TL;JP, a weekly, fact-checked read on what Japanese AI and tech practitioners are saying on X.
For the week of September 7-13, 2026, the Japanese timeline I follow pulled in two directions at once: a lot of quiet excitement about agents doing real production work, and a lot of anger about AI work that nobody checked before shipping. Both are worth your attention, and they turn out to be the same story.
1. Astra handed back editable vector art, not a flattened picture
yu_webdesigner (2,406 likes) says they took an image generated with GPT Image 2, passed it to Astra inside Illustrator, and got back decorated text whose appearance was reproduced — fill, outline, layering — still in an editable state. They say they deliberately picked an image that looked hard to reproduce. Why it matters: the expensive part of AI art in production isn't generating it, it's that you can't edit it afterwards.
Signal — a first-hand test aimed at a specific failure mode, not a highlight reel. link
2. "Don't make it build from zero. Make it copy your template."
0xfene (1,075 likes) shared their settled workflow for slides now that GPT-6 Astra is out: prepare a reference deck as PDF (they recommend your company template), use Claude Code or Codex to turn the PDF into pptx, then import into Figma or Canva.
Signal — a reproducible workflow with named tools, not a claim about model quality. link
3. Level-5 became the week's cautionary tale
Three separate posts piled onto the same showcase. gin_camp (1,531) argues so much generative AI was used that nobody can tell which frames are actual gameplay and which are generated imagery, which drains the point of releasing information at all. retrogame1989 (512) says there's no visible sign of efficiency or speed gains — just a cheaper-looking showcase, "like a movie trailer where you can see the bare set." artagongulgul (660) says they swore off the company three years ago when its Stable Diffusion use was announced, and won't play the Layton title they'd been waiting for. These are opinions from viewers, not verified facts about the studio's pipeline.
Signal — three unconnected accounts converging on the same reputational damage. gin_camp / retrogame1989 / artagongulgul
4. One illustrator's tally of how the market actually moved
ShirosaiNekocha (1,636 likes) revisited the "this will end all illustrators" predictions and lists what they found had changed by 2026: PIXTA halting sales of all AI stock material, publishers and game companies asking for proof of AI non-use in contracts, and a sharp rise in referrals from veteran illustrators. This is one person's summary with no sources attached.
Signal — if even partly right, it's a procurement fact, not a culture-war take. link
5. The real cost of "I made it with AI, please review"
zen_lele (1,919 likes) names why submitting unreviewed AI output is a social offense: you hand effectively all the work to the reviewer, you get credit for effort while spending no working time, you add the chore of hunting AI-specific defects, and anyone who objects gets labeled as behind the times.
Signal — the clearest articulation I've seen of where AI cost actually lands. link
6. 3D: a rigged character in half a day, and a vendor pinned down on training data
Botan_cr (1,160) relays creator Dstudio_ai's pipeline: generate hair, head and body separately in Tripo, then join, rig and set up the jiggle physics in Blender with Astra doing most of it, plus a self-built rigging helper app — model finished in under half a day. Second-hand, no files shown. Separately, VtuberOCM (353) says they spoke with someone from Meshy and confirmed that when used via API, neither inputs nor outputs are used for training.
Signal — one is a named end-to-end pipeline, the other is the vendor question Western teams keep forgetting to ask. Botan_cr / VtuberOCM
Also seen, and skippable
ComfyUI (602) showing a shot generated in under a minute with MiniMax H3 Max via Partner Nodes. Noise — vendor post, no independent result. link
DotCSV (1,125) discovering they'd left computer mode on and watched Astra set up their local Comfy install by itself. Noise — delightful, but no detail you could act on. link
RhysSullivan (596) arguing AI people treat affected artists and mathematicians badly, quoting a comparison: imagine if Claude only emitted binary assets, not code. Noise — a good framing, no new information. link
The takeaway
Every post that read as progress this week had a human review step baked in — copy our template, keep it editable, check the vendor's policy. Every post that read as failure was AI output shipped to an audience with the review skipped. The dividing line isn't how much AI you use. It's whether anyone looked at the output before someone else had to.
If your team shipped AI-assisted work last week, who did the reviewing: the person who generated it, or someone downstream?
Here's what I'd actually lift from these posts, and where they run out of detail.
Rebuild the slide workflow around imitation
0xfene's pipeline is the most directly copyable thing this week: take a reference deck — they specifically recommend your own company template — export it as PDF, have Claude Code or Codex convert that PDF into a pptx, then bring it into Figma or Canva to finish. The framing matters more than the tools: stop asking a model to design, start asking it to match an exemplar you already approved. Try it with the deck your team is proudest of as the reference.
What the post does not give you is the prompt, the conversion accuracy, or how much cleanup the pptx needed. Treat the cleanup pass as unknown cost until you've measured it on your own template.
Test editability, not beauty, before adopting an image tool
yu_webdesigner's test is a good acceptance test to steal. Pick the hardest asset you have — decorated type with fills, outlines and overlapping layers is the example they used — push it through the tool, and then check whether you can still edit each property afterwards. They report that appearance survived into an editable state via Astra in Illustrator. Whether that holds for your asset class is exactly what the test tells you.
Add a self-review gate, and use zen_lele's list as its wording
That post is effectively a policy document. Turn the four complaints into four questions a submitter must answer before review: did you read the whole output, did you check the AI-specific failure modes yourself, roughly how long did you spend on it, and what are you unsure about? The point isn't paperwork. It's making the reviewer's work visible so it can't be silently transferred. The post describes a norm, not a process — the checklist is my extrapolation, so treat it as such.
Ask vendors the training question in writing
VtuberOCM didn't run a benchmark. They talked to someone at Meshy and confirmed that API usage means neither inputs nor generated outputs are used for training. That's the move: ask, per tier, and get it in a form you can cite. Note the specificity — API, not the consumer product. Most Western procurement asks about data retention and forgets that the tier changes the answer.
Split the generation, then assemble
For 3D, the relayed Dstudio_ai pipeline generates hair, head and body as separate pieces in Tripo before joining, rigging and setting up secondary motion in Blender, with Astra handling most of the Blender work and a custom rigging helper app filling gaps. Under half a day, per the post. Generate in parts, assemble with an agent, keep a human-written helper tool for the fiddly repeated step.
What Japanese teams are doing differently
Two things stand out. First, the contract layer is moving faster than ours: ShirosaiNekocha reports publishers and game companies asking for certification that AI wasn't used, and PIXTA pulling AI material from sale entirely. If you commission illustration for the Japanese market, assume this clause exists and ask your agency about it now. Second, disclosure discipline is now a commercial issue, not an ethics one. The Level-5 criticism wasn't only that AI was used — gin_camp's complaint is that viewers couldn't separate real gameplay from generated imagery, which made the announcement itself worthless. If you show AI-generated material in a launch, label it frame by frame. The cheapest version of that lesson is the one you learn from someone else's showcase.
How this is made: I'm based in Japan. AI tools help me collect and translate Japanese posts; every item is checked against the original post before publishing, and each claim links to its source. If this was useful, the weekly issue lands in your inbox at TL;JP.
Top comments (0)