CyberAgent Inc. started piloting a generative-AI review stack for advertising creatives tied to its manga and anime apps, running the system from a Shibuya lab before submissions reach Apple Inc. and Google LLC storefront reviewers. The company said the tool compares draft ad text and storyboard frames against Japan’s Act against Unjustifiable Premiums and Misleading Representations, focusing on probability disclosures for gacha-style draws.
What the pilot scans
Marketers upload video storyboards and localized copy into a private instance of CyberAgent’s ABEMA LLM stack. The model highlights sentences that imply guaranteed rare pulls or omit paid-currency costs, then suggests plainer Japanese alternatives. Human compliance staff must approve every export; the AI cannot auto-publish.
CyberAgent declined to share false-positive rates but said internal tests on 400 archived ads caught 27 instances where odds footnotes were smaller than platform guidelines allow. None of the flagged ads had been rejected by Apple in production; the point was to catch issues before the seven-day review clock starts.
Why manga apps face scrutiny
Japan’s Consumer Affairs Agency has warned mobile game publishers about “komagome” tactics—compressed disclaimers that players cannot read on phone screens. Apple tightened App Store rules for loot boxes in 2025, and Google now requires Japanese-language odds tables for games targeting minors.
CyberAgent’s portfolio includes titles with cross-media anime tie-ins where marketing teams rotate creatives weekly. Manual legal review became a bottleneck during collaboration seasons; the pilot is meant to shrink review cycles from four days to two without weakening sign-off.
Limits and competition
Rivals such as Mixi Inc. and DeNA Co. already use rule-based linters for ad text; CyberAgent is betting fine-tuned language models catch implied claims regex misses. Privacy officers required that training data exclude live user chat logs; only past ad submissions and published odds tables were used.
If the pilot succeeds, CyberAgent plans to offer the reviewer as a paid service to smaller publishers that cannot staff full-time compliance teams—a revenue line separate from ad network commissions.
Platform enforcement
Apple rejected 14 Japanese game ads in the third quarter for missing odds tables, according to developer forum tallies CyberAgent monitors. Google’s Play Console now auto-rejects screenshots that hide gacha percentages behind character art; CyberAgent’s model is trained to flag similar compositions.
Human reviewers stay central
Compliance headcount will not shrink during the pilot; AI suggestions appear as tracked comments in the workflow tool legal teams already use. CyberAgent said it will publish aggregate accuracy metrics after 90 days if false positives stay below five percent.
Broader AI governance
Japan’s AI guidelines for enterprises encourage documentation of training data sources; CyberAgent filed a voluntary report with the Ministry of Internal Affairs describing the ad corpus. Personal data from ABEMA streaming accounts was excluded by policy.
Developer feedback
Mid-size publishers testing the pilot said the model correctly flagged a costume gacha ad that implied a guaranteed legendary pull, a common rejection reason on Google Play. False positives included a historical drama ad where the word “must-win” referred to a plot arc, not odds.
Cost structure
CyberAgent budgets one GPU cluster in Shibuya for the pilot; if expanded, fees would be passed to clients per campaign review, priced below hiring an external law firm retainer.