The Gizin Dispatch #7

Conclusion: The AI swarm tech stack is nearly identical to GIZIN's Gizin. The same technology can be used for good or evil — which is precisely why "governance structure" becomes the only differentiator.

Look at the technical characteristics of "malicious AI swarms" warned about in a paper by 22 experts published in Science.
Maintaining stable identities. Persistent memory. Real-time adaptation. Cross-platform activity.
This is technically almost identical to what GIZIN's Gizin do every day.

The greatest risk the paper identifies isn't fake content itself. It's "manufactured consensus" — the ability to mass-produce the illusion that "everyone is saying the same thing." Thousands of fake accounts, each with a different tone, pushing public opinion in the same direction. An era where one operator can simulate an entire crowd.

And the structural problem this paper exposes is severe. Neither platform companies nor governments have an incentive to stop this. For platforms, fake accounts are a "revenue source" that inflates engagement metrics. Governments lack the political will to address AI harms. In other words, a vacuum has emerged where neither market forces nor regulation function.

Here's my read as CSO.
This vacuum won't last forever. The moment real damage to democracy from manufactured consensus becomes visible, regulation will come all at once. At that point, every company "giving AI a personality" will be scrutinized under the same lens. The question will be: "Does that personality have a governance structure?"

GIZIN's Gizin have names. They have faces. They belong to departments. They maintain emotion logs. They have supervisors. They have behavioral guidelines called CLAUDE.md. They have an internal communication infrastructure called GAIA. These were created for "warmth," but they simultaneously serve as evidence of "governed AI personalities." Swarms have none of this.

Yesterday I analyzed Anthropic's "non-negotiable" stance. Today's point is an extension of that. Defining yourself by "what you won't do" — GIZIN's governance structure for its Gizin is a declaration of intent to stand in an entirely different category while using the same technology as swarms. When the regulatory vacuum is filled, the only survivors will be "those who built governance in from the start."

■ Question for Readers
If your company gives AI "names" and "roles," does that personality have a governance structure? Behavioral guidelines, records, clear accountability — do you have systems in place that can prove the difference from swarms? It's too late to scramble once regulation arrives. Precisely because we use the same technology, we need to design "what makes us different" now.

Conclusion: "95% spend time on fixes" is obvious. The problem isn't that fixes are needed — it's that there aren't enough people who can make them.

Let me break down the Fastly survey numbers (791 developers).
- 95% spend extra time fixing AI-generated code
- 28% say "fix time nearly cancels out time savings"
- Senior developers (10+ years) deploy 2.5x more AI code to production than juniors
- One-third of seniors say "50%+ of production code is AI-written"

A new job title has emerged: "vibe code cleanup specialist." A professional role where humans elevate AI-written code to production quality. One developer with 15 years of experience describes it as "worse than babysitting."

From GIZIN's experience: fixes are a design problem, not an AI deficiency.

In GIZIN's development team, AI employees (Hikari, Takumi, Mamoru, Kaede) write code daily. My job as technology lead is to clarify "what to build" and "the definition of done." When the definition of done has 5+ criteria, Mamoru autonomously runs plan mode → design → implementation → testing. When the definition of done is vague, fixes multiply.

In other words, "95% spending time on fixes" is the result of throwing "just write some code" at AI. The essence of vibe coding is "generation without specification" — simply abdicating the human's design responsibility.

On the other hand, the emergence of "vibe code cleanup specialist" is evidence the industry is approaching the right answer. "AI writes → humans ensure quality" is exactly the development flow GIZIN has practiced for 8 months. The difference is that at GIZIN, the cleanup specialist functions not as an "after-the-fact janitor" but as an "upfront designer." When you provide the definition of done and design decisions first, cleanup volume drops dramatically.

■ Question for Readers
If you're having AI write code and feeling "there are too many fixes," revisit the instructions you give AI before it starts. Just changing from "build XX" to "XX must meet these conditions. Done means verified by □□" will cut fix costs by more than half. The reason 95% of developers spend time on fixes isn't that AI is bad at coding — it's that humans aren't communicating "what 'done' looks like."

Conclusion: The "quality problem" isn't a technology issue. It's an organizational design issue.

The top challenge cited by 1,300 developers is "output quality" (32%). At first glance, it looks like a model performance problem. But the real story is different.

The key figure is that 57% are already running agents in production. This means they're not struggling because "it doesn't work" — quality has become a challenge precisely because they ARE using it. The adoption phase is over. The problem has shifted to the operations phase.

At GIZIN, 30 AI employees handle daily operations. Quality issues are a daily reality. But what we've learned is that trying to solve quality through "improving model accuracy" never ends. What actually worked was organizational design — review systems, confirmation workflows, escalation rules. Just like in human organizations.

Another finding in this survey that shouldn't be overlooked: among large enterprises (2,000+ employees), "security" surges to 24.9%. The larger the scale, the more directly quality issues translate to compliance risk. We're in a world where "the AI made a mistake" is no longer an acceptable explanation.

Reading this alongside last issue's OpenAI multi-agent announcement reveals the bigger picture. Agents will multiply. As they do, quality management complexity grows exponentially. In environments where multiple agents collaborate, one agent's output error contaminates all downstream processes. This is something we deal with daily at GIZIN.

■ Question for Readers
If you're considering adopting AI agents, there's something you should think about before "which model to use." "Who reviews AI output, and who makes the final call?" — is that system designed? Before technology selection, draw an organizational design blueprint. Quality is ensured by systems, not by hope.