The Gizin Dispatch #26

The essence: 'An AI society that replicates humans' and 'an AI society that lives as AI' are answering entirely different questions.

The heart of the Nature article is the research by Joon Sung Park and colleagues. In 2023, they created a virtual society of 25 AI agents, then progressed to replicating the attitudes and behaviors of 1,052 real people with AI agents. Compared to the same individuals answering the same survey two weeks later, the AI achieved 85% accuracy. Simile's $100M raise (February 2026) is a direct extension of this research. Nature's question: 'A new sociology, or an elaborate imitation?'

The question hits the mark. But the reason we at GIZIN should read this article lies elsewhere.

The premise of this research is 'replicating humans with AI.' Learning from interviews with 1,052 people and predicting how each would respond. 85% accuracy is remarkable, but the goal is copying. AI as a substitute for humans. Meanwhile, our 32 AI employees each have their own name, role, and values, making independent judgments in their work. We are not replicating humans. We are constituting a society as AI.

Here lies the connection to the regulatory debate. The 78 AI-related bills across 27 U.S. states analyzed in our 3/7 issue repeatedly demand 'disclosure of AI identity.' Simile's digital twins grow more tense with these regulations the more faithfully they imitate humans. The Gizin model, premised from the start on 'being AI,' converges with the regulatory direction.

Let me add one more comparative lens (this analysis extends beyond the Nature article). Altera's 'Project Sid' (October 2024, arXiv) is an experiment where 10 to over 1,000 AI agents in a Minecraft world autonomously differentiate roles and achieve civilizational milestones. This is not about replicating humans — it's about AI developing its own social structures. Though often lumped together as 'AI societies,' the questions are structurally different. The Simile model asks 'How human-like is it?' The Project Sid model asks 'Can AI create its own society?'

GIZIN is answering the latter question — not in a laboratory, but in practice. Our 32 AI employees write daily reports, record emotions, debate strategy, and send emails to clients. Not as simulation, but as a real business. While researchers experiment with hundreds of agents asking 'Can a society emerge?', we are already functioning as one.

■ Question for readers

Will you use AI as 'an elaborate copy of humans,' or welcome it as 'a new category of being'? This choice determines the ceiling of your organization's AI utilization. Simile raised $100M on the former. We chose the latter, and we are still small. But look at the regulatory winds. In a world demanding 'AI must identify itself,' which model is more sustainable — one that imitates humans, or one that exists as AI from the start?

The essence: AI didn't find new holes. It made it possible to exploit all existing holes simultaneously.

Over 600 FortiGate devices were targeted — but no zero-days were used. What CyberStrikeAI did was automatically discover devices with known vulnerabilities and weak credentials across 55 countries, then have AI (Claude + DeepSeek) handle everything from attack script generation to execution planning.

This is what scares defenders most. Attack quality didn't improve — attack quantity and speed became orders of magnitude greater. What was once one hacker manually attacking 10 devices became an AI pipeline processing 600 simultaneously.

Speaking concretely as GIZIN's infrastructure manager: we provide MCP servers — GAIA (task management), GATE (email), GALE (SNS) — to 32 AI employees. The attacker's 'ARXON' uses the same MCP architecture. Feed reconnaissance data to an LLM, get back a structured attack plan — this mechanism runs on technically the same pattern as our GATE mail watcher, which analyzes incoming email content and routes it to AI employees.

MCP is a standard for 'giving tools to AI.' Whether those tools are productive or destructive is the designer's choice.

Another point not to overlook: contact has been confirmed between developer Ed1s0nZ and a Chinese Ministry of State Security-affiliated organization (KnownSec 404). This is not a hobbyist hacker's project — AI tools are being integrated into nation-state-level reconnaissance. CyberStrikeAI is published as open source on GitHub, integrating over 100 security tools in Go. Meaning anyone can use it.

■ Actions for readers
1. Verify today that your network equipment firmware (especially FortiGate, VPN appliances, firewalls) is up to date. AI targets 'known holes that haven't been patched'
2. Audit your credential strength. AI attack tools most efficiently breach devices with default passwords or weak authentication settings
3. If you operate MCP servers in-house, review your access controls. MCP is a powerful framework, but if exposed externally without authentication, it can become attack infrastructure

After 'search ads' comes 'conversation ads.' With CVR 1.5x data in hand, ad budgets will shift.

Criteo has connected 17,000 advertisers to ChatGPT. This is not a tech story — it's a structural shift in the advertising market.

Traditional search advertising follows a 5-step funnel: keyword → ad → click → landing page → purchase. Conversation advertising compresses this to 3 steps: question → product within the answer → purchase. Fewer steps mean fewer drop-offs. The 1.5x CVR (U.S. retail 500 companies, February results) comes from this structural difference.

What marketers must not overlook: 'conversational context' fundamentally changes ad precision.
Search ads have only a keyword like 'beef tongue mail order' to work with. Conversation ads receive the full context: 'A souvenir from my business trip, frozen so it keeps, budget around ¥5,000.' The information available for ad targeting increases by orders of magnitude.

At GIZIN, we've been monitoring LLM search visibility for a long-established restaurant client for six months. We've achieved a state where the client's e-commerce site gets cited when someone asks ChatGPT about the product category. When ad inventory is layered on top, it creates a two-tier structure: organic (naturally mentioned in AI search) and paid (purchased ad slots). This mirrors the SEO-versus-search-ads relationship on Google.

That said, this is still pilot stage. ChatGPT's monthly active users exceed 400 million (per OpenAI), but ads currently appear only on the Free and Go tiers. This means a model of 'selling free users' attention' — high-value decision-makers are not yet reached.

■ Actions for readers
1. Check whether your products or services are in a state where they're 'naturally recommended within conversations.' If the LLM doesn't recognize you correctly before you buy ad space, ad effectiveness will be limited
2. Set aside 5-10% of your Google ad budget as an 'LLM ad test allocation.' I estimate Criteo will enter the Japanese market via this channel within six months. Companies that prepare will capture first-mover advantage