The news carries the bitter taste of a lesson learned too late. Krafton, the South Korean gaming giant, agreed to pay bonuses to the entire staff of Unknown Worlds Entertainment, the studio behind Subnautica 2, after reaching a legal settlement with the company’s co-founders. The fallout claimed CEO Ted Gill, who announced his resignation as part of the deal. “We mutually agreed to part ways,” he told Bloomberg, closing a chapter that intertwines corporate governance with reckless use of artificial intelligence.

The algorithmic shortcut

Reports indicate that Gill had used ChatGPT to find legal arguments to bypass contractual bonus obligations. The idea, however creative, backfired: it triggered employee and founder pushback, eroded trust in management, and prompted direct intervention from Krafton. Using a public LLM for such a sensitive decision is no footnote. It shows how opaque, unauditable tools can enable ethically dubious choices, with immediate legal and reputational damage.

Public LLMs and governance: opening Pandora’s box

The incident is a symptom, not an outlier. More companies are weaving language models into decision-making without clear rules. ChatGPT and similar services run on third-party cloud infrastructure, where interaction histories, prompts, and sensitive data travel outside the corporate perimeter. That means no data sovereignty, no way to verify reasoning chains, and potential exposure to GDPR violations or litigation. In the Unknown Worlds case, a manager used a public chatbot to justify a move that harmed workers’ rights — a scenario any compliance officer would find alarming.

The on-premise answer: control and transparency

For those evaluating on-premise deployment, this story offers a concrete lesson. A self-hosted LLM, running on corporate hardware under full IT control, keeps data inside the perimeter and enables logging, audit trails, and access policies that make every AI-assisted decision traceable. On-premise solutions allow organizations to integrate open-source models, apply quantization to match available resources, and build documentable inference pipelines, reducing the risk of AI being used as a smokescreen for opaque actions. AI-RADAR examines exactly these trade-offs: from VRAM requirements for local inference to orchestration frameworks like vLLM or Ollama, as well as the TCO of an internally managed infrastructure.

Beyond the case

Ted Gill’s downfall is not just a labor relations story; it’s a wake-up call for anyone outsourcing high-stakes decisions to external models. Algorithmic sovereignty isn’t a luxury — it’s a safeguard against ethical and legal drift. Pipeline transparency and the ability to audit LLM behavior retroactively become essential requirements when reputation and criminal liability are on the line. In an era where generative AI enters the boardroom, the only defense is to bring control back within the boundaries of your own infrastructure, where every token can be traced and every decision stays under the lens of human governance.