Grievance as Vector: Inside the China-Japan Influence Operation

The Setup

On February 25, 2026, OpenAI published its latest threat intelligence report documenting how state actors and criminal networks have attempted to exploit ChatGPT. Among the cases disclosed, one entry stands out for what it reveals about operational design rather than just technical abuse.

An individual identified as associated with Chinese law enforcement submitted prompts to ChatGPT requesting assistance planning "a covert IO" targeting Japanese Prime Minister Sanae Takaichi. The context: Takaichi had recently made public statements criticizing human rights conditions in Mongolia. The operational angle being explored was fanning online anger about US tariffs on Japanese goods, with the intent of directing that anger toward the prime minister as the visible face of Japanese policy.

Further reading: National Institute of Mental Health

ChatGPT refused to assist with the operation planning. The account was subsequently banned. But what the attempt reveals about influence operation architecture is worth examining closely.

The Mechanism: Grievance Injection

The tactic being deployed here has a specific structure. It is not disinformation in the conventional sense. No false facts are being invented about Takaichi. No fabricated events are being seeded. Instead, the operator is locating a real, existing grievance in the target population, tariff anger is genuine, widely held, and emotionally charged, and using it as a delivery vehicle to concentrate attention on a specific named target.

This is grievance injection: the selection of a pre-existing emotional charge in a population and the deliberate steering of that charge onto a chosen object. The operator borrows credibility from the authentic grievance. The anger existed before the operation began. The operation simply redirects where it lands.

The strategic logic is sound. Manufactured grievances require infrastructure, maintenance, and time. They also tend to collapse under scrutiny because the foundational claim is false. Authentic grievances carry none of those liabilities. The emotion is already present, the believers already exist, and the underlying complaint has real-world anchors that make it resistant to debunking. What the operator adds is not the grievance itself but the targeting layer: the suggestion, amplified through coordinated accounts, that this particular leader is the appropriate object of blame.

The Evidence: Scale and Exposure

The OpenAI report describes the broader infrastructure within which this specific operation existed. The program counted "at least hundreds of staff, thousands of fake accounts across scores of platforms, the use of locally deployed AI models, and a playbook of dozens of tactics." Those tactics ranged from mass online posting and abusive reporting of dissident social media accounts to forging documents and impersonating US officials.

The exposure itself came through an operational error of a specific kind. The individual being tracked was asking ChatGPT to "edit and polish periodic status reports" on this broader program. This routine administrative task, the kind of clerical work that makes large bureaucratic operations function, handed OpenAI a window into the entire operation's scope and methodology. The operator was not breached. They disclosed.

This matters because it reveals a structural vulnerability in any large-scale covert operation: the larger and more bureaucratic the program, the more administrative overhead it generates, and the more surface area exists for exposure through that overhead. The weak link was not the operational personnel conducting influence activities. It was a staff member polishing internal reports using a commercial AI tool under the assumption that the tool was neutral infrastructure.

The Counter-Read

Two things deserve calibration here. First, the target selection was not arbitrary. Takaichi's Mongolia criticism gave the operator a political context in which criticizing Japan's government could be framed as principled rather than hostile. The operation was designed to look like organic domestic discontent, not foreign interference. The tariff grievance provided that cover: it is a legitimate economic complaint, not a fabricated one, and it routes through entirely Japanese political concerns rather than anything that reads as Chinese in origin.

Second, when ChatGPT refused the initial request, the user continued operating within the same tool. This persistence is consistent with how large bureaucratic programs function. The individual operator had a task to complete. Refusal was a friction point, not a termination signal. The program was designed with enough redundancy that a single tool's refusal would not stop the broader effort.

The Takeaway

Grievance injection is harder to counter than disinformation because the foundational complaint is true. Debunking the anger fails because the anger is legitimate. The intervention has to happen at the targeting layer: demonstrating that the blame assignment is manufactured even when the underlying emotion is not.

For anyone reading influence operations in real time, the diagnostic question is not "is this grievance real?" Most effective influence operations use real grievances. The question is: who benefits from this particular grievance being aimed at this particular target, right now? That framing separates authentic organic anger from engineered concentration of authentic organic anger. The difference is the targeting layer. That is where the operation lives.

---