Claude Code: The Discreet Marking That Divides Anthropic

The context, in plain terms

Anthropic, a US company specializing in AI, has integrated discreet marking into its 'Claude Code' tool since April 2026. This mechanism aims to identify Chinese developers via proxies and targeted domains, according to a general press article. It is reported that this mechanism was discovered through reverse engineering and is linked to the fight against model distillation and unauthorized resale of APIs or accounts.

Several uncertainties surround this decision. The exact date of its introduction, initially mentioned as April 2, 2026, by a secondary source, is not corroborated by a major business press source. Similarly, the claim that the mechanism was withdrawn on July 1, 2026, remains unconfirmed by a reference business source. Kapari cannot establish an exact date or confirm that this decision was formally announced by Anthropic, and the precise wording regarding the identification of Chinese developers is also not confirmed by a major business source.

A Marked Front of Hostility

The Kapari test bench engaged a simulated panel of 37 voices to evaluate Anthropic's decision. The range of reactions reveals a clear predominance of hostility: 23 voices openly declare against the initiative, while 9 voices support it. A smaller group of 5 voices expresses doubts, indicating an overall tense reception. This distribution shows that the decision, as formulated, struggles to convince a majority of the simulated stakeholders.

Internal and External Divisions

Several fault lines run through the simulated panel. An important signal is the presence of dissenting voices: an ML security researcher from Anthropic Engineering, although belonging to a largely opposed group, declares in favor of the decision, but with execution doubt. This suggests that it is important to first listen to these profiles to understand the underlying motivations for their support and the perceived operational obstacles, in order to refine the proposal.

Another signal reveals that external voices receive the decision less favorably than internal voices. This indicates the need for distinct messages for internal and external audiences: the decision-maker must prepare specific communication to reassure external stakeholders, while managing internal expectations and concerns. Furthermore, the presence of 'Chinese Developers' (13% of voices) is noisy but without significant weight on this panel, suggesting not to overreact to background noise without real impact on the balance of forces here.

Execution Doubt, a Shared Friction

The dominant friction identified by the test bench is execution doubt. This doubt is particularly notable because it is shared by opposing camps: Anthropic Leadership, which favors the decision, and Anthropic Engineering, which is largely opposed, both raise this same concern. This signal illuminates a potential tipping point: to defuse this friction before any announcement, the decision-maker must proactively address questions of feasibility and operational impact, by presenting concrete guarantees on the implementation of the marking.

Another relevant signal is the stability of the verdict over three independent passes of the engine. This means that the result is not random and that the decision generates a solid and consistent reaction in the simulated panel. This strengthens the validity of the analysis and the need for a serious reevaluation before proceeding.

Hold Off: A Prudent Path Forward

The verdict calculated by the Kapari engine is 'Hold off,' with a 'High' reception risk. This verdict directly stems from the range of reactions, which are mostly hostile, and the identified tensions. The engine recommends reworking the decision before any real-world exposure, due to the strong simulated tension and the perceived high risk. The execution doubt shared between Anthropic Leadership and Anthropic Engineering (dominant friction) is a key signal that justifies this hold off, calling for a revision of the practical aspects and consequences of implementation.

To move forward, the path involves several actions. It is recommended to listen carefully to dissenting voices, such as that of the ML security researcher from Anthropic Engineering, to understand the execution doubts and the motivations behind nuanced support. The disparity in reception between internal and external voices (internal vs. external signal) requires formulating distinct and adapted messages for each audience. Finally, the stability of the verdict (stability signal) confirms the robustness of the analysis and the imperative need to reconsider the decision as a whole, by integrating feedback on the identified friction points.

Questions about this case

Related cases