Anthropic updated its Acceptable Use Policy, effective 12 November 2026.
Among other things, the policy prohibits “sustained and needless abusive or cruel behaviour towards” their models (R129). However, there is no further elaboration in the policy, such as what is considered to be abusive or cruel. While this change gained attention, (see for example: Guardian, Tech Crunch, The Verge) the policy involves numbers of significant rules that have real-life implications.
The new policy prohibits:
- producing, distributing, or facilitating real or AI generated non-consensual intimate imagery (NCII), and tools to make it (R124);
- building or improving surveillance tools and systems in cases of “facial recognition, battlefield management, predictive policing, monitoring in law enforcement investigations or criminal justice processes” (R169, R170);
- identifying individuals from anonymous activity and creating or collecting personal data for surveillance purposes or to cause harm (R172).
- as a universal standard, the use of Anthropic products and services in law enforcement and criminal justice decision processes (R252).
Additionally, Anthropic has expanded the list of cases in which the use of AI is permitted, but subject to further requirements. These requirements are:
- a qualified human reviewer who evaluates the AI output and assumes responsibility for the decision (human-in-the-loop);
- disclosure to individuals affected by the AI-generated output;
- a human controller who monitors the process and can intervene anytime to stop the physical actions during execution.
The “High-Risk Physical Actions” list includes control of safety systems, industrial processes, hazardous energy or materials, and so on (R254–267).
Crucially, the “High-Risk Areas” for AI recommendations list now allows:
- decisions or recommendations on legal status (e.g., immigration, asylum, or citizenship) and judicial processes for civil, administrative, or arbitral matters (R250).
- decisions related to credits, loans, mortgages, credit score production, or creditworthiness assessments (R243).
Author: Ayşe Darakçı