In its work report of March 2026, the Supreme People’s Court stated, in the section on promoting the orderly development of AI, that courts should properly hear AI-related cases and accurately grasp the room for “tolerance” for technological innovation.

The report cited a case: a generative AI service produced an error, but the developer had exercised due care and the plaintiff’s rights were not actually harmed, so the court found no infringement. The report also addressed the other side: conduct that uses AI to infringe others’ lawful rights or disrupt social order will be firmly regulated in accordance with law.

“Tolerance” is not “exemption”

The most commonly misread point is treating “tolerance” as “no liability”.

The reasoning is restrained: the court did not say that AI errors are always excused. It said that where two conditions are met, liability does not attach — first, the developer exercised due care; second, no rights were actually harmed. Tolerance applies to innovation that erred despite diligence, not to harm caused by a failure of diligence.

The Court’s general office further explained that an AI model is not a civil-law subject, has no capacity for civil conduct, cannot independently make legally meaningful commitments, and therefore cannot bear civil liability. Liability continues to attach to people.

What the duty of care covers

Since liability turns on whether due care was exercised, companies must translate “duty of care” into demonstrable engineering actions. Judicial rules on AI disputes are becoming more granular — see our note on the Supreme People’s Court’s 24 rules on AI disputes.

From the existing reasoning, due care covers at least three layers:

First, notice and disclosure. Prominently flagging the uncertainty of generated content is often the key fact establishing that care was exercised.

Second, basic safeguards. Filtering obviously harmful output, restricting high-risk scenarios and fixing known defects.

Third, harm avoidance. Acting promptly on a rights-holder’s notice is the most direct way to contain damage.

Three implications for companies

First, turn due care into a traceable process. Notice wording, filtering strategy, response times and disposal records are all evidence later, and they also support the wider governance posture — see our note on China’s AI governance and development arrangements.

Second, distinguish “service error” from “using AI to infringe”. The former may be tolerated where care was exercised; the latter is firmly regulated. Criminal and civil liability for using AI to infringe others’ copyright or portrait rights is not displaced by “technological neutrality” — see our note on the Beijing High Court typical case on AI-processed artwork and our criminal compliance service.

Third, treat “tolerance” as a pacing reference. Judicial tolerance for innovation does not conflict with regulatory requirements on filing, labelling and assessment. Companies can use it to judge which innovations can be piloted first and which must first complete their compliance baseline — see our data compliance service.

The precise meaning of “room for tolerance” is certainty for those who exercise care. The price of that certainty is making the duty of care real.

This article is based on the Supreme People’s Court’s March 2026 work report and related public explanations. It is for general reference only and does not constitute legal advice.