When a person asks AI about the weather, a formula, or a schedule, an error usually remains an error of information. When they ask about another person, a company, a conflict, or a controversial publication, AI increasingly does something fundamentally different: it doesn't just retell facts, but evaluates the participants.

It may say that a person is "manipulating," "offended," "trying to justify themselves," "acting out of commercial interest," "deserving of more trust," or conversely, that a source is unreliable. Sometimes such conclusions may be reasonable. The problem arises when the model turns an unprovable interpretation into a confident conclusion.

Let's imagine a regular dispute between two people. There are public actions, documents, correspondence, and a long history of relations. An independent observer might carefully say: "Here are the confirmed facts. Here is the first party's version. Here is the second's. The participants' motives are not reliably known." But a generative system tends to build a coherent narrative. If the story involves conflict, loss of status, and sarcasm, it may take the next step: "that means the person is offended." After that, all their arguments are perceived through the psychological framework created by the machine.

This is precisely where AI ceases to be a simple information assistant and becomes a reputational arbiter.

An arbiter—because its response influences who the user will trust. But without an arbitration procedure—because the person being evaluated by the system usually has no normal mechanism to participate in this process. They do not know what exactly the model said to a specific user, do not see the complete evidentiary base, cannot object in advance, and have no guarantee that correcting one error will change future responses to other people.

For an individual, this is a new form of reputational risk. Previously, a controversial opinion had an understandable author: a journalist, blogger, competitor, or commentator. One could see who was speaking, what their interests were, and what the conclusion was based on. AI looks different. The user often perceives it as a neutral synthesizer of multiple sources, even though part of the final judgment may have been generated by the model itself.

The law is already beginning to grapple with this transition.

In May 2026, a court in Munich ruled that Google can be held directly liable for incorrect statements generated by AI Overview. The key meaning of the ruling is not just about a specific error: the court distinguished between regular link search results and the new text that the system itself generated from sources.