AI tools conduct hundreds of millions of conversations every day with users on a variety of topics, and while the companies responsible for these tools – whether OpenAI, Google, or Anthropic – claim that their tools rely on online sources, sometimes they may provide false information to users.
However small or very small this percentage of error may be, the volume of conversations conducted with AI tools makes the actual number of conversations containing errors larger than we expect.
Read also
list of 2 itemsend of list
These errors vary greatly, as some tools hallucinate and mix facts and data and transmit them to the user incorrectly, while other tools invent completely new facts that have no relation to reality, so that the information becomes completely false.
The matter is made worse by the fact that users have great confidence in the answers that come to them from artificial intelligence tools, as a YouGov statistics report reveals that 69% of artificial intelligence users in the United States trust its answers.
This confidence raises a pivotal question about the legal responsibility for incorrect answers provided by artificial intelligence, especially with the widespread use of various tools and the increasing rate of adoption and ease of access.
A small percentage with a big impact
A report published by the New York Times reveals the extent of errors made by Google’s artificial intelligence search summaries feature, as the feature provides one wrong answer out of every 10 answers, and this puts the tool’s accuracy at more than 90%.
However, even the correct answers that appeared in the artificial intelligence summaries feature were not based on clear and explicit sources that the user can refer to, according to the New York Times report, as the percentage of correct answers that were not based on clear sources reached 56% in February with the launch of the new “Gemini 3” model.
Despite this, the impact of this small percentage, which does not exceed 10%, may be very large thanks to the volume of usage operations and conversations that users engage in with Google, as users conduct more than 5 trillion searches annually in the tool, according to the report, which makes the volume of daily errors exceed hundreds of millions.
This also applies to other artificial intelligence tools, as a tool such as “GBT Chat” conducts more than 2.5 billion requests daily with its users, who exceeded 900 million users weekly, according to a report by the American “Omni Bond” artificial intelligence company.
The latest GPT5 model has an accuracy rate of up to 87%, according to the latest study compiled by the American website ChatBiz for testing artificial intelligence models, which also makes its daily errors exceed hundreds of millions.
Overwhelming confidence in artificial intelligence
Although a large portion of users have been exposed to artificial intelligence hallucinations and were able to discover them, according to a study published by Exploding Topics, a digital marketing company, the majority of users still trust artificial intelligence and its answers.
The study confirms that only 18.6% of users doubt the results of artificial intelligence and its answers and search for additional sources, even though 42% of users had previously received wrong answers and hallucinations from artificial intelligence.
The study suggests that the reason for this confidence is due to the ease of using artificial intelligence and accessing the information it provides, as the tools usually provide the user with the answer he is looking for in a form that is easy to understand and read.
Who is behind the AI answers?
Global technology companies deal with the artificial intelligence tools they develop in a way that mimics Google’s treatment of its search engine when it first launched it, as companies see their tools as summarizing and republishing what was found in various sources online.
This point of view has emerged in several legal cases in which technology giants have been accused of incorrect answers to their models, including “Google” in the recent case in the German Munich court, and “Open AI” in the case brought against them by radio broadcaster Mark Walters in 2025 and which is still in circulation until now.
But the crisis lies in the fact that American and international laws do not have clear frameworks for dealing with such types of cases that may fill courtrooms in the coming years with the widespread use of artificial intelligence technologies, according to a report by the American newspaper “Bloomberg Law.”
The report attributes the crisis in such cases to four main pillars, including identifying the speaker, because with artificial intelligence you cannot specifically determine who is speaking, whether it is the company or not.
Determining the speaker behind the artificial intelligence answers depends on the court’s point of view, whether it will view the artificial intelligence tools as a special product belonging to the companies that own them, or whether these companies are merely a “platform” on which the answers are published.
In the first case, this does not protect it from legal liability and will put it under penalty of law as it is in fact the speaker, no matter how much these companies try to add warnings to users or warnings that the existing information may be wrong.
But if the court considers artificial intelligence tools to be platforms that indirectly publish others’ answers, this places them under the umbrella of communications ethics laws and does not hold them responsible at all.
Disparity in the courts
Recent cases against artificial intelligence tools reveal a clear difference in the position of different courts.
While the Munich court held Google responsible after its artificial intelligence falsely accused several publishers, the Georgia court held that GBT Chat was not responsible and did not intend to insult radio broadcaster Mark Walters when he accused him of embezzlement.
The Munich court relied on the fact that Google’s artificial intelligence summaries feature provided new and original content to it, and therefore it was responsible for the things mentioned in this content.
But the Georgia court found that GPT Chat lacked intent and chose a defamatory meaning, so what it did could not be considered a case of defamation.
It is expected that the size of this discrepancy will increase between different courts with the diversity of cases and their different details, as well as the difference in each judge’s point of view and how he looks at artificial intelligence technologies.
Does the model question itself?
The legal and liability issues arising from AI hallucinations point to another question, which is whether AI itself can be held responsible for its wrong answers.
A Macedonian academic paper presented at the AI Smart 2025 conference examines this same question, under the title “Legal Liability for Artificial Intelligence?” By researchers Raškowski and Veronika Raškowska.
The study concludes that artificial intelligence is currently not suitable to be legally accused, because the law does not recognize it as a person, in addition to its lack of three pillars on which responsibility is based, which are intent, awareness, and financial liability.
Despite this, the paper warns that artificial intelligence models will become a scapegoat for companies to hide behind to escape the legal responsibility they may face due to what their tools produce.
The importance of this question increases after the legislative steps that several countries around the world are considering taking, starting with Estonia, which decided to grant artificial intelligence agents a personal identity.
