Can You Sue for AI Misinformation? Defamation and Liability

This article is not legal advice. It is provided for general information only; the legal assessment of, and the response to, any specific matter must be discussed with a qualified attorney. Case law and legislation in this area move quickly, and the statements here reflect the position as of the source verification date of July 16, 2026.

When AI generates false information about you or your company, there are cases in which you can challenge it legally. Betting on that in practice, however, is risky. The reasons fall into three groups. First, who is responsible is not yet settled. Traditional search engines have escaped liability on the footing that they merely arrange other people's information, and whether that shield, the limitation of liability granted to intermediaries, extends to output in which an AI assembles the sentences itself and asserts them has become a contested question worldwide. Second, jurisdictions diverge. In Germany, in May 2026, a court assessed Google's generative AI summaries (AI Overviews) as Google's own statements and granted a preliminary injunction accepting its liability, although this is not a final judgment and is on appeal. In the United States, by contrast, the evidentiary bar for defamation is high, and in the Walters matter at least, the AI provider prevailed. That said, Wolf River and Starbuck, both against Google, remain pending, and the direction of US law is not settled. Third, it takes time. In one US matter, the threshold fight over which court would hear the case consumed roughly ten months on its own. Litigation should therefore be preserved as a last resort while day-to-day practice gets ahead of the problem through detection, prevention, and measurement.

Heads of legal, executive, and communications functions whose company is being described inaccurately, or unfavorably, inside AI answers, or who want to prepare for that risk. The article assumes a reader who needs to know whether legal action is available and how to position it realistically.

• On May 28, 2026, the Regional Court of Munich I granted a preliminary injunction under which Google may be subject to an administrative fine (Ordnungsgeld) of up to EUR 250,000 per violation if it repeats the false statements produced by its generative AI summaries, and ordered that Google bear 80% of the procedural costs (LG München I, 26 O 869/26). This is a preliminary injunction, not a final judgment, and Google has stated that it will appeal.
• A Minnesota solar installation company is seeking USD 110 million to 210 million in damages over misinformation in Google's AI summaries (LTL LED, LLC v. Google LLC). Both figures are amounts claimed by the plaintiff in a pending matter, not findings by any court.

What "AI misinformation and misattribution" means in the first place, including the definitions of its four types and the argument that AI citation is structurally imperfect, is handled in the hub article on AI misinformation and misattribution. This article is its legal spoke, narrowed to a single question: can you fight it in court?

When people talk about challenging AI misinformation legally, the first thing they trip over is a basic question: whom do you sue? There is a structural difficulty here.

To begin with, you cannot sue the AI itself. An AI is not a subject of legal rights and duties, so the only possible defendants are the company that develops and provides the AI, such as OpenAI or Google, or the original publisher of the information the AI cited incorrectly. In the typical misinformation case, however, those two do not line up. Where the AI states that a given source says something when that source says nothing of the kind, which is to say where the AI has confused its sources or fabricated a citation that does not exist (misattribution), the original publisher cannot be held responsible. The error was produced by the AI's own synthesis.

This is what decisively separates the situation from conventional online defamation. In the conventional pattern, someone posts something unlawful on a forum or a social platform, and that human poster is the primary bearer of responsibility. The platform, having merely provided the venue, was in principle shielded as an intermediary. With AI misinformation, the unlawful statement was not written by a human poster; it was written by the platform's own AI. This is where the doctrinal core discussed below originates.

The question of whether AI misinformation can be challenged legally therefore resolves into a different question. For statements an AI generates, can the provider be shielded in the same way as a traditional intermediary, or does it bear direct liability as the speaker? That is the subject of this article.

Internet service providers have long enjoyed strong limitations of liability, as intermediaries, for information originated by third parties. The framework goes by different names in different countries, but the underlying idea is shared.

• United States: Section 230 of the Communications Decency Act, enacted in 1996. It provides that a provider of an interactive computer service shall not be treated as the publisher or speaker of information provided by another information content provider. This is why Facebook or Google are generally shielded from liability as publisher or speaker for what users post. The shield is not absolute: there are exceptions for intellectual property, federal criminal law, and sex trafficking related claims under FOSTA, among others, and a suit can still be filed.
• European Union: the Digital Services Act (DSA). A hosting provider that does no more than store third-party information is in principle exempt from liability, subject to conditions such as acting expeditiously once it becomes aware of illegality. The traditional exemption for search engines sits in the same lineage.
• Japan: the Information Distribution Platform Act (Jōhō Ryūtsū Purattofōmu Taisho Hō, formerly the Provider Liability Limitation Act; in force since April 1, 2025). It limits, on certain conditions, the damages liability of service providers with respect to information distributed through what the statute calls specified telecommunications.

Each of these was built on the premise that a business is passively transmitting or storing information originated by someone else. The problem is whether generative AI output fits that premise.

Legal commentary has largely converged on a single axis. The analysis by the Congressional Research Service (CRS) places generative AI products on a spectrum running between a search engine, which is more likely to fall within the shield, and a creation engine, which is less likely (CRS Legal Sidebar LSB11097). An analysis by the Center for Democracy & Technology, a US nonprofit, likewise takes the view that where an AI generated the problematic content in whole or in part, courts are unlikely to extend the Section 230 shield. Senator Wyden and former Representative Cox, who drafted Section 230, themselves stated in 2023 that the provision does not protect generative AI output. All of these, however, are policy views or legal analyses, not settled judicial determinations.

The same question has been taken up squarely in the academic literature. "Where's the Liability in Harmful AI Speech?" by Peter Henderson, Tatsunori Hashimoto, and Mark A. Lemley of Stanford University (Journal of Free Speech Law, 2023) analyzes several categories of liability, including defamation, and concludes that whether the Section 230 shield reaches AI output is closely bound up with the technical details of how the algorithm is designed. The authors argue against granting AI a blanket exemption in these settings, while also noting that there are many obstacles to actually imposing liability on models and the parties behind them. In other words, the shield does not apply as a matter of course, but establishing liability is not easy either, and it is this two-sidedness that makes the subject genuinely hard.

It was the Munich matter, examined next, that first put this abstract question to a court as a concrete judicial determination.

All four matters below are ongoing or recent as of the time of writing, and they must be read with strict attention to legal status: judgment, preliminary injunction, pending, settled, on appeal. None of them may be generalized as settled doctrine.

3-1. Regional Court of Munich I: a generative AI summary assessed as Google's own statement (preliminary injunction, on appeal)

The matter: LG München I (Regional Court of Munich I, 26th Civil Chamber), case no. 26 O 869/26, May 28, 2026. The claimants were Verlagshaus24, a Munich publishing house that brings together twelve publishing brands, and its subsidiary, which publishes books and magazines under the GeraMond brand.

What happened: in early 2026, when users combined the publisher's name in Google Search with the German word for a fraud scheme (Betrugsmasche), a phrase Google's own autocomplete was suggesting, the AI summary opened with an affirmative sentence to the effect that yes, the company was known for dishonest business practices, listed the hallmarks of a fraud scheme, and went on to advise consumers on what to do. According to the court's findings, none of these central assertions was supported by any of the sources the AI cited. The AI had confused the claimants with a different company that did have problems, manufacturing a connection that appeared in none of the cited sources. The claimants gave notice on February 2, 2026 through a cease-and-desist letter from counsel and an email the same day, and separately through Google's reporting form, but Google did not respond adequately.

The core of the reasoning: the court assessed the AI summary not as a display of third-party search results but as Google's own independent statement. Three points carried the reasoning. First, unlike a conventional list of links, which arranges results without summarizing them, the AI summary summarizes and presents those results in its own words and its own structure. Second, it contained assertions found in none of the sources. Third, if such output were treated identically to search results, victims of machine-generated falsehoods would be left with no remedy at all. On this assessment, because the AI summary was not the mere storage or display of third-party information but content Google itself had generated, the court held that Google could not straightforwardly invoke the intermediary liability exemption under the DSA, and placed Google in the position of a directly responsible party (unmittelbarer Störer). The court also stated that Google is able to verify the content of an AI summary against the sources it rests on without contacting any third party, rejecting Google's argument that users should establish the truth for themselves.

Status (read carefully): this is a preliminary injunction (einstweilige Verfügung), not a final judgment. It is a first-instance determination, and Google has already stated that it will appeal (Berufung). An administrative fine of up to EUR 250,000 per violation may be imposed, and Google bears 80% of the procedural costs. The determination rests on German law, specifically the framework of personality rights and injunctive claims, and does not extend automatically to other countries.

3-2. Wolf River Electric v. Google: misattribution and secondary spread (United States, pending)

The matter: LTL LED, LLC (trading as Wolf River Electric) v. Google LLC. A Minnesota solar installation company filed a defamation suit in Ramsey County District Court, a state court, on March 11, 2025. Google removed the case to the District of Minnesota, a federal court, where it was assigned to Judge Jeffrey Bryan (federal case no. 25-cv-02394).

What happened: Google's AI summary stated that the company was being sued by Minnesota Attorney General Keith Ellison. The Attorney General had in fact sued four solar loan companies (GoodLeap, Sunlight Financial, Solar Mosaic, and Dividend Solar Finance), but Wolf River Electric was not among the defendants. The AI summary cited four sources, including a local newspaper article, a review site, and an announcement from the Attorney General, none of which said the company had been sued. This is a textbook case of misattribution: presenting an assertion the actual sources do not support as though it rested on them.

Damages and secondary spread: the plaintiff alleges that a contract worth USD 39,680 was cancelled on March 3, 2025, one worth USD 150,000 on March 5, and a USD 174,044 project with a nonprofit on March 11. It estimates its 2024 damages at USD 24.7 million and claims USD 110 million to 210 million. What matters is that the misinformation did not stay inside the AI answer. Competitors raised the falsehood in sales conversations to steer customers away, search autocomplete began suggesting "Wolf River Electric lawsuit Minnesota Settlement," and posts disparaging the company appeared on Reddit. Misinformation spreads outward from the AI, and the range of potential legal claims widens with it. That point is taken up in section 6 below.

Status (read carefully): this matter is pending and remains at the stage of the plaintiff's allegations. Google is reported to be preparing to assert the Section 230 shield, and commentators, including Dean McGeveran of the University of Minnesota Law School, described the plaintiff as facing an uphill battle. On January 9, 2026, however, Judge Bryan determined that the case should be remanded to state court on the ground that Google's removal to federal court had been untimely; the order was reported on January 12. In other words, the threshold fight over which court would hear the case took roughly ten months from filing on its own. The merits are still ahead.

3-3. Walters v. OpenAI: the high bar of defamation (United States, summary judgment for OpenAI, individual case)

The matter: Walters v. OpenAI, L.L.C., case no. 23-A-04860-2, Superior Court of Gwinnett County, Georgia. On May 19, 2025, the court granted OpenAI's motion for summary judgment. The plaintiff was not a company but Mark Walters, an individual who hosts a nationally syndicated radio program.

What happened: while a journalist was having ChatGPT summarize the complaint in a separate, real lawsuit, ChatGPT produced a false summary stating that Walters had been accused of embezzlement. Walters sued for defamation.

The core of the reasoning: the court ruled for OpenAI on three independent grounds. First, no defamatory meaning. Applying a reasonable reader standard, the court gave weight to the several cautions present in that exchange, including ChatGPT's references to being unable to access the link and to its knowledge cutoff, and OpenAI's communication through its terms and elsewhere that output may be inaccurate, and held that a reasonable reader in that position could not have taken the output as a statement of actual fact. The journalist himself confirmed within a short time that the content was not true. Second, no negligence or actual malice. The court held that the showing of negligence or actual malice required in this matter had not been made; OpenAI works to suppress errors, and knowing that errors can occur does not amount to malice. Third, no damages. Walters testified that he had suffered none.

Status and implications (read carefully): this is an individual case, and its premises differ from those of reputational harm to a company. What it shows is that US defamation doctrine, and in particular the requirement that a reasonable reader take the statement as fact together with the wall of proving negligence or actual malice, can be a powerful defense for an AI provider. As the court itself noted, this is one court's determination on one set of facts.

3-4. Starbuck v. Meta and Starbuck v. Google: a settlement and a pending matter

The matter: in April 2025, the conservative activist Robby Starbuck sued Meta, alleging that the company's AI had falsely stated that he was involved in the attack on the United States Capitol of January 6, 2021. That matter was settled in August 2025, and Starbuck was reported to have taken on an advisory role regarding the company's AI, with the terms of the settlement undisclosed. On October 22, 2025, Starbuck then sued Google in the Superior Court of Delaware, alleging that Google's AI products (Bard, Gemini, and Gemma) had repeatedly described him as a child sexual abuser and a serial sexual assailant while citing fabricated sources, and seeking at least USD 15 million. According to the complaint, Google's AI stated that it had delivered false and defamatory information to approximately 2,843,917 unique users. All of these are the allegations of a plaintiff in a pending matter, not facts found by any court.

Status (read carefully): the Meta matter is settled, which is not a judicial determination. The Google matter is pending, and Google has filed a motion to dismiss on three grounds: that because a user's query is the trigger, the AI cannot be said to have published the statement; that the plaintiff has not identified any specific person who saw and believed the output; and that the tools were experimental and expressly flagged the possibility of inaccuracy. Note that in both matters the plaintiff is an individual.

On the same underlying subject, AI misinformation, Germany produced a preliminary injunction favorable to the claimant, while in the United States there is at least one matter (Walters) in which the AI provider prevailed, and the corporate suits against Google remain pending with no settled direction. The divergence is not accidental; it follows from structural differences between the two legal systems.

The United States: a double wall. In the United States, a claim against a platform runs into two barriers. One is the Section 230 intermediary shield; the other is the stringency of defamation doctrine derived from the First Amendment. Even if a court were to hold that Section 230 does not reach AI-generated content, the plaintiff must still prove defamation itself. A public figure must show actual malice; a private figure must show at least ordinary negligence; and the plaintiff must also establish that a reasonable reader took the output as a statement of actual fact. As the Walters matter shows, so long as an AI provider warns that output may be inaccurate and works to suppress errors, these elements tend to work in the provider's favor. Damages must be proven as well. The result is that plaintiffs in the United States face an uphill battle.

Germany: personality rights and injunctive claims. Germany, and much of the civil law world, protects personality rights and commercial reputation more robustly, and the central remedy is the claim for injunctive relief (Unterlassungsanspruch). An injunction does not necessarily require proof of a damages figure or a heavy subjective element; it prohibits repetition of the unlawful statement going forward. The framework the Regional Court of Munich I applied was precisely this one, which narrowed the dispute almost entirely to a single question: is the AI summary Google's own statement, or merely an intermediation? Because the court answered that it was the former, it held that the AI summary could not be characterized as the mere storage or display of third-party information and that Google could not straightforwardly invoke the intermediary liability exemption under the DSA.

The shared watershed. Even so, the question at the root of both systems is one and the same. Does the traditional intermediary shield extend to statements the AI itself generated? Munich answered no, clearly. US courts have not yet given a final answer, but commentary, from the drafters of Section 230 to the CRS to Henderson, Hashimoto, and Lemley, leans toward the view that AI output is something the AI itself produced and that the shield does not apply as a matter of course. In the United States, however, the independent wall of defamation doctrine still stands beyond it. Losing the shield and being held liable are two different things.

What follows concerns Japanese law, and one caveat has to be stated first and stated strongly. Japan has no established case law on defamation arising from AI misinformation. What follows is an organization of the existing statutory framework; it does not support any conclusion to the effect that a claim brought in Japan would succeed. The assessment of any specific matter requires separate consideration and must be discussed with a qualified attorney.

There is no special statute defining who is liable or on what elements. Japan currently has no special statute that directly establishes the responsible party or the elements of liability for defamation or damage to credit caused by AI misinformation. Where a problem arises, the response has to be assembled from tort law (Article 709 of the Civil Code), contract law, and in some cases the Information Distribution Platform Act. For defamation, the criminal offenses of defamation (Article 230 of the Penal Code) and insult (Article 231), and in the case of a company damage to credit (Article 233), together with damages in tort, are the bases one would expect to be invoked.

Primary responsibility tends to point to the person who used and published the output. The mainstream view in Japan is that the party who actually uses and publishes AI output bears primary responsibility. Someone who reposts AI output verbatim to social media or the web in a way that disparages another person can be held liable for defamation whether or not AI was involved. In this respect nothing has changed from conventional disparagement.

The provider's liability is not settled. Where the AI's own synthesis produced the error, by contrast, whether the provider, meaning the developing company, can be held responsible is not necessarily settled under current law. One question is whether an AI chatbot's answer constitutes the distribution of information through specified telecommunications within the meaning of the Information Distribution Platform Act. The assessment may differ depending on the form of the service. An answer displayed only to the user who asked, and generated content published on the open web, may occupy different legal positions; the former does not sit comfortably with the statute's premise of distribution directed at an unspecified audience. It is also said that joint tort liability may arise where a provider intentionally facilitates infringement, but this too depends on the facts.

The direction of policy. The Information Distribution Platform Act, formerly the Provider Liability Limitation Act, which came into force on April 1, 2025, strengthened the framework for victim relief relative to its predecessor. It obliges large platform operators to speed up their handling of removal requests, including notifying a requester of the outcome within seven days in principle, and to make their operations more transparent, and it requires them to formulate and publish criteria for implementing measures to prevent transmission, that is, removal criteria (Article 26 of the Act). That framework, however, is aimed primarily at the removal of third-party posts and the disclosure of sender information; it was not built with statements an AI itself generates squarely in view. The government's basic posture is to rely for now on voluntary industry guidelines while considering legislation if a serious situation arises.

Where the victim is a company, damage to credit or injury to commercial reputation may be at issue, but here too there is no established case law on how to construct that as the provider's liability. In short: in Japan, the path to challenging AI misinformation legally is not closed, but there is no established body of case law to rely on, and which theory supports which claim against whom requires careful, case-by-case examination. That is the accurate description of where things stand.

What deepens the difficulty of handling AI misinformation legally is secondary spread. Misinformation does not sit still inside the AI answer that first generated it. As the Wolf River matter shows vividly, a falsehood an AI states surfaces in search autocomplete, is copied onto forums, is quoted in a competitor's sales pitch, and is absorbed into the training data of other systems.

From the standpoint of legal remedies, this produces two consequences. First, the set of potential defendants disperses. Even if you began with the AI provider in mind, each point of secondary spread brings its own responsible party into view: the person who reposted, the operator of the forum, and so on. Second, removing one output does not end the harm. Even if the provider corrects or deletes the answer in question, no judgment or injunction can sweep up the copies that have already spread, the memories of the people who saw them, or the traces absorbed into other services. The Munich injunction prohibits future repetition; it does not roll back what has already spread.

Litigation, in other words, can achieve the result that this provider will not repeat this statement in future, but that is a different thing from managing the whole picture of misinformation in motion. That recognition leads into the practical design in the next section.

The shape of legal action emerges from everything above. There are cases in which you can fight. But it is slow, it is expensive, and the outcome is unpredictable.

• Slow: in the Wolf River matter, the jurisdictional fight alone, before reaching the merits, took roughly ten months. Judicial procedure cannot keep pace with the speed at which misinformation spreads in real time.
• Unpredictable: Munich favored the claimant but is a preliminary injunction on appeal; Walters went to the provider; Wolf River is pending; Starbuck against Google is pending with a motion to dismiss filed. Outcomes swing widely with jurisdiction and facts, and no settled rule has emerged.
• Not erasable: as section 6 explains, a win or an injunction stops future repetition but cannot recover the secondary spread that has already occurred.

None of this makes legal action pointless. Injunctions and damages should be preserved as the last resort for egregious and serious cases. As the Munich matter shows, there are certainly situations in which a cease-and-desist letter from counsel, or an injunction, does its work. The point is the ordering: keep litigation in reserve as the final card, and let day-to-day practice get ahead of the problem through detection, prevention, and measurement.

Concretely, think in three layers. Detection, meaning monitoring on a continuing basis how AI writes about your company. Prevention, meaning preparing the points most likely to be misunderstood as accurate, primary information of your own. And measurement. Misinformation cannot be captured in a one-off screenshot: cited sources rotate from day to day, and the same question yields answers that vary. That is exactly why you need a layer that works across multiple AI engines and measures continuously, not only whether you were cited but how you were described, meaning whether it is accurate, whether the comparison is unfavorable, and whether there is misattribution. Where an official tool such as the generative AI performance report in Google Search Console shows only impressions on Google's own surface, and shows nothing about how you were described or whether you were misattributed, what is at stake here is instead the idea of managing perception in AI space itself, which is the domain covered in AI Perception Management (AIPM). Precisely because litigation is inferior to getting ahead of the problem, continuous measurement becomes the foundation of practice.

The reporting and removal routes each provider offers once you have found misinformation, and the question of whether it can be erased technically, are handled in a companion article on the correction and removal of AI misinformation (in preparation). This article keeps its weight on courts and legal claims.

If legal action is in view, the quality of the first response shapes the options that remain later. What follows is limited to the legal first response; the mechanics of each provider's reporting form belong to the companion article noted above.

1. Preserve the evidence. The moment you find the answer in question, take a screenshot and record the date and time, the query you entered, and the URLs displayed. AI answers change constantly, and you may not be able to reproduce the same screen later. That the Munich claimants could submit emails with the AI summary attached, and printouts of the search results page, as evidence became the foundation of their proof.
2. Check reproducibility. Confirm whether the same or a similar query brings the erroneous output back repeatedly. If it reproduces, it is easier to treat as a continuing display rather than an accidental one-off error. In Munich, the fact that similar output reproduced on a fresh search after the warning letter carried weight.
3. Identify what is false, and in relation to which source. Check which sources the AI cited and whether those sources actually support the assertion in question. If they do not, it is misattribution, which means the cause lies in the AI's synthesis rather than in the original publisher.
4. Keep a record of your notice. After consulting counsel, give notice where appropriate through a cease-and-desist letter, sent by certified content mail or an equivalent method. Notice fixes the point at which the provider became aware of the illegality and can ground liability if it then fails to act. In Munich, the letter from counsel and Google's failure to respond adequately to it were important elements of the determination.
5. Have criteria for when to involve counsel. Weigh the scale of the harm, meaning whether there are concrete losses such as cancelled contracts or suspended business, together with the egregiousness, the reproducibility, and the jurisdiction, meaning where the parties are located and which country's law may apply. Then work with a specialist to judge whether legal measures including injunctive relief and damages are realistic. An injunction is a legal remedy worth considering as a way to stop future repetition.

Every one of these is a useful first response whether or not you proceed to litigation. Preserving evidence and recording reproducibility are continuous with the ongoing measurement described above; they are the starting point for both the legal and the operational response.

Q1. AI has produced false information about my company. Can I sue?

In some cases, yes. But the defendant is not the AI itself; it is either the company that provides the AI or the original publisher of the information the AI cited. Where the misinformation was born in the AI's own synthesis (nothing in the sources supports it), you cannot claim against the original publisher, so the question becomes the provider's liability, and that framework is being contested country by country. Germany has a preliminary injunction accepting a provider's liability (not a final judgment, and on appeal), while in the United States the provider prevailed in the Walters matter at least, and matters brought against Google remain pending. Japan has no established case law, so whether a claim is viable is a case-by-case judgment. Consult a qualified attorney first.

Q2. Who ends up being responsible?

The question at the center is whether, for statements an AI generates, the provider is shielded like a traditional intermediary or bears liability as the speaker. The Regional Court of Munich I assessed Google's AI summary as Google's own statement and accepted its liability. In the United States, whether the Section 230 shield reaches AI output turns on whether the output is information provided by another party or content the AI itself created, and no final judicial determination has been issued. In Japan as well, the treatment of an AI provider's liability under current law is not settled.

Q3. Why is liability treated differently for a traditional search engine and an AI summary?

Traditional search was regarded as an intermediary that merely arranges links to other people's pages, and it escaped liability on that footing. An AI summary, by contrast, summarizes several sources in its own words and its own structure and presents them as a single conclusion. The Regional Court of Munich I seized on that act of asserting something in its own words and held that this is no longer a display of third-party information but Google's own expression. What the user experiences is not a list of references but an answer, and that difference is changing the legal assessment.

Q4. Could I win if I sued in Japan?

It is not possible to assert that you would win. Japan still has no established case law on defamation arising from AI misinformation, and which legal theory supports which claim against whom requires careful case-by-case examination. Criminal defamation (Article 230 of the Penal Code), tort (Article 709 of the Civil Code), and damage to credit are the bases one would expect to be invoked, but the treatment of an AI provider's liability is not settled. Determinations abroad, such as Munich, rest on German law and do not extend automatically to Japan. Always consult a qualified attorney.

Q5. How much time and money does litigation take?

As a general matter, it takes considerable time and money. In the US Wolf River matter, the threshold fight over which court would hear the case took roughly ten months from filing, before the merits were even reached. Judicial procedure is inherently slow relative to the speed at which misinformation spreads in real time, and outcomes swing with jurisdiction and facts. Specific costs and prospects vary widely with the content of the matter and the forum, so they need to be confirmed with an attorney.

Q6. How should I preserve evidence?

The moment you find the answer in question, take a screenshot and record the date and time, the query you entered, and the URLs displayed. AI answers change easily, and you may not be able to reproduce the same screen later. It is also useful to check and record whether the same query brings the erroneous output back repeatedly, which is to say its reproducibility. In the Munich matter, records with the AI summary attached, and the fact that similar output reproduced on a fresh search after the warning letter, became the foundation of the proof.

Q7. What is the difference between a preliminary injunction and a judgment? Which was Munich?

The Munich matter is a preliminary injunction, an interim prohibition, and not a final judgment on the merits. It is a first-instance determination, and Google has already stated that it will appeal. A preliminary injunction has the effect of provisionally stopping future repetition, but it may still be overturned on appeal. This determination therefore cannot be generalized as settled precedent. The accurate way to understand it is as an instance in which such a claim was accepted by a court.

Q8. If a statement is attributed to a source that does not exist, how does that differ from ordinary defamation?

Where an AI states that source A says B when A contains no such statement, we call that misattribution. In ordinary defamation, you can identify the person who wrote the falsehood and hold them responsible; with misattribution, because the error was produced by the AI's own synthesis, you cannot claim against the original publisher. The arrow of responsibility has nowhere to point except at the AI provider, and that is what differs from the conventional pattern. The Wolf River and Starbuck (against Google) matters both contain this element of misattribution.

Q9. Do the Walters and Starbuck matters, where the plaintiffs are individuals, apply to companies?

They cannot be applied as they stand. In both Walters and Starbuck the plaintiff is an individual, and the premises differ from reputational harm to a company. In the United States in particular, a public figure faces the extremely high evidentiary hurdle of actual malice. A corporate matter such as Wolf River instead runs into different walls: the Section 230 shield, and proof of damages and negligence. Care is needed not to generalize the outcome of an individual case mechanically to a company's situation.

Q10. Can I seek an injunction?

An injunction is a legal remedy that prohibits repetition of an unlawful statement going forward, and it is an option worth considering. It was also the framework the Regional Court of Munich I applied, German law being structured so that injunctive relief can be sought without necessarily proving a damages figure. In Japan, injunctive claims are likewise one of the general remedies, and a company may raise damage to credit or injury to commercial reputation, but there is no established case law on how to construct that as an AI provider's liability, so it is a case-by-case judgment. Note also that a court-ordered injunction and voluntary removal through a platform's reporting form are separate routes; the latter is covered in a companion article.

Q11. Does the US Section 230 extend to AI?

No final judicial determination has been issued. Section 230 is a shield for information provided by another party, and the contested question is whether content an AI itself generated falls within it. The drafters of Section 230 have themselves said that it does not protect generative AI output, and several legal analyses take the view that the shield is unlikely to reach content an AI generated in whole or in part. These are policy views and legal analyses, however, not settled judicial determinations. And even if the shield falls away, the United States leaves a separate wall in place: proof of defamation itself, including negligence, actual malice, and damages. Losing the shield and being held liable are different questions.

Q12. Apart from litigation, what can a company do?

There is a great deal, and the foundation of practice is in fact there. Because litigation is slow, its outcome is hard to predict, and it cannot erase misinformation that has already spread, the realistic approach day to day is to get ahead of the problem through detection, prevention, and measurement. Monitoring how AI writes about your company across multiple engines on a continuing basis, preparing the points most likely to be misunderstood as accurate primary information, and measuring how you were described as well as whether you were cited: this posture also builds the evidentiary foundation you would need if legal action ever became necessary. See the hub article and the companion article on measurement (in preparation) for details.

AI misinformation can, in some cases, be challenged legally. The preliminary injunction from the Regional Court of Munich I assessed a generative AI summary as the provider's own statement and, for the first time, put into a concrete judicial determination the proposition that the traditional intermediary shield does not apply as a matter of course. But it is not a final judgment and it is on appeal; in the United States the provider prevailed in the Walters matter at least, while the corporate suit against Google is pending with no settled direction; and Japan has no established case law. Responsibility is unsettled, jurisdictions diverge, and procedure is slow.

That is exactly why the ordering of practice is clear. Keep litigation in reserve as the last resort for egregious and serious cases, and get ahead of the problem day to day through detection, prevention, and measurement. The legal first response, preserving evidence and recording reproducibility, is continuous with ongoing measurement, and the two do not conflict. Since no single judgment can stop misinformation that keeps moving, you need a layer that continuously measures and manages how your company is perceived in AI space.

Next action: start by combining your company name with words such as fraud, lawsuit, and reputation, check what the major AI systems actually return, and preserve the evidence immediately if there is a problem. Then determine whether the misinformation is transient or continuing, and consult counsel as needed. For the broader picture of countermeasures, see the hub article; for the thinking behind continuously measuring and managing perception in AI space, see the explainer on AIPM.

Disclaimer (repeated): this article is general information, not legal advice. The legal assessment of, and the response to, any specific matter must be discussed with a qualified attorney.