Who Is Responsible When ChatGPT Gets Your Brand Wrong? Legal and Compliance Considerations for 2026

Key takeaways

• Under current US case law, the odds of holding an AI company liable for a wrong statement about your brand are low. The one ruling we have, Walters v. OpenAI (May 2025), went for OpenAI, though it is a single trial-court decision on unusual facts.
• OpenAI's terms put responsibility for output on the user. Your own marketing and legal claims about AI stay your problem too.
• The thing that changes your position is evidence: provable harm, a documented correction request, and a record of what the model actually said and when.
• ChatGPT's answers about a brand lean on a small number of sources. Fixing those sources is usually faster than any legal route.
• Build a repeatable process: monitor, document, correct at the source, escalate to counsel only when harm is real.

The short answer: nobody has clearly accepted the blame

If ChatGPT tells a buyer your product has a security breach it never had, or that your pricing starts at three times the real number, who pays for the damage? Right now, mostly you do. Not because the law says so in some settled way. The law hasn't settled. But the practical incentives all point at the brand that was misdescribed.

OpenAI's Usage policies say it plainly: "Our rules are no substitute for legal requirements, professional duties, or ethical obligations that should influence how people use AI." The terms of use say output may not always be accurate and should not be treated as a sole source of truth. A lawyer reading that sees a company that has pre-disclaimed the thing you want to sue over.

This guide walks through what courts have done so far, what the terms say, which regulators might care, and what a sensible compliance process looks like for a marketing or legal team. I'm not a lawyer and this isn't legal advice. Treat it as a map for the conversation you'll have with counsel.

What the courts have said so far

Walters v. OpenAI

The closest thing to a ruling is Walters v. OpenAI (Ga. Super. Ct., Gwinnett County, No. 23-A-04860-2), decided May 19, 2025. ChatGPT had falsely told a journalist that radio host Mark Walters was accused of embezzling from the Second Amendment Foundation. The journalist caught the error and never published it. The court granted summary judgment to OpenAI.

The reasoning matters more than the outcome, because a brand's situation differs from his on several points:

• No defamatory meaning. The court held that no reasonable reader would take the output as a literal factual assertion, given ChatGPT's well-known limitations and disclaimers.
• Public figure status. Walters had to show actual malice and had no evidence of it.
• No negligence. OpenAI took reasonable steps to reduce errors and warned users.
• No damages. Walters admitted he suffered no harm. He also never asked for a correction or retraction, which Georgia law requires before punitive damages.

One caution: this is a single trial-court ruling, not binding anywhere else. Don't read it as "AI can't defame." Read it as "this plaintiff had no harm, no correction request, and no malice evidence."

My own inference, not a holding from the case: a company with provable lost sales and a paper trail of correction requests would be arguing from a stronger position. Every element the court leaned on is something a prepared brand can try to fix.

Wolf River Electric v. Google

The more interesting test for brands is Wolf River Electric v. Google. The company sued after a Google AI Overview allegedly fabricated a lawsuit against it, and it claims lost sales and reputational harm. Coverage from The New York Times and a Quinn Emanuel client alert describe the claims. I haven't verified the current procedural status or damages sought, so check the docket before citing numbers. For a broader read on the theory, Haynes Boone has a useful piece, Defamation in the AI Age.

Why it matters: it is a business plaintiff, with claimed commercial damage, against an AI answer surfaced inside a search product. That is much closer to the brand problem than a radio host's reputation.

What OpenAI's own terms put on you

The terms are the first thing a defense lawyer will cite, so it pays to know them.

• The Usage policies prohibit tailored advice that needs a license, such as legal or medical advice, without a licensed professional involved. They also prohibit automating high-stakes legal decisions without human review.
• The Terms of Use, updated January 1, 2026, say output may be inaccurate and shouldn't be relied on as a sole source of truth or a substitute for professional advice.
• OpenAI says it holds people accountable for inappropriate use and offers an appeals process for enforcement decisions. That is about policy enforcement against users, not a channel for brands to dispute what the model says about them.

If you want to read the business-use angle from another firm, Windes puts the 2026 picture simply: ChatGPT can support compliance, but it doesn't make you compliant, and that responsibility stays with your organization.

Two separate problems that get mixed up

People blur these, and the legal answers differ.

This guide is mostly about the first row. But the other three are where regulators actually act, so ignoring them would be a mistake.

Which regulators care

FTC and state consumer protection

The attorneyx.law 2026 startup compliance guide makes the point that the FTC has been clear that overstated AI claims are deceptive practices, and it doesn't need an AI statute to act. State regulators work from the same playbook. If your own content, built with ChatGPT or not, makes claims you can't support, there is no small-company exemption from deception.

The same guide notes that state AI laws apply based on where your users are, not where your office is. Texas's Responsible Artificial Intelligence Governance Act took effect January 1, 2026, with penalties reaching $200,000 per uncurable violation. Illinois amended its Human Rights Act to reach discriminatory AI in employment. Utah requires disclosure when consumers interact with generative AI in regulated contexts. None of these directly give you a claim against ChatGPT for a bad brand answer, but they show where enforcement attention sits.

Advertising rules if you buy ChatGPT placements

If you run ads inside ChatGPT, ordinary advertising law applies. The Adventure Media compliance guide says that ignorance of the rules isn't a defense, and that OpenAI has platform-level policy commitments advertisers must respect. See Promptwatch's ChatGPT ads data for how often ads appear in responses, and the Ads Radar page for what competitors are bidding on.

Privacy and accuracy rights in the EU

I didn't research EU AI Act or GDPR accuracy obligations for this guide, and there is a real gap. GDPR includes an accuracy principle for personal data, and privacy group noyb has complained to regulators about ChatGPT hallucinating facts about individuals. Those rules are about people, not companies, so a brand's route is weaker. If you operate in the EU and the wrong statement concerns a named founder or employee, bring in privacy counsel early.

The evidence problem

The Walters ruling turned partly on missing evidence: no damages, no correction request. That's a process failure you can avoid. Here is what a brand would need.

1. A dated record of the exact prompt, the model, the country and language, and the full response. Screenshots alone are weak. Repeatable monitoring is better, since a one-off answer can be dismissed as a fluke.
2. Proof that the answer repeats. A claim that appears in 40% of responses across weeks is a different story from one weird reply.
3. A link to harm. Lost deals, support tickets quoting the false claim, sales calls where the prospect repeated it.
4. A correction request on file, sent to the AI provider and, where relevant, to the source site that fed the error.

This is also where AI visibility platforms earn their keep outside marketing. Tools like Promptwatch track what ChatGPT, Gemini, Perplexity, Claude, and AI Overviews say about a brand from the real interfaces, in each country and language. Its Brand Narrative feature pulls out the individual claims AI makes about you, such as "its mobile app lacks features", and shows how often and in which models they appear. Fact Check lets you mark a claim inaccurate and then finds the pages on your site that state it or its opposite.

Other monitors worth a look for evidence-style logging include [tool:profound], [tool:peec-ai], [tool:otterly-ai], [tool:athenahq], and [tool:scrunch-ai]. Most are prompt trackers, so check whether they store full responses with dates and locations. For a wider view of the category, the directory at bestgeosoftware.com lists GEO software side by side.

Where wrong answers come from, and why that is your best lever

Legal routes are slow. Fixing the source is faster. Promptwatch Data found that for brand-specific prompts, ChatGPT narrows its sources: the most common count is 5 domains per response, and 40.7% of such responses cite 5 or fewer, against 52.5% of organic prompts citing 10 or more (Promptwatch Data, Sep 28 to Oct 4, 2026). With so few sources, one outdated review page or forum thread can dominate how ChatGPT describes you.

The type of page matters too. In the same week, Promptwatch's content type report shows that for branded prompts, product pages were 23.3% of citations and landing pages 22.8%. Reviews rise to 5.6%, more than three times their overall share, and FAQ pages to 4.8%. So the pages you control carry a lot of weight, and third-party reviews carry more than you'd expect.

A practical reading of that:

• Make your homepage, product pages, and docs answer pricing, features, and audience questions in plain language.
• Add a brand FAQ that states the facts AI keeps getting wrong.
• Keep reviews current on trusted sites.
• Check which domains are cited for prompts about your brand, and contact owners of pages with errors.

Crawler access is the other thing teams miss. If AI bots can't read your corrected page, the old version keeps winning. IP-verified crawler logs, like Promptwatch's AI crawler logs, show whether GPTBot or OAI-SearchBot are actually reaching it. You can also test access for free with the Agentic Readiness Check.

A response playbook

Step 1: Triage

Not every error is a legal matter. Sort them.

Step 2: Preserve evidence

Save the prompt, model, location, date, and full response. Run the same prompt repeatedly to show frequency. Keep sales and support records that show real-world impact. Don't rely on memory or a single screenshot.

Step 3: Correct at the source

Find the pages the model cited. Fix your own. Email the owners of third-party pages with a short, factual correction. Publish a clear statement of the true facts on a crawlable page.

Step 4: File correction requests with the provider

I didn't find a documented OpenAI process built for brands to report inaccurate output, so this is a gap I can't fill with certainty. Use whatever feedback and content-report channels the provider offers, and keep copies. Even if nothing happens, the record shows notice, which is exactly what the Walters plaintiff lacked.

Step 5: Decide on escalation

Talk to counsel when you have repeated false statements, measurable harm, and a documented notice. Counsel will weigh defamation, trade libel, and consumer protection angles, plus whether Section 230 or the provider's disclaimers apply. I didn't research how courts treat Section 230 for generated output, so don't assume either way.

Internal compliance: your own use of ChatGPT

Brand-error risk has a mirror image. Your staff might create the false statements. A few basics:

• Write a short AI use policy. Sprintlaw's Australian guide asks the right question: does your business need one? For most, yes.
• Ban pasting confidential contracts and customer data into public chatbots unless your plan and settings allow it.
• Require human review of any published claim, especially numbers, legal statements, and comparisons.
• Remember that conversations aren't privileged. Sam Altman said in July 2025 that confidentiality for ChatGPT chats hasn't been figured out, and that OpenAI could be required to produce them in a lawsuit.
• For law firms, the ABA's Formal Opinion 512 (July 2024) sets competence, confidentiality, and verification duties, and courts have documented well over a thousand filings with fake AI citations.

Choosing monitoring that holds up as evidence

If the goal includes documentation, the tool matters. Questions to ask any vendor:

Promptwatch covers these, including CSV and Excel exports of prompts, responses, and citations, plus an audit log on Business and above. Pricing runs from $95/month (Essential) to $579/month (Business), and you can confirm current details at promptwatch.com/pricing. A free AI visibility report is available at promptwatch.com/ai-visibility-report if you just want a first look at what ChatGPT says about you. Other options with similar tracking include [tool:peec-ai], [tool:semrush-ai-visibility-toolkit], and [tool:ahrefs-brand-radar]. These are usually broader SEO suites with narrower AI data, so test whether they preserve full responses.

If you'd rather hand the whole program to a team, 1001 SEO Media runs AI search and GEO work for brands, including cleaning up the sources that shape AI answers. Counsel still handles the legal side.

Open questions to watch in 2026

• Whether Wolf River v. Google produces a ruling that treats an AI answer as a publisher statement.
• Whether other courts follow the Walters reasoning on disclaimers, or distinguish it for business plaintiffs with real damages.
• How the FTC treats AI-generated output when it is wrong about a third party, which I haven't found a clear position on.
• Whether EU accuracy rights extend in practice to organizations.
• Pending state bills, such as a Digital Provenance Act on synthetic content disclosure introduced in June 2026, which is not law yet.

Bottom line

Assume you carry the risk of wrong AI answers about your brand, and act as if you will need to prove it later. Monitor the prompts that matter, keep dated records, fix the few sources the model leans on, send correction requests, and tie any harm to business numbers. That sequence is cheap, and it is the exact set of things the plaintiff in the only major ruling so far didn't have. Start with a list of 20 prompts about your brand, run them this week, and write down what you find.