Does Section 230 protect a company from its own chatbot's output?

It is a much weaker shield than most teams assume. Section 230 immunizes providers from liability for information provided by another information content provider. Text your own deployed model composes is not obviously someone else's content, and courts have been skeptical of extending the immunity to generative output. Plan as though it does not apply.

Does a disclaimer that the chatbot may be inaccurate protect us?

Not reliably. A disclaimer helps show a statement was not offered as settled fact and can undercut reasonable reliance, but it does not immunize a specific false factual claim about a named third party. It is a mitigating factor, not a defense — and it does little where the harmed party never saw the disclaimer at all.

Are we bound by a promise the chatbot made to a customer?

Often, in practical terms. If a bot on your own site quotes a discount, states a refund policy, or promises a delivery date, a customer who reasonably relies on it has strong arguments in contract, misrepresentation, and consumer-protection law. Businesses have been held to policies their assistants invented rather than allowed to disown them after the fact.

What is the biggest defamation risk in a customer-facing bot?

Competitor comparisons. A bot asked why it beats a named rival will readily generate specific-sounding claims about that rival's security incidents, lawsuits, or practices. Those statements are false statements of fact about an identifiable business made in a commercial context, which is both defamation and Lanham Act false advertising territory.

How do we reduce the risk without turning the chatbot off?

Constrain the domain rather than trusting the model. Ground answers in retrieval over approved documents, refuse questions about named third parties, block commitments about price, refunds, medical, and legal matters, log every conversation, and route anything outside the safe domain to a human. Most real incidents came from an unconstrained bot answering a question it never should have taken.

AI Chatbot Defamation Liability 2026: When Your Bot's Hallucination Becomes Your Lawsuit

Q: Can a business be sued for something its AI chatbot said?

Yes. When your bot generates a false statement of fact about an identifiable person or company and a third party reads it, the elements of defamation are present and your business is the publisher. There is no separate legal category for statements a machine produced on your behalf — the deployer is the party in the caption.

Defamation Does Not Care How the Sentence Was Produced

The classic elements are unchanged: a false statement of fact, about an identifiable person or entity, published to a third party, causing harm, with some level of fault. Nothing in that list asks whether a human or a language model composed the words. If your bot tells a prospect that a named competitor "lost customer data in a breach last year" and that is not true, you have published a false factual claim about an identifiable business.

The fault element is where teams hope to escape and usually cannot. Deploying a system you know produces confident fabrications, in a public channel, without guardrails or review, is not an obviously innocent state of mind. Internal evaluations documenting a known hallucination rate cut both ways — they show diligence, and they show knowledge of the risk you shipped anyway.

Four Distinct Claims From One Bad Answer

Statements About Others

•Defamation for false factual claims about a person or company
•Trade libel where the target is a competitor's product or service
•Lanham Act false advertising for commercial comparative claims
•Right-of-publicity issues when a real person is invoked in promotion

Statements To Customers

•Negligent misrepresentation when a customer relies on invented terms
•Contract and promissory theories for prices, refunds, and guarantees
•State UDAP and FTC Act exposure for deceptive statements at scale
•Regulated-advice risk in health, legal, financial, and housing contexts

The Section 230 Assumption Is the Dangerous One

Platform teams carry a reflex that user-generated content is somebody else's problem. Section 230 protects interactive computer services from liability for information provided by another information content provider — the operative words being "another" and "provided by." Text your own model composes in response to a prompt is not obviously supplied by someone else, and the trend in commentary and early rulings runs against stretching the immunity to cover generative output. Treat it as unavailable when you plan controls, and consider anything you get from it a bonus.

Comparison Questions Are the Highest-Risk Prompt You Serve

Sales-assist bots are explicitly built to handle "how do you compare to X." That is precisely the prompt that pushes a model toward specific, confident, unverifiable claims about a named third party — a rival's outage history, pricing practices, litigation, or security posture. The output is commercial speech, about an identifiable competitor, published by you, with an obvious economic motive. It is difficult to construct a worse fact pattern on purpose.

The fix is not better prompting. It is a hard refusal boundary: the bot does not make factual assertions about named third parties, full stop, and instead returns an approved comparison page maintained by marketing and reviewed by legal. Deterministic refusal beats probabilistic good behavior every time.

Why the Disclaimer Does Less Than You Think

"This assistant may produce inaccurate information" is worth including. It can undercut the reasonableness of a customer's reliance and signal that answers are not offered as settled fact. What it cannot do is neutralize a specific false statement about a third party — the defamed competitor never saw your disclaimer and never agreed to anything. Nor does boilerplate reliably defeat a consumer-protection claim when a business systematically misstates its own terms. Disclaimers reduce the severity of some claims. They do not remove the category.

Chatbot Liability Controls Checklist

Immediate Actions

☐Block factual assertions about named competitors and individuals
☐Ground answers in retrieval over approved, versioned source documents
☐Prohibit commitments on price, refunds, eligibility, and delivery dates
☐Log every conversation with the model version and retrieved sources
☐Red-team with adversarial competitor and complaint prompts before launch

Within the Quarter

☐Add a clear notice that the user is talking to an AI assistant
☐Build a takedown and correction path for a reported false statement
☐Escalate regulated topics to a human rather than answering
☐Confirm media liability or tech E&O covers AI-generated statements
☐Review vendor terms for who bears liability for model output

Frequently Asked Questions

Can a business be sued for something its AI chatbot said?

Yes. A false statement of fact about an identifiable person or company, published to a third party, is defamation regardless of whether a human or a model composed it. The business that deployed the bot is the publisher and the defendant.

Does Section 230 cover our own chatbot's output?

Assume not. The immunity is for information provided by another information content provider, and text your deployed model generates is not clearly someone else's content. Build your controls as though the shield is unavailable.

Will an accuracy disclaimer protect us?

Partially at best. It can weaken a reliance argument and signal that answers are not authoritative, but it cannot cure a specific false claim about a third party who never saw it, and it rarely defeats a consumer-protection claim about your own terms.

Are we bound if the bot promises a discount or refund?

Frequently, in practice. A customer who reasonably relies on a commitment made by your own assistant on your own site has real arguments in contract, misrepresentation, and consumer-protection law. Blocking commitments is far cheaper than litigating them.

Where does defamation risk concentrate?

Competitor comparison prompts. They invite confident, specific, unverifiable claims about a named rival in a commercial context — defamation and false-advertising exposure in the same sentence. Refuse the category rather than trying to prompt around it.

What single control reduces the most risk?

A hard refusal boundary combined with retrieval grounding. Constrain the bot to a defined domain backed by approved documents, refuse everything outside it, and log what was retrieved for each answer so you can reconstruct any disputed conversation.

Treat the Bot as a Publishing Channel, Because It Is One

No marketing team would push thousands of unreviewed public statements a day about customers, competitors, and company policy. A customer-facing chatbot does exactly that. Constrain the domain, refuse third-party claims, ground answers in approved sources, and log everything — and the channel becomes defensible rather than merely fast.

The same applies to everything else your site says in public. Run a free scan of your site to see what's live right now.