AI Text Chatbots and the TCPA: Where Conversational SMS Gets Risky

An AI chatbot that texts customers follows the same TCPA rules as any other text, but a typed message isn’t a recorded voice, and the FCC’s AI declaratory ruling covers voice calls. Its AI disclosure proposal for texts hasn’t been adopted. Most of the risk is a bot that misses an opt-out in a customer’s own words or slips a pitch into a service chat. Both are fixable, and a company facing one of these claims often has more defenses than the demand letter suggests.

Is a chatbot’s text a call under the TCPA?

Courts have treated it as one for the autodialer and recorded-voice provisions. The statute bars any call “using any automatic telephone dialing system or an artificial or prerecorded voice” to a cell phone without prior express consent, under 47 U.S.C. § 227(b)(1)(A). The Supreme Court wrote in Campbell-Ewald Co. v. Gomez that “A text message to a cellular telephone, it is undisputed, qualifies as a ‘call’ within the compass of §227(b)(1)(A)(iii).” In California, the Ninth Circuit held in Satterfield v. Simon & Schuster, relying on deference to the FCC’s interpretation, that “a text message is a ‘call’ within the meaning of the TCPA.”

What the AI doesn’t do is turn the text into a prerecorded voice, the statute’s term for a recorded or synthetic voice message played on a call. In Trim v. Reward Zone USA, the Ninth Circuit held that pre-drafted marketing “text messages did not use prerecorded voices under the TCPA, because they did not include audible components.” A typed reply from a language model has no audio either.

That leaves the autodialer theory as the main route to consent claims over texts. After the Supreme Court’s decision in Facebook, Inc. v. Duguid, an autodialer needs “the capacity to use a random or sequential number generator to either store or produce phone numbers to be called.” A bot that only replies to the number that just texted it isn’t using a number generator to decide whom to text, though plaintiffs often plead the platform’s capacity instead. Our January post covers how to attack an autodialer allegation after Duguid.

Does the FCC’s AI voice ruling reach chatbot texts?

No. In its February 2024 Declaratory Ruling (FCC 24-17), the FCC confirmed that the TCPA’s restrictions on the use of an artificial or prerecorded voice reach “current AI technologies that generate human voices.” That ruling addresses cloned and synthetic voices on calls, not typed text.

In its August 2024 Notice of Proposed Rulemaking (FCC 24-84), the FCC proposed defining an AI-generated call to include one that uses technology “to generate an artificial or prerecorded voice or a text” through machine learning or large language models to communicate on an outbound call. It also proposed requiring callers who send “autodialed text messages that include AI-generated content” to disclose, when getting consent, that the consent may cover AI-generated content.

As proposed, the text disclosure reaches only messages sent with an autodialer, so a bot that isn’t one under Duguid wouldn’t be covered. And nothing has been adopted. The proposal was published in the Federal Register on September 10, 2024. As of this post, the FCC hasn’t adopted a final rule in that docket.

How much do FCC rulings matter in a lawsuit now?

Less than before. In June, the Supreme Court held in McLaughlin Chiropractic Associates v. McKesson Corp. that a district court “should interpret the TCPA under ordinary principles of statutory interpretation, affording appropriate respect to the agency’s interpretation.”

That cuts both ways: an FCC ruling that helps you, like the on-demand text ruling below, isn’t binding, and one that hurts you can be challenged. I’d still build to the FCC’s rules. They remain the agency’s enforcement position, and I think following them helps defeat a claim that a violation was willful or knowing.

When the customer texts first, does the bot need consent to reply?

Usually not for a direct answer. In its 2015 TCPA order (FCC 15-72), the FCC ruled that a one-time text sent immediately in response to a consumer’s specific request doesn’t violate the TCPA. The reply must be one the consumer asked for, and it must contain “only the information requested by the consumer with no other marketing or advertising information.” The FCC added that the consumer’s first text is consent “to an informational reply in fulfillment of the consumer request.” Our post on text-to-enter sweepstakes and TCPA consent applies the same ruling to contest keywords.

A chatbot conversation stretches that ruling, because it’s a back-and-forth, not one reply. My view is that each answer to a question the customer just asked is the easiest message in the TCPA to defend. The risk starts when the bot texts first, like a follow-up nudge two days after the customer went quiet. Those are messages the business initiated, and they need their own consent analysis.

Can the bot tell when a customer is opting out?

It has to. Since April 11, 2025, the FCC’s revocation rule at 47 C.F.R. § 64.1200(a)(10) lets a person revoke consent “by using any reasonable method to clearly express a desire not to receive further calls or text messages from the caller or sender.” Replying with one of seven standard words (stop, quit, end, revoke, opt out, cancel, or unsubscribe) is reasonable per se. For any other wording, the rule provides that “the caller must treat that reply text as a valid revocation request if a reasonable person would understand those words to have conveyed a request to revoke consent.” Requests must be honored within a reasonable time, and no later than ten business days after receipt.

The rule applies to consent for texts that need it under the FCC’s robocall rules, and to written permission used to text a number on the National Do Not Call Registry, the national list of numbers whose owners don’t want telemarketing. Even where it doesn’t apply, a company that texts marketing to consumers’ personal phones must keep its own do-not-call list and honor requests within a reasonable time, capped at the same ten business days, under § 64.1200(d)(3). Our April post explains what a defensible internal do-not-call policy contains.

A language model should beat a keyword filter here, catching a reply like please stop messaging me. The FCC’s 2024 order (FCC 24-24) says the fact finder “will conduct a totality-of-circumstances analysis” of whether the request was reasonable. The sender gets a chance to explain why other wording wasn’t a reasonable request, and the order notes that when a business texts in another language at the consumer’s request, that language’s equivalent of the standard words may count. I think a company that pays for a conversational AI and still listens only for STOP is ignoring the best feature it bought.

Under § 64.1200(a)(12), one confirmation text is allowed if it only confirms the request, is the only message sent after it, and “does not include any marketing or promotional information.” A confirmation sent within five minutes is presumed to fall within the consumer’s consent. A bot trained to save the relationship will want to add a discount offer or ask the customer to reconsider.

The FCC’s Consumer and Governmental Affairs Bureau, in DA 25-312, delayed until April 11, 2026 the part of the rule that would make an opt-out from one type of message apply to all future robocalls and robotexts from that caller on unrelated matters. Until then, a STOP reply to an appointment reminder can be applied to that category alone, but I’d log it for when the broader rule takes effect.

What happens when a service chat turns into a sales pitch?

The rules can change with it. The FCC’s rules define telemarketing as initiating a message to encourage “the purchase or rental of, or investment in, property, goods, or services.” A billing answer that adds a pitch for the premium plan risks being treated as telemarketing, though a business can argue that a reply to the customer’s own text isn’t a message it initiated.

The on-demand ruling no longer fits, because its third condition is a reply with no other marketing. If the platform were found to be an autodialer, marketing texts would need prior express written consent under § 64.1200(a)(2), meaning a signed agreement (an electronic signature counts) that authorizes marketing messages to a named number and includes the required disclosures. And the internal do-not-call rules apply to telemarketing, so a customer who earlier asked to stop marketing shouldn’t get the upsell.

Registry claims have defenses that fit chat well. The FCC’s definition of a telephone solicitation excludes a message sent “To any person with that person’s prior express invitation or permission,” and one sent to a person with an established business relationship, which lasts 18 months after a purchase or three months after an inquiry and ends with a do-not-call request. I’d argue that a customer who asks the bot about upgrade pricing has invited the answer. The Registry rule’s separate permission exception requires a signed written agreement, so that argument rests on the definition. The cleanest control is a scope rule in the bot’s instructions: service threads stay on service unless the customer asks about products.

In July, a federal court in Illinois dismissed Do Not Call claims built on texts in Jones v. Blackstone Medical Services, concluding that “under a plain reading, Section 227(c)(5) of the TCPA does not regulate text messages.” It’s one district court’s ruling, without prejudice, and the FCC’s rules say the Do Not Call protections reach texts, but a business sued over texts alone should raise it.

Do any state laws require a bot to say it’s a bot?

In some states. California’s and Utah’s laws are the two to know, and neither creates its own private damages claim.

California. Business and Professions Code § 17941 makes it unlawful to use a bot to communicate with a person in California online “with the intent to mislead the other person about its artificial identity” and knowingly deceive the person about the content of the communication, to incentivize a purchase or sale or to influence a vote. A business isn’t liable if it discloses that it is a bot in a clear and conspicuous way. The statute’s definition of “online” refers to websites and applications, and whether it reaches an ordinary text thread is an open question.

Utah. Since May 7, 2025, Utah law, as enacted by Senate Bill 226, requires a supplier using generative AI in a consumer transaction to disclose that the person is interacting with AI if the person asks. The law covers AI that produces non-scripted responses with limited or no human oversight, so a menu-driven bot may fall outside it. The question must be “a clear and unambiguous request to determine whether the interaction is with a human or with artificial intelligence.” Licensed or state-certified professionals in regulated occupations must disclose up front in high-risk interactions. A safe harbor from enforcement of the disclosure section protects a business whose AI discloses that it is AI at the outset and throughout the interaction, if the disclosure is clear and conspicuous. The state’s Division of Consumer Protection enforces the chapter with fines of up to $2,500 per violation, and a violation also counts as a violation of Utah’s Consumer Sales Practices Act.

How the theories line up against a chatbot program

Securing the chatbot record when a claim arrives

1. Preserve the full conversation transcripts and platform send logs under a litigation hold, along with the consent record for the number and the customer’s first inbound text.
2. Gather every opt-out and internal do-not-call entry for the number, time-stamped, plus the bot’s log showing how it classified each inbound message.
3. Find the Registry scrub reports, with the version date of each Registry download, for any number that got a marketing message.
4. Collect the AI vendor and texting-platform contracts, the bot’s prompts and instructions, the model version in use, and any message templates.
5. Have a lawyer review the file before anyone answers a demand letter, contacts the customer, or offers to settle.

When a claim involving automated text conversations and missed opt-out requests is filed as a proposed class action, the firm’s page on defending TCPA class actions in California covers what comes next.

A bot log showing it read a message as a question, not an opt-out, is strong evidence of good faith and of how the message read in context. Carrier rules add their own opt-out expectations, and our June post covers CTIA’s messaging principles and 10DLC registration.

Questions businesses ask about AI chatbots and the TCPA

Does federal law require our text bot to say it’s AI?

No TCPA rule requires it yet. The FCC has only proposed AI disclosure rules, and its proposal ties text disclosures to autodialed messages. For Utah and California, a clear disclosure at the start of every conversation goes a long way, and Utah’s safe harbor asks for one that continues throughout.

Can the bot send a reminder if the customer stops replying?

Yes, but it’s a new message the business initiated. If the reminder is informational and the customer gave consent for those messages, it’s usually fine. If it pushes a sale, it’s telemarketing, and the do-not-call rules apply. Cap follow-ups in the bot’s instructions, and stop them once the customer says anything a reasonable person would read as no.

What should the bot do with an unclear reply?

For a reply like not now or maybe later, the cautious move is to treat it as an opt-out for the topic at hand. The rule only requires honoring words a reasonable person would read as a request to stop. If the customer consented to several categories of messages, the confirmation rule lets the business ask, without marketing, which ones the opt-out covers, and it must stop those texts until the customer answers.

Defending claims over AI-driven texts

If your company is accused of texting through a chatbot after a customer asked it to stop, the bot’s transcripts and classification logs usually decide the case. Rosenberger + Kawabata represents businesses defending TCPA lawsuits and class actions through its business litigation practice, including claims involving automated text conversations and missed opt-out requests. Contact Rosenberger + Kawabata online for a free and confidential consultation, or call (310) 894-6921.

Sources

The statute is the Telephone Consumer Protection Act at 47 U.S.C. § 227. The FCC’s rules are at 47 C.F.R. § 64.1200, as in effect on September 7, 2025. The FCC documents are the AI Declaratory Ruling, FCC 24-17; the AI Notice of Proposed Rulemaking, FCC 24-84; the revocation order, FCC 24-24; the Bureau’s waiver order, DA 25-312; and the 2015 Declaratory Ruling and Order, FCC 15-72. The state laws are Cal. Bus. & Prof. Code § 17941 and Utah Code §§ 13-75-101 to -106, enacted by S.B. 226 (2025).

The court decisions are Campbell-Ewald Co. v. Gomez, 577 U.S. 153 (2016); Facebook, Inc. v. Duguid, 592 U.S. 395 (2021); McLaughlin Chiropractic Associates, Inc. v. McKesson Corp., No. 23-1226 (U.S. June 20, 2025); Satterfield v. Simon & Schuster, Inc., 569 F.3d 946 (9th Cir. 2009); Trim v. Reward Zone USA LLC, 76 F.4th 1157 (9th Cir. 2023); and Jones v. Blackstone Medical Services, LLC, No. 1:24-cv-01074 (C.D. Ill. July 21, 2025).

This post is general information about the Telephone Consumer Protection Act and related laws as of September 7, 2025. It is not legal advice, and reading it does not create an attorney-client relationship. Courts and the FCC change the rules in this area often, and the outcome of any claim depends on its facts. Past results do not guarantee future outcomes.