California AB 2905 Artificial Voice Disclosure Checklist for Robocalls: A 2026 Compliance Guide

California AB 2905, signed September 20, 2024 and effective January 1, 2025, amended Public Utilities Code §2874 to require every robocall using an AI-generated or significantly altered voice to disclose that fact in the natural-voice opening announcement at the start of the call — closing a gap that the January 2024 New Hampshire deepfake robocall mimicking President Biden made impossible to ignore. This is the operational compliance checklist for marketers, political campaigns, debt collectors, healthcare appointment-reminder systems, and any other operation running California-bound robocalls through an automatic dialing-announcing device. We walk through who is covered, what the disclosure has to say, where AB 2905 fits with the federal TCPA, and the integration patterns that production-grade voice campaigns are using to satisfy the requirement without breaking the rest of their compliance posture.

The Biden deepfake robocall and why AB 2905 exists

The clearest way to understand AB 2905 is to start with the event that made it inevitable. In January 2024, New Hampshire Democratic primary voters received robocalls that used a synthetic voice cloned to sound like President Biden, telling them to skip the primary. The calls reached an estimated 5,000 to 25,000 voters, and the deepfake-voice attack pattern they demonstrated — clone any public figure's voice, push the resulting audio through a high-volume automated calling system, deliver it to thousands of recipients in hours — was suddenly visible to every state legislature watching for AI-related election integrity threats. California's response was AB 2905, drafted by Assemblymember Evan Low and signed by Governor Newsom that September.

The statutory mechanism is narrow but pointed. California's existing automatic dialing-announcing device law, Public Utilities Code §2874, already required every robocall to begin with a natural-voice opening announcement before playing the prerecorded message. The opening announcement had to state the nature of the call, identify the calling business, and ask the recipient for consent to continue. AB 2905 added one more element to that opening announcement: if the prerecorded message that follows uses an AI-generated or significantly altered synthetic voice, the natural-voice opening announcement must disclose that fact. A recipient who picks up the phone and hears "This message uses an artificial voice generated by AI" before any synthetic content plays has been informed in time to evaluate what comes next; a recipient who hears synthetic Biden audio without that warning has not.

Who is covered, and the threshold question

AB 2905's scope is anyone operating an automatic dialing-announcing device that places telephone calls through California's telephone network. The categories that show up most often are commercial telemarketers, political campaigns, debt collectors, healthcare appointment-reminder systems, customer service callback systems, and survey research firms. The law does not have a size or volume threshold — a single robocall using an AI voice triggers the disclosure obligation, the same way a single non-AI robocall already triggers the existing PUC §2874 opening-announcement obligation. The law applies to calls received in California regardless of where the calling operation is physically based, which means out-of-state operators calling into California are subject to AB 2905 even if their home state imposes no equivalent rule.

The threshold question that comes up repeatedly in compliance review is what counts as an "automatic dialing-announcing device" in modern voice-tech architecture. Traditional ADADs were physical hardware that played a recorded message. Modern systems are software platforms that synthesize, schedule, and place calls programmatically. The CPUC has interpreted the statute to cover any system that automatically delivers a prerecorded message to a called party regardless of underlying technology, which means cloud-based voice platforms, AI agent systems, and software-defined dialers all fall within scope. The exemption everyone tries to claim — "our system is too advanced to be an ADAD" — does not work; the regulatory frame is functional, not technological.

What the disclosure has to say

AB 2905 does not prescribe exact wording, but the structural requirement is clear: the AI-voice disclosure must be part of the natural-voice opening announcement that PUC §2874 already requires, before any AI-generated audio plays. Compliant phrasings that have emerged in production include "This message uses an artificial voice generated by AI," "The voice you are about to hear is artificially generated using AI," and "The following message uses an AI-synthesized voice." The phrasing should be clear enough that an ordinary listener understands the disclosure without specialized knowledge — "This message uses generative AI text-to-speech synthesis" is technically accurate but linguistically inaccessible to most call recipients.

The disclosure must be in a natural human voice, not in the AI voice itself. That is the structural point of the opening announcement requirement: a real human speaks the disclosure, including the AI-voice element, before the synthetic audio begins. Operators that try to satisfy the requirement by having the AI voice itself deliver the disclosure are not complying with the statute's structure; the statute treats the natural voice and the synthetic voice as separate components of the call, with the natural voice carrying the disclosure responsibility. Most production implementations record the disclosure phrase in the campaign's standard human-voice opening template, then concatenate the AI-generated content as the body of the call.

What "significantly altered" means in practice

The phrase "generated or significantly altered" defines the scope of the disclosure trigger. Voice cloning, full text-to-speech synthesis, and substantial AI-driven voice transformation clearly fall within scope. Standard audio editing — background noise removal, volume normalization, brief pitch correction — does not. The hard cases are in the middle: AI-driven accent neutralization, synthetic emotional tone overlay, voice age or gender alteration, and similar techniques that change the perceived character of the voice without being full synthesis.

The bright-line test most compliance programs are using is whether the voice would be perceived by a typical listener as a different person from the speaker who recorded the original audio. If a recorded message of a 50-year-old male campaign volunteer is AI-altered to sound like a 25-year-old female candidate, the alteration is significant — the perceived speaker has changed. If the same recorded message is AI-cleaned to remove background noise but the speaker remains recognizable, the alteration is not significant. Edge cases get triaged conservatively, with the disclosure included whenever there is reasonable doubt; the cost of including the disclosure unnecessarily is small, and the cost of omitting it when required is enforcement exposure plus the reputational risk of a privacy or election-integrity headline.

How AB 2905 integrates with the federal TCPA

AB 2905 operates alongside the federal Telephone Consumer Protection Act, not in place of it. The TCPA governs robocalls broadly through restrictions on prior express consent, calling-window restrictions, and do-not-call registry compliance, with statutory damages of $500 per violation that can be trebled to $1,500 for willful violations. AB 2905 adds a California-specific disclosure layer on top: even a robocall that fully complies with TCPA still violates AB 2905 if it uses an AI voice without the opening-announcement disclosure.

Most compliance programs are integrating AB 2905 into their existing TCPA infrastructure by treating the AI-voice disclosure as a mandatory element of the natural-voice opening template for any campaign that uses synthetic or substantially altered voices. The implementation is straightforward in the systems where TCPA compliance is already operationalized — dialer platforms have settings for opening-announcement content, and adding the AI-voice disclosure to those settings is a configuration change rather than a re-architecture. The harder integration is for campaigns that have historically run consent and disclosure compliance manually rather than through their dialer platform, because manual processes scale poorly when the AI-voice production pipeline starts producing campaigns at higher volume.

The federal-state interaction with AB 2655 and election deepfakes

Political campaigns running AI-voice robocalls into California face two state-law obligations simultaneously: AB 2905's artificial-voice disclosure and AB 2655's election deepfake regime. AB 2655 governs materially deceptive AI-generated content depicting candidates and elected officials during specified pre-election and post-election windows; it imposes labeling and removal obligations on large online platforms and creates causes of action for affected candidates. The two laws operate at different points in the content lifecycle — AB 2905 at the moment of call delivery, AB 2655 at the moment of content distribution — and a single political robocall using a deepfake voice to depict a candidate during an election window can violate both.

For non-political robocalls (commercial telemarketing, debt collection, customer service), AB 2905 is generally the only state-specific AI-voice regime in play, with TCPA as the federal floor. Healthcare appointment-reminder systems also need to consider AB 3030 when the AI-voice call delivers clinical content; the AB 3030 disclosure for clinical communication is separate from the AB 2905 disclosure for AI-voice technique, and a healthcare robocall that uses an AI voice to deliver a clinical reminder needs both.

The compliance checklist

For operators running California-bound robocalls, the practical compliance posture has four elements. First, audit your voice production pipeline: identify every campaign that uses AI-generated or significantly altered voices. Second, update your natural-voice opening announcement template to include the AI-voice disclosure for the affected campaigns, recorded in a real human voice rather than synthesized. Third, configure your dialer platform to use the updated template for any campaign flagged as AI-voice. Fourth, document the audit and configuration in a written compliance memo so the audit trail exists when a CPUC inquiry or TCPA litigation arrives. The substantive work is small if your dialer platform already handles opening-announcement content cleanly; most of the work is policy and audit-trail discipline rather than engineering.

For operators running campaigns at scale, the additional layer is monitoring: a system that periodically samples outbound calls, transcribes the opening announcement, and verifies that the AI-voice disclosure is present where required. The monitoring data is what defends compliance posture in retrospective enforcement inquiry; without it, the question "did your disclosure actually play on every applicable call" cannot be answered with evidence.

Sources

The primary statute is AB 2905 on California Legislative Information. For the broader regulatory context, EPIC's legislative session roundup and the California Attorney General's legal advisory on AI place AB 2905 in the broader 2024 California AI legislation cycle. For practitioner-grade analysis, ZwillGen's overview and EM3 Law's explainer are useful references. Watch the CPUC for any pre- or post-effective-date guidance on what counts as "significantly altered" under the statute.