AI Regulation: Musk wants training to be speech  

Curated by Alexis Gonzalez-del-Valle, Editor, AI Weekly

AI regulation this week stopped waiting for Congress and split into three parallel tracks: constitutional litigation, a state-by-state chatbot patchwork, and platform-level rulemaking by private gatekeepers. Each track moves on its own timeline, answers to a different authority, and produces incompatible obligations for the same deployed system. Compliance programs built around a single federal framework are now structurally behind, and the gap between where rules are written and where they bind is widening every week.

Watch & Listen First

The AI Policy Podcast -- Unpacking Russian Military AI with Kateryna Bondar -- CSIS's Gregory Allen on Russia's drone ecosystem, command-and-control AI, and battlefield autonomy, April 14. (CSIS)

EU AI Act Podcast: RegInt -- Decoding AI Regulation -- Article 12 record-keeping and Article 19 logging -- what deployers need before August 2. (Spotify)

Key Takeaways

  • Brief your litigation team on First Amendment preemption theories now. The argument that model training and output are protected speech is no longer fringe -- it is a live federal case that, if it survives a motion to dismiss, reopens every state anti-discrimination AI statute. Assume the industry-wide playbook is being drafted this quarter.

  • Redraw your US chatbot compliance map by deployment state, not by federal baseline. Companion-chatbot disclosure, crisis-response protocols, AI-therapy bans, and algorithmic pricing restrictions are accreting in seven-day cycles. A national privacy policy is no longer a substitute for state-specific product configuration.

  • Treat every prompt, system message, and model output as a discoverable business record. Attorney-client privilege does not attach to conversations with a non-human counterparty; update retention schedules, legal hold processes, and FRCP 34 response playbooks before your next litigation trigger, not after.

  • Add an app-store compliance track alongside your statutory one. Private distribution gatekeepers can now demand content moderation plans and threaten removal on their own timeline, with no notice-and-comment and no appeal. Map which guidelines bind your model and who inside the company owns the relationship.

  • Lock your EU AI Act Annex III readiness on the August 2 date. Roughly 15 weeks out, a possible Digital Omnibus extension to December 2027 is not a plan. Stand up logging, record-keeping, and sandbox engagement against the current deadline and treat any slip as a bonus.

The Big Story

xAI Sues Colorado in Federal Court to Block Consumer AI Protections Act -- April 10, 2026 -- Colorado Sun

xAI filed suit in the US District Court for Colorado on April 9, naming AG Phil Weiser and seeking to block SB24-205 before its June 30 effective date. The complaint argues training and deploying an AI model is expressive conduct protected by the First Amendment, and that Colorado's "reasonable care against algorithmic discrimination" standard unconstitutionally compels speech. xAI contends the law would force Grok to "abandon its disinterested pursuit of truth and instead promote the State's ideological views." The filing also argues extraterritorial reach violates the Dormant Commerce Clause because obligations attach any time a Colorado resident interacts with the system, regardless of where xAI operates. If xAI wins on First Amendment grounds, every state AI anti-discrimination statute is in jeopardy. Parallel IAPP read on xAI's training-data case: xAI v. Bonta.

Also This Week

Heppner Privilege Ruling Hardens Into Nationwide Compliance Assumption -- April 15, 2026 -- The Daily Record
-> Judge Rakoff's February order in In re Heppner compelling production of 31 Claude-generated documents reached full industry absorption: major firms including Debevoise & Plimpton issued client advisories April 15-16 warning chatbot transcripts carry no attorney-client privilege -- "no attorney-client relationship exists, or could exist, between an AI user and a platform such as Claude," Rakoff wrote. Deployers should treat prompt logs as discoverable business records under FRCP 34 and update retention schedules.

Apple Threatens Grok Removal Over Deepfake Generation -- April 14, 2026 -- NBC News
-> A January letter from Apple to Senators Wyden, Markey, and Luján -- obtained by NBC News this week -- shows Apple demanded xAI submit a content moderation plan after Grok produced non-consensual sexualized deepfakes. Apple found both X and Grok in violation of App Store guidelines. App Store Review Guidelines 1.1.4 and 5.1.2 now function as the binding content rulebook for 1.5 billion iOS devices, with no due process, no notice-and-comment, and no appeal.

Nebraska LB 525 Signed, 49-0 Final Reading -- April 14, 2026 -- Nebraska Legislature
-> Governor Pillen signed the bundled Agricultural Data Privacy Act and Conversational AI Safety Act. The chatbot provisions require disclosure of non-human status, bar claims of licensed mental-health care, and mandate crisis-response protocols for suicidal-ideation prompts -- the fourth state this cycle to regulate companion chatbots directly.

Maryland HB 895 Clears Legislature With AI Pricing Restrictions -- April 13, 2026 -- Troutman Pepper
-> Targets algorithmic and surveillance pricing; retail, travel, and rideshare deployers should start mapping dynamic-pricing models now.

Maine LD 2082 Bars AI-Only Therapy Services -- April 13, 2026 -- Troutman Pepper
-> Prohibits offering therapy or psychotherapy through AI unless licensed; "not a therapist" disclaimers are no longer sufficient.

China's CAC Opens Consultation on Human-Like Interactive AI Rules -- April 2026 -- Mayer Brown
-> Covers AI companions and emotional interaction with prohibitions on "emotional manipulation" and self-harm encouragement that have no US analogue.

Deadlines & Compliance

  • April 22, 2026: Updated COPPA Rule compliance deadline for anyone knowingly collecting data on children under 13. (O'Melveny)

  • April 24, 2026: ADA Title II digital accessibility deadline for state and local governments serving 50,000+ residents -- covers AI-generated content. (Flockler)

  • June 30, 2026: Colorado AI Act (SB24-205) takes effect, pending xAI's injunction. Developers must exercise reasonable care against algorithmic discrimination; deployers must complete impact assessments. (Clark Hill)

  • August 2, 2026: EU AI Act Annex III high-risk obligations trigger; Member States must designate notifying and market-surveillance authorities and stand up a national AI regulatory sandbox. Digital Omnibus trilogue may push to December 2027. Build for August. (European Parliament)

Worth Reading

Musk wants a federal judge to rule that training a model is speech. Apple wants a private letter to regulate what a model can generate. The statutes haven't caught up to either -- and half of what compliance teams are currently building assumes they will.