A complaint that keeps growing
On July 7, 2026, plaintiffs' counsel at Lieff Cabraser Heimann & Bernstein and Baehr-Jones Law filed an amended complaint in the Doe litigation against xAI in the Northern District of California, adding Stability AI as a defendant and two new minor plaintiffs, Jane Doe 4 and Jane Doe 5 (Lieff Cabraser). According to the amended pleading, Jane Doe 4's stepfather used a photo of her at age 11 to generate roughly 7,000 sexually explicit images through Grok and traded them online; Jane Doe 5's abuser used her eighth-grade graduation photo the same way (CyberScoop). The complaint also alleges xAI obstructed a law enforcement investigation by withholding user and IP information, and that Stability AI trained open-weight models on datasets containing CSAM and stripped safety restrictions to chase adoption. This case sits alongside a related suit, Doe v. X.AI Corp., No. 5:26-cv-00772 (N.D. Cal.), filed in January 2026 over Grok's non-consensual sexualized deepfakes of adult women (N.D. Cal. case docket).
Why Section 230 is actually in play
Section 230 of the Communications Decency Act shields a platform from liability for content posted by someone else. Historically that has meant Grok's operator would need to look like a mere conduit for a third party's abusive prompt. But the plaintiffs' theory doesn't hinge on the prompt — it hinges on the image. Grok's diffusion model is the thing that renders the CSAM pixel by pixel; the user supplies a request, not the output. That distinction is exactly what a California federal court explored in Bouck v. Meta (C.D. Cal., Mar. 24, 2026), holding Meta could be treated as the creator of fraudulent ads its generative tools produced, rather than a neutral host, because the platform's own AI "literally generat[ed]... the images and text" (National Law Review). If a court applies that same creator-versus-host logic to Grok, Section 230 does not need to be repealed or amended for AI to lose the shield here — it was arguably never designed to reach a company's own generative output in the first place.
Steelmanning the case for stripping immunity
Child-safety advocates have the stronger argument on the facts of this specific case. CSAM is not protected speech, has no First Amendment value, and Congress has already legislated around its production and distribution independent of any platform-immunity debate — Masha's Law, 18 U.S.C. § 2255, entitles a minor victim to $150,000 in liquidated damages without needing to prove actual harm (Cornell Law). If a company's own model is the direct instrument that manufactures the abusive image — not a bystander hosting someone else's post — then treating that company as a neutral publisher stretches Section 230 past its statutory text, which immunizes treating a defendant "as the publisher or speaker of any information provided by another information content provider." When the company's model is the content provider, the predicate for immunity is arguably missing.
Where the pro-innovation case still holds
The risk is not that xAI and Stability AI should escape liability for CSAM their tools generate — they shouldn't, and existing product-liability, negligence, and federal CSAM statutes already reach that conduct without touching Section 230 at all. The risk is that a case this sympathetic becomes the vehicle for a much broader legislative rewrite of Section 230 aimed at all AI features, including the unglamorous ones — spam filters, live captioning, translation, and content-ranking tools that route around individually reviewing every output (Cato Institute). EFF has made a similar case for restraint, arguing it would be "exceedingly hard to write a law that would protect users' rights before courts have a chance to clarify these issues" and favoring case-by-case judicial line-drawing over a blanket statutory carve-out (EFF). That is the right instinct: the Bouck creator-versus-host test, applied narrowly to direct AI-generated output, gets courts to the right result in the Grok litigation without collapsing the broader immunity that lets smaller AI startups operate without pre-clearing every generated token.
The proportionate path
Congress already showed it can legislate narrowly here: the TAKE IT DOWN Act criminalized publishing non-consensual intimate imagery, including deepfakes, and gave platforms a 48-hour takedown clock, without amending Section 230's text. That is the model — target the harmful conduct and the specific content category, not the immunity statute wholesale. Courts applying Bouma-style creator analysis to genuinely AI-generated CSAM, combined with existing federal criminal exposure under Masha's Law, can hold xAI and Stability AI accountable for what their own models produced. What the Doe litigation should not become is the test case that persuades Congress to gut Section 230 for every AI feature merely because one company's flagship product handled child safety recklessly.