A 39-page complaint landed with four days on the clock

On Monday, July 27, 2026, a complaint hit the docket of the U.S. District Court for the District of Minnesota. Plaintiff: X.AI LLC. Defendant: Keith Ellison, Minnesota Attorney General. Case number 0:26-cv-03425. Length: exactly 39 pages, captioned "Complaint for Declaratory and Injunctive Relief."

The timing isn't subtle. The law xAI is attacking is Minnesota's HF 1606, known as the nudification technology ban, and it takes effect August 1, 2026 — four days after the filing. This is the classic pre-enforcement play: get a judge to freeze a statute before it ever gets a chance to bite.

Here's the deal with what the law actually says. Anyone who owns or controls a website, application, software, or other service must not allow a user to access, download, or use that service to "nudify" an image or video. Break it and the state AG can seek up to $500,000 in civil penalties per violation. Separately, anyone depicted in a nudified image gets a private right of action worth up to triple their actual damages. Minnesota is the first state in the country to build a ban like this from scratch.

xAI's opening paragraph doesn't hedge. The statute, it says, "imposes an overbroad, content-based ban on free speech and the tools of visual expression in a clumsy attempt to prohibit 'nudification.'" But in that same paragraph, xAI plants a flag: the company "strictly prohibits its users from generating nude or sexualized images of people without their consent and has indeed filed suit against users who evade its extensive technological blockers." Translation: we're not here to defend deepfake porn, we're here because this law was drafted badly.

Ellison's response was short and pointed. He noted his office hadn't been served yet, then said: "I know that using AI to generate nude images of people against their will is appalling. There are plenty of worthy debates to have about AI policy. This is not one of them. AI nudification robs the target of their dignity and can cause immense harm on an emotional, personal, and professional level." Governor Tim Walz was even shorter: "See you in court, creep."

Three parties in the ring — xAI, Ellison, and a law that passed 132 to 1

Start with xAI. Paragraph 11 of the complaint lays out the corporate facts: a limited liability company organized under Nevada law, principal place of business in Austin, Texas. And then, pointedly, "xAI maintains no offices in Minnesota." That's not filler. It's setting up the argument that a company with zero physical presence in the state is being forced to reshape a nationwide product because of one state's statute.

The product at issue is Grok Imagine. Paragraphs 18 through 20 walk through its lineage in unusual detail. Model development started around March and April 2023. Grok-1 got a limited public release in November 2023. On December 9, 2024, xAI announced Aurora, its own autoregressive image generation model that "excels at photorealistic rendering" and shipped with native multimodal input so it could directly edit user-supplied images. Grok Imagine rolled out as a dedicated experience in late July 2025, and roughly a week later — August 5, 2025 — Elon Musk posted on X: "Grok Imagine usage is growing like wildfire. 14 million images generated yesterday, now over 20 million today!" After that came Imagine v0.9 in October 2025, Grok Imagine 1.0 in February 2026 (10-second videos, 720p), a "Quality Mode" in May 2026, and Grok Imagine Video 1.5 just last month.

One piece of background worth flagging: xAI was folded into SpaceX in February 2026, so rockets, satellites, and frontier AI now sit under one roof in Musk's empire. You can see it in the complaint's own footnotes — older posts are cited to @elonmusk, recent product announcements to @SpaceXAI. The plaintiff is still named X.AI LLC, but functionally this is a subsidiary of a much larger holding structure filing a constitutional challenge against a state government. That's why CNBC's headline read "SpaceX's xAI sues Minnesota."

Second party: Keith Ellison. He's sued exclusively in his official capacity, so this isn't about personal liability. xAI's hook is that HF 1606 expressly vests enforcement authority in the Attorney General under § 325E.91, subd. 5(a), and that Ellison has never disclaimed any intent to enforce it. Paragraph 17 leans on Eighth Circuit precedent (281 Care Committee v. Arneson) for the proposition that even "some connection" between an attorney general and a challenged statute makes him a proper defendant. Worth knowing: Ellison is an incumbent running for a third term on November 3, 2026. It's hard to see how this lawsuit hurts him politically.

The third party is the one people keep skipping — the Minnesota Legislature. HF 1606 was not a party-line fight. Chief House author was Rep. Jessica Hanson (DFL-Burnsville); the Senate author was Sen. Erin Maye Quade (DFL-Apple Valley). Introduced February 26, 2025. Passed the House 132–1 on April 23, 2026, and the Senate 65–0 on April 29. Presented to the governor May 6, signed by Walz May 7, and filed with the Secretary of State the same day as Chapter 72 of 2026. The single no vote came from Rep. Drew Roach (R-Farmington), whose objection was that the bill goes after software makers instead of perpetrators. Which, awkwardly, rhymes with what xAI is arguing now.

The legislative record shows up in the complaint too. At a February 19, 2026 Senate Judiciary and Public Safety Committee hearing, when a committee staffer pointed out that the prohibition applies to consensual images, Maye Quade said "that is intentional." At the same hearing, Sen. Kreun asked whether anything in the bill would protect a company using its best efforts to prevent prohibited generations — and according to paragraph 33, she deflected rather than answer. xAI's lawyers are using both moments as evidence that lawmakers knew exactly what they were building.

What's actually written in those 39 pages

The dispute boils down to three things: no intent requirement, an enormously broad definition of "intimate part," and no dissemination requirement. Let's take them in order.

One: strict liability. HF 1606 contains no knowledge, intent, or purpose element at all. In the complaint's words, liability is "keyed solely to whether a user succeeded in creating a covered image using the AI provider's platform" — regardless of whether the provider forbids that use, how many mitigations it deployed, or how aggressively it polices abuse. There is no safe harbor for good-faith effort. xAI backs this up with its own numbers. Its acceptable use policy explicitly bans "undressing or nudifying real persons" and circumventing safety mitigations. And in 2026 alone, the company says it has suspended more than 50,000 accounts, filed more than 70,000 reports to NCMEC, and that those reports produced at least 244 arrests. The source is xAI's own separate complaint filed July 14, 2026 in the Northern District of Texas (X.AI LLC v. Harwood, 7:26-cv-00078-O), where it sued a user for breach of contract after he allegedly evaded technical controls to generate child sexual abuse material.

Two: the definition problem. This is where the drafting really gets interesting. HF 1606 doesn't define "intimate part" itself — it borrows from Minn. Stat. § 609.341, subd. 5, a criminal sexual conduct provision. That statute says intimate parts "includes the primary genital area, groin, inner thigh, buttocks, or breast of a human being." For a touching offense, that scope makes sense. Port it into an image-generation statute and, as xAI argues, you've suddenly criminalized generating pictures of people in shorts, shirtless men, and competitive swimmers in speedos. The definition also opens with "includes," making it non-exhaustive — so it can arguably stretch beyond the listed body parts.

And here's the smoking-gun detail. The original draft of HF 1606 cross-referenced a much tighter definition from Minn. Stat. § 604.32: "the genitals, pubic area, partially or fully exposed nipple, or anus." Discrete anatomical features, not general regions. Per paragraph 30, the legislature considered both definitions and chose the broader one. For a plaintiff arguing overbreadth, that legislative history is gold.

Three: no dissemination requirement. Liability attaches at creation. As paragraph 64 puts it, an image "generated once, seen only by the user who made it, and immediately deleted incurs the same $500,000 penalty as an image broadcast to the world." And the word "consent" appears nowhere in the statute. An adult editing a photo of themselves is treated identically to someone building a deepfake of an unwilling stranger.

xAI also did the arithmetic. Because the penalty is per image, ten violating images means up to $5 million; a thousand means up to $500 million; and a hundred thousand — which the complaint calls "not at all unlikely for a publicly available program with millions of users generating billions of images" — means up to $50 billion. That's a theoretical statutory ceiling, not a prediction of any judgment. But it's the engine of the chilling-effect argument. In paragraphs 7 and 45–46, xAI says it plans to restrict Grok Imagine's image-editing features in various ways before August 1 for Minnesota users, without specifying which ones.

The legal architecture is deliberately narrow: one count, First Amendment, brought under 42 U.S.C. § 1983 and 28 U.S.C. § 2201(a), seeking both facial and as-applied invalidation. xAI says the statute is content-based, triggers strict scrutiny, and can't survive it. One deliciously ironic detail: when the complaint describes strict scrutiny as "fatal in fact absent truly extraordinary circumstances," it cites Free Speech Coalition v. Paxton, 606 U.S. 461 (2025) — the case the state of Texas won last year. Other citations stack up fast: Brown v. Entertainment Merchants (violent video games), Reed v. Town of Gilbert (signs), Ashcroft v. Free Speech Coalition (virtual child porn), Counterman v. Colorado (subjective mental state), and Erznoznik v. Jacksonville from 1975, which struck down an ordinance barring films with "any uncovered buttocks or breasts" as overinclusive.

The complaint also offers less restrictive alternatives, and they're real. Minnesota already criminalizes intentional dissemination of nonconsensual deepfakes under § 617.262, with a parallel private action and a civil penalty up to $100,000 under § 604.32. Federally, the TAKE IT DOWN Act makes it a crime to knowingly publish a "digital forgery" and requires platforms to run notice-and-removal processes. All of those use narrow anatomical definitions, all require dissemination plus nonconsent plus scienter, and the federal statute carves out matters of public concern and legitimate medical, scientific, or educational purposes. xAI's point: the tools already exist, so HF 1606 isn't necessary.

Item Detail
Case X.AI LLC v. Ellison, 0:26-cv-03425 (D. Minn.)
Filed July 27, 2026
Defendant Keith Ellison, Minnesota AG (official capacity)
Statute challenged HF 1606 / 2026 Chapter 72 / Minn. Stat. § 325E.91
Effective date August 1, 2026
Claim First Amendment — facial and as-applied, plus preliminary and permanent injunction
Civil penalty Up to $500,000 per violation (AG enforcement)
Private remedy Treble compensatory damages, punitive damages, injunction, attorney fees
Legislative votes House 132–1 (2026-04-23), Senate 65–0 (2026-04-29)
Signed Gov. Tim Walz, May 7, 2026
Counsel for xAI Stinson LLP (Minneapolis) + Eimer Stahl LLP

Who wins what — and the question xAI didn't answer

What xAI gets is straightforward. Time, first: a preliminary injunction means Grok Imagine's editing features stay live in Minnesota past August 1. Precedent, second: Minnesota being first means more states are coming, and if "strict liability plus broad definition" gets knocked down here, the drafting template changes everywhere else. Framing, third: "we oppose deepfake porn but this law is wrong" is a genuinely useful position for a company currently fighting on several fronts at once.

What Ellison gets isn't bad either. He's defending a law that passed 132–1 and 65–0. There aren't many easier hills in politics. And his opponent is Elon Musk. For an incumbent AG running for a third term, this lawsuit reads more like an asset than a risk — Walz's one-liner pretty much captures the calculus.

What Minnesota victims get is conditional. If the law survives, the U.S. gets its first channel for holding the tool provider directly liable, which matters because the recurring practical problem has been that identifying the abusive user is often impossible. If the law gets frozen before August 1, the first attempt dies at the first hearing, and legislative momentum in other states cools with it.

Now the uncomfortable part. There's a fact the complaint doesn't confront head-on. In a suit filed March 24, 2026 in the Circuit Court for Baltimore City, the City of Baltimore alleged that Grok generated roughly 3,000,000 sexualized images between December 29, 2025 and January 8, 2026 — about 23,000 of which appear to depict children. Those are allegations, not findings; no court has adopted them. But when xAI writes "we suspended 50,000 accounts and filed 70,000 reports," whether those numbers are proportionate to the problem is an entirely separate question. The "strict liability is unfair" argument depends on "our safeguards actually work." If Baltimore's numbers hold up, that premise wobbles.

There's a second soft spot. The complaint illustrates overbreadth with political satire: an image President Trump posted to his own account on May 1, 2026 showing him, the Vice President, the Interior Secretary, and others shirtless in the National Mall reflecting pool; a sumo-match composite of a former and current governor; and Trump standing on a Chinese street with Xi, Musk, and Tim Cook. The argument lands — is that really worth $500,000? But those three examples also scrub the ugliest use cases off the screen entirely. If Minnesota walks into the hearing with examples pointed the other way, the picture the judge sees changes completely.

Musk's side has fought this fight three times already

This isn't a first attempt, and the scorecard is genuinely mixed.

Start with the loss. In April 2025, X Corp. sued this same Attorney General over Minnesota's 2023 election deepfake law, Minn. Stat. § 609.771 (X Corp. v. Ellison, 0:25-cv-01649). The claims were First and Fourteenth Amendment violations plus preemption under Section 230 of the Communications Decency Act. It went badly. On December 2, 2025, Judge Laura Provinzino held that X Corp. lacked Article III standing to bring a pre-enforcement Section 230 challenge, and stayed the constitutional claims pending the appeal in a parallel case, Kohls v. Ellison. So Musk's side has already been stopped at the threshold in this exact courthouse against this exact opponent. This new complaint looks like it learned the lesson: no Section 230 theory at all, a single First Amendment count, and standing carefully pleaded with case citations across paragraphs 15 through 17.

Now the win. In August 2025, Judge John Mendez of the Eastern District of California struck down California's AB 2655, the Defending Democracy from Deepfake Deception Act of 2024, which had required large platforms to block and label materially deceptive election content. Mendez didn't touch the First Amendment question — he killed the statute purely on Section 230 preemption, and California's own Attorney General ultimately agreed the law conflicted with Section 230. The lesson buried in that outcome: when suits against platform mandates win, they tend to win on federal preemption, not on constitutional grounds. And here's the thing — HF 1606's subdivision 7 says the section "shall be construed consistent with 47 U.S.C. § 230." Minnesota's drafters read the California case and built a wall. Which may be exactly why xAI left Section 230 out.

And the case where the state won. In Free Speech Coalition v. Paxton, decided June 27, 2025, the Supreme Court upheld Texas HB 1181, the age-verification law for adult sites. The whole fight was about the standard of review: Justice Thomas's majority held the law imposes only an incidental burden on adults' protected speech, so intermediate scrutiny applies rather than strict scrutiny — and the law survives it. Justices Kagan, Sotomayor, and Jackson dissented. That case is the real battleground here. xAI says HF 1606 is content-based because enforcing it requires examining what an image depicts. Minnesota can reach for Paxton and argue its law targets the conduct of producing nonconsensual sexual imagery and only incidentally burdens protected expression. Whichever standard the judge picks basically decides the case.

One paragraph to connect this with yesterday's story, because they're separate matters that people will conflate. On July 29 we covered UK Labour MP Jess Asato's action against xAI in the High Court in London — filed June 3, 2026, and amended this month to seek an order compelling xAI to implement permanent technical measures so Grok can no longer manipulate her image. That case runs in the opposite direction from this one. Asato is an individual suing xAI under UK data protection law and misuse of private information, targeting the model's design and deployment. Minnesota is xAI suing a state government under the U.S. First Amendment to void the regulation itself. Same product, two continents, opposing pressure vectors — and xAI is the defendant in one and the plaintiff in the other.

Why OpenAI, Google, and Meta are staying quiet

Everybody else is silent, but the silence has different flavors.

OpenAI and Google have little reason to jump in. Both apply notably more conservative policies to editing photos of real people, and both are known for high refusal rates on prompts aimed at real-person nudity. But here's the cold read: HF 1606 doesn't ask about intent or policy. It asks whether a user succeeded. A company with a 99.9% refusal rate still eats $500,000 of exposure every time the remaining 0.1% gets through. Staying quiet isn't a sign of safety — it's a sign that xAI is about to generate a precedent for the whole industry at its own expense. Win, and everyone wins. Lose, and you respond then.

Meta's position is genuinely awkward. In June 2025 Meta sued the operator behind a nudify app called CrushAI. It has been a plaintiff against this ecosystem. Standing beside xAI now to argue that state regulation goes too far would collide with its own litigation posture. Silence is the only coherent option.

Open-weight model providers sit in this law's strangest blind spot. Subdivision 3 says the prohibitions "do not apply when the website, application, software, program, or other service requires the technical skill of a user to nudify an image or video," and subdivision 1(e) defines technical skill as "substantial application of individualized technological or artistic skill and judgment by a human creator in directing, shaping, or controlling the output." Read that literally and the exemption spares hard-to-use tools while punishing easy ones. Someone running a local Stable Diffusion checkpoint with a custom LoRA may qualify; Grok Imagine, where one sentence does the job, does not. Given where actual harm originates, it's fair to wonder whether that's the right line at all. xAI makes the same point in paragraph 40, calling the exemption something that "does little to narrow the Act's broad reach."

And the safety side. The strongest counterargument in this case probably won't come from the state — it'll come from victims. At the February Senate hearing, women testified about how their mental health, employment, and sense of safety collapsed after a man used their Facebook photos to build pornographic deepfakes. How well "but shirtless men are covered too" holds up against that testimony depends heavily on which judge is listening.

So what actually changes for you

If you build software, this isn't abstract constitutional theater. HF 1606 targets any "person who owns or controls a website, application, software, program, or other service" — no size threshold anywhere. A three-person team wrapping the Grok Imagine API or any other image-editing model is, on the statute's plain text, exposed to $500,000 per incident the moment one Minnesota user produces a covered image. Once August 1 passes, the practical moves are obvious: geo-gate person-photo editing, add face-detection refusals, or block Minnesota traffic outright. Until an injunction lands, that's the conservative choice. And the math doesn't stop at one state — first states are always followed by second and third ones.

If you invest, watch two signals. One is that state-by-state regulatory fragmentation is now a real line item in the cost of shipping AI products in the U.S. Regional feature gating, jurisdiction detection, and legal response all land on the roadmap. The other is the shape of the exposure. xAI's $50 billion figure is a theoretical statutory maximum, not a forecast, but the underlying issue is real: per-violation penalties scale terribly against generative products. When a product mints thousands of images a second, "per unlawful use" fines are arithmetically uncontrollable. That said, nothing is decided yet, and it could break either way.

If you run AI in a company, it's time to reread your vendor terms. Image generation APIs are already embedded in marketing, real estate, and e-commerce workflows — paragraph 22 of the complaint describes exactly those use cases, like a candle brand compositing an autumn background or an agent virtually staging an empty room. Those are mostly safe because no human body is involved. The risk lives in apparel try-on, fit simulation, and any creative with people in it. If your vendor geo-gates a feature, a Minnesota campaign can quietly break. Check whether your contract says anything about regional feature availability.

If you're just a user in Minnesota after August, expect Grok Imagine's editing features to behave differently — xAI said in its own filing that it plans to restrict them, though it never specified which. There's a flip side too: if the law survives, Minnesotans become the first residents in the country who can sue the company that built the tool for treble damages and attorney fees when their photo gets nudified.

Last thing, the calendar. xAI has asked for a preliminary injunction and the statute goes live August 1. That's an extremely tight window. The court could rule within days, or let the law take effect while briefing continues. Either way, expect news in the first week of August.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? If you don't live in Minnesota, almost nothing directly. But if you build or operate anything on top of an image-generation API, the era of region-dependent feature availability has started. When one state goes first, the second one usually arrives fast.

— Will the law actually be blocked on August 1? Too early to call. xAI asked for a preliminary injunction, but no judge is obligated to rule before an effective date. And Musk's side already failed to clear the standing threshold in this same court against this same defendant in December 2025. Narrowing to a single First Amendment count helps, but the outcome is genuinely open.

— If xAI wins, does deepfake regulation collapse? No. What xAI is attacking is the specific design — ban the tool, skip the intent requirement. The laws that do require dissemination, nonconsent, and scienter (Minnesota's § 617.262, the federal TAKE IT DOWN Act) are cited in the complaint as better alternatives. A win here looks less like deregulation and more like a forced rewrite of how these statutes get drafted.

Sources

Numbers are as of announcement and may change.