The Argument Nobody Made

On 25 August 2026 the United States Court of Appeals for the Seventh Circuit decided United States v. Anderegg, No. 25-1354. Within two days the headlines had settled into a shape: a federal court has ruled that AI-generated child sexual abuse material is protected by the First Amendment.

That is not what the opinion says. What it says is narrower, and the narrowness matters, because the part everyone skipped is the part that could change the next case.

We are going to walk the opinion. We are also going to tell you where we got it wrong ourselves before we read it, because we did, twice, and this record exists so that our mistakes are as visible as anyone else's.

What actually happened

Steven Anderegg was charged in the Western District of Wisconsin, before Chief Judge James D. Peterson, on four counts: producing, distributing, and possessing visual depictions of minors engaged in sexually explicit conduct, and transferring such material to a minor under the age of sixteen. According to the government he produced the images using generative software.

He moved to dismiss the whole indictment. The district court granted the motion as to the possession charge only and denied the rest. The government appealed that one dismissal. The Seventh Circuit affirmed it.

So the count sheet after 25 August reads: production, still charged. Distribution, still charged. Transfer to a minor under sixteen, still charged. Possession in his own home, dismissed as unconstitutional as applied under 18 U.S.C. § 1466A(b)(1).

One count of four. The other three are pending. A person reading only the headline would not know that making this material and sending it to a child remain federal crimes that this defendant is still facing.

Why the one count fell

Three cases do the work, and they are easier to follow than their reputations suggest.

Stanley v. Georgia (1969) held that the government cannot criminalize the private possession of obscene material in your own home. The holding turns on the location, not the content.

New York v. Ferber (1982) held that child pornography is its own category of speech, outside First Amendment protection entirely. Read how the Court got there, because everything downstream depends on it. Ferber is built on harm to the actual children in the images. The image is a record of a crime against a real child, and the state's interest is in those children.

Ashcroft v. Free Speech Coalition (2002) took that reasoning to its end. Remove the real child, and the Ferber rationale goes with it. Material that depicts no actual child is not in the Ferber category, though it can still be obscene.

Stack those and the result follows mechanically. Anderegg was charged with possessing, at home, material that is obscene but depicts no actual child. Obscene material at home is Stanley. So the count fell.

The panel did not enjoy this. In the majority opinion, Judge Lee wrote:

Given the relentless advancement in artificial intelligence models, we have some concerns about the lines these cases draw, but we are not free to redraw them ourselves.

That is a court saying it thinks the line is in the wrong place and that moving it is not its job.

The door we walked into, and the correction we owe

There is a fourth case, and it is the one that looks like the way out.

Osborne v. Ohio (1990) held that Stanley does not protect the private possession of child pornography at home. Ohio criminalized simple possession, Osborne argued Stanley, and the Supreme Court disagreed. The distinction it drew is that Georgia in Stanley had been trying to police what a man thought in his own house, while Ohio was trying to destroy a market and the material itself, and that interest reaches inside the home.

Reading that, the transfer looks obvious. Destroying a market does not require an identifiable victim. A market in synthetic material is still a market.

We made that argument internally on 1 September, with some confidence, before reading the opinion. It is wrong, and the Seventh Circuit had already explained why a week before we said it:

Because the images for which Anderegg is charged do not depict an actual child, Ferber and Osborne do not directly apply to this case.

Osborne is anchored in the participants. As the court put it elsewhere in the opinion, Osborne rests on concern for those it called the victims of child pornography. Where there is no such victim, the anchor does not hold. The door was already closed, and we had recommended walking through it.

That is correction one. There are two more at the end.

The part almost nobody has reported

Judge Lee wrote the majority. Then Judge Lee wrote a separate concurrence to his own opinion, joined by Judge Kolar.

Judges do not do that idly. A concurrence to your own majority is a signal, and the audience is usually one court up. Lee quotes Justice O'Connor and Justice Thomas from Free Speech Coalition in 2002, both of whom warned that technology would eventually make virtual images indistinguishable from real ones and that the day might come when the government's interest changed. Lee writes:

According to some experts, that day may have arrived.

Then comes the passage that we think is the actual story of this case:

What is more, AI-generated CSAM may also present other challenges not raised by the government in this appeal. For example, researchers have found that "a major dataset used for training image-generating AI contained hundreds of CSAM images."

Thus, unlike the virtual CSAM at issue in Free Speech Coalition, the market for AI-generated virtual CSAM images may impact the demand for actual CSAM on which the AI models can train. Such a finding would implicate many of the concerns raised in Osborne.

Read that slowly.

The judge who wrote the opinion striking the count is saying, in print, in the Federal Reporter, that there was an argument available which might have put this material back inside Osborne, and that the government did not make it.

Ashcroft rests on a factual premise: no real child was involved. If the model learned what it knows partly from real children's abuse imagery, that premise is not a legal question at all. It is an evidentiary one, and it can be litigated without asking any court to overrule anything.

The premise is checkable. In December 2023 the Stanford Internet Observatory reported finding thousands of suspected abuse images inside LAION-5B, the dataset behind widely used image generators, and verified more than a thousand of them using PhotoDNA. LAION took the datasets down and published a cleaned version, Re-LAION-5B, in August 2024. That cleanup matters and it cuts both ways: models trained after it may be clean, and models trained before it are still in circulation and cannot be untrained.

We want to be careful here, because the discipline applies to arguments we like. This is not a proven fact about the specific model in this specific case. It is an argument with a factual burden attached, and somebody would have to carry it. The point is that the burden was never taken up. The court noticed. The court said so.

The bill everyone is pointing at

The recommendation we have seen repeated, and which we ourselves made before checking, is to call the House about the ENFORCE Act.

So we read it. Not a summary of it. The engrossed text as the Senate passed it on 16 December 2025, S.3021. The House version is H.R.4831, introduced by Rep. Ann Wagner with Rep. Steve Cohen. Here is every operative provision:

• Section 2 amends 18 U.S.C. § 2252A(a)(7), which concerns production of material covered by § 2256(8)(C), and adjusts the interstate commerce hooks.
• Section 3(a) removes the statute of limitations for § 1466A offenses.
• Section 3(b) adds § 1466A to sex offender registration under the Adam Walsh Act.
• Section 3(c) keeps the depictions in government or court custody during proceedings and gives an identifiable minor the same access rights a victim has.
• Section 3(d) creates a presumption of pretrial detention for offenses under § 1466A(a).
• Section 3(e) adds § 1466A to the supervised release provision.

Every one of those is worth having. Not one of them touches § 1466A(b).

Section 1466A(b) is the possession subsection. It is the exact provision the Seventh Circuit held unconstitutional as applied. The bill amends subsection (f), and references subsection (a) by name in the detention provision, and never reaches subsection (b) at all.

And it could not have fixed it by trying. Stanley is a constitutional holding. A statute cannot overrule one. Closing that particular hole takes either the Supreme Court revisiting Free Speech Coalition, which Judge Lee just went out of his way to invite, or a prosecutor proving the training-data premise false in a real case.

The bill has passed the Senate and is held at the desk in the House, which means it needs a floor vote and nothing else. It should get one. It is also not the answer to this case, and saying that it is has the effect of telling people the problem is being handled when the specific hole is not.

The practical thing

If you take one action item from this, it is not a phone call about a bill that does not reach the provision.

It is this. The next United States Attorney who charges one of these cases should put the training data in the brief. A federal appellate judge has now written, in a published opinion, that this argument exists, that it may implicate Osborne, and that the government did not raise it. That is about as close to an engraved invitation as an appellate court gets.

And if you write about this case, the honest summary is not that a court blessed the material. It is that four counts were charged, three survive, one fell on a doctrine from 1969, and the judge who applied that doctrine wrote separately to say he thinks the Supreme Court should look at it again and that the government left its best argument on the table.

Our corrections on this piece

Three, all logged in the errata record with dates and how each was found.

We forecast that Congress was likely to move quickly after this ruling. It had already passed the Senate unanimously in December 2025 and had been sitting at the House desk ever since. The forecast was disputed on the spot by the person we said it to, and he was right.

We argued that Osborne transferred cleanly to synthetic material and was the strongest argument available. The panel had considered and rejected exactly that, and we had not read the opinion when we said it.

We recommended the ENFORCE Act as the remedy for this specific hole. It is not, and reading the engrossed text rather than its summaries is what showed us that.

All three ran in our own favour, in the sense that each made our analysis sound more confident and more actionable than the record supported. That is the direction errors usually run, which is why we log the direction.