When the Evidence Is AI-Generated

On April 7, 2026, the United States Attorney for the Southern District of Ohio announced the first federal conviction under the Take It Down Act, the bipartisan federal statute signed by President Trump in May 2025 that criminalizes the non-consensual publication of intimate visual depictions, including AI-generated forgeries. The defendant, James Strahler II of Columbus, pleaded guilty to four counts: two counts of cyberstalking, one count of producing obscene visual representations of the sexual abuse of children under existing federal law, and one count of publishing digital forgeries under the Take It Down Act. He pleaded a second time on July 13, 2026, to an amended information. He was sentenced on September 8, 2026.

The court imposed 180 months on the Section 1466A count, 60 months on each cyberstalking count, and 24 months on the Take It Down Act count, all running concurrently, for an aggregate of fifteen years. The fifteen years came from a statute enacted in 2003. The new one contributed twenty-four months, which is its maximum for a publication offense involving an adult victim. The count that makes the case historic is the count that carried the least exposure in it.

The case has been reported in technical and legal press as a milestone for prosecution of AI-generated material generally. That framing is partly right and partly worth unpacking. The Strahler case actually surfaces three distinct legal regimes each addressing a different category of conduct, each with its own statutory text, its own evidentiary requirements, and its own defense issues. For the criminal defense bar, untangling those regimes is the first step. The technical forensic questions that will run alongside them (how synthetic media is identified, authenticated, and preserved as evidence) are the second.

What the Take It Down Act Actually Criminalizes

The Take It Down Act (formally, the Tools to Address Known Exploitation by Immobilizing Technological Deepfakes on Websites and Networks Act) was authored by Senators Ted Cruz (R-TX) and Amy Klobuchar (D-MN) and passed both chambers of Congress in 2025 with rare unanimity (unanimous consent in the Senate on February 13, 2025, and 409 to 2 in the House on April 28, 2025). President Trump signed it into law on May 19, 2025. It is codified at 47 U.S.C. § 223(h) and § 223a.

The Act has two main operative pieces:

First, a criminal prohibition on the non-consensual publication of intimate visual depictions. The Act creates federal criminal liability for knowingly publishing intimate visual depictions of an identifiable individual without the depicted individual’s consent. The Act’s reach includes both authentic intimate imagery and “digital forgeries,” which the statute defines as an intimate visual depiction of an identifiable individual created through software, machine learning, artificial intelligence, or other computer-generated means, including by adapting or altering an authentic depiction, that a reasonable person viewing it as a whole would find indistinguishable from an authentic depiction of that individual. The penalties are tiered four ways:

• Publishing a depiction of an adult carries not more than two years
• Publishing a depiction of a minor carries not more than three
• Threatening to publish carries eighteen months where the depiction is of an adult
• Threatening to publish carries thirty months where the depiction is of a minor

The Strahler conviction is, specifically, a conviction under the digital-forgeries provision.

Second, a notice-and-takedown regime for covered platforms. The Act requires online platforms that allow user-uploaded content to establish a process for identifiable individuals to request removal of non-consensual intimate visual depictions, and to act on those requests within 48 hours, along with reasonable efforts to find and remove identical copies. Platforms had one year from enactment (until May 19, 2026) to have that process in place. The duty is administrative rather than criminal and is enforced by the Federal Trade Commission as an unfair or deceptive act or practice.

What the Take It Down Act does not do is criminalize all AI-generated sexual content, all synthetic media generally, or all CSAM in synthetic form. Each of those areas is governed by different (and in some cases older) statutes. Conflating the Act with the broader synthetic-media regulatory landscape is one of the more common errors in the press coverage and in the lay legal commentary that has followed the Strahler plea.

The Other Statute in the Strahler Case: 18 U.S.C. § 1466A

Of the counts Strahler pleaded guilty to, the most consequential for the wider defense bar is not the Take It Down Act count. It is this one: producing obscene visual representations of child sexual abuse under 18 U.S.C. § 1466A.

Section 1466A is a federal statute enacted in 2003 as part of the PROTECT Act. It criminalizes the production, distribution, receipt, and possession of obscene visual representations of the sexual abuse of children including representations created without the use of any real child, such as drawings, cartoons, paintings, sculptures, and, of particular contemporary relevance, computer-generated and AI-generated material. The statute applies whether the visual representation depicts an identifiable real minor or a wholly synthetic minor. The mens rea is knowledge.

The statute offers two alternative routes, not a single combined test. The first requires that the material depict a minor engaged in sexually explicit conduct and be obscene, which means the full Miller inquiry. The second requires that it depict certain enumerated conduct and lack serious literary, artistic, political, or scientific value, which is the third Miller prong standing alone. A charge under the second route does not require the government to prove obscenity at all. Subsection (c) then removes the remaining question by providing that it is not a required element that the minor depicted actually exist.

Section 1466A exists because of a constitutional decision the United States Supreme Court reached in 2002. In Ashcroft v. Free Speech Coalition, 535 U.S. 234 (2002), the Court struck down portions of the Child Pornography Prevention Act of 1996 that had criminalized purely virtual child-pornography material (imagery that appeared to depict minors engaged in sexual conduct but that had been created without the use of any actual minor). The Court held that the statutory prohibition swept too broadly into protected speech and could not be sustained on the theory that the material itself was (by definition) unprotected child pornography, because the constitutional basis for the categorical exclusion, the harm to real children, was absent in cases involving purely virtual content.

The Court struck the definition reaching material that appears to be a minor and the definition reaching material conveying the impression that it depicts a minor. It left the real-child definition standing. And it expressly declined to decide anything at all about the morphing provision, noting that the respondents had not challenged it and that the Court would not consider it. That reservation is the hinge of everything below.

Congress’s response was § 1466A, which uses a different constitutional theory: obscenity. Under longstanding First Amendment doctrine, obscene material is not constitutionally protected. The PROTECT Act’s drafters concluded that visual representations of the sexual abuse of children (even where no real child was involved in production) could be prosecuted under obscenity principles. Lower federal courts have generally upheld § 1466A against constitutional challenge, applying the three-part obscenity test articulated in Miller v. California, 413 U.S. 15 (1973). The leading appellate decision is United States v. Whorley, 550 F.3d 326 (4th Cir. 2008), in which the Fourth Circuit affirmed a § 1466A conviction involving obscene Japanese animation depictions of minors and held that obscenity is unprotected speech regardless of whether real children were involved in production. The Supreme Court denied certiorari in Whorley and has not, in the years since, squarely revisited the Ashcroft holding in the AI-generation era. Two cautions on that case. The Fourth Circuit denied rehearing en banc, with Judge Gregory dissenting from the denial and urging a certiorari petition, and the conviction there was under the obscenity route, so the decision is not authority on the second one.

The Strahler plea involved both real and synthetic depictions of minors. As reported by the Department of Justice, the AI-generated material in the case involved morphed imagery, meaning the faces of known minor children placed on other bodies. The usual telling treats that as the harder case doctrinally because identifiable real children are involved. For the government it is the easier one, and the distinction is worth understanding before you build a defense on the wrong half of it.

Morphed images fall inside the federal definition of child pornography at 18 U.S.C. § 2256(8)(C), which is the provision Free Speech Coalition expressly left alone. They can therefore be charged as child pornography under § 2252A, with no obscenity element at all, and Section 1466A is not the only route available to a prosecutor. Every federal court of appeals to reach the question has sustained the prosecution.

The rationales do not line up, though, and the fault line is whose body was used. United States v. Bach, 400 F.3d 622 (8th Cir. 2005), involved a real minor’s body and is the strongest of them. United States v. Hotaling, 634 F.3d 725 (2d Cir. 2011), involved adult bodies and held the images unprotected anyway. United States v. Anderson, 759 F.3d 891 (8th Cir. 2014), is the one to read first, because the Eighth Circuit declined to extend the categorical rationale of Bach after United States v. Stevens and upheld the conviction only under as-applied strict scrutiny. Shoemaker v. Taylor, 730 F.3d 778 (9th Cir. 2013), reads stronger than it is, because it comes up through habeas and holds only that the Supreme Court has not clearly established protection. And there is contrary state authority: State v. Zidel, 156 N.H. 684 (2008), reversed a possession conviction on First Amendment grounds where the bodies were adult and nothing had been distributed.

A purely synthetic case, imagery generated by AI that does not depict any identifiable real child, raises the Ashcroft questions in cleaner form. That case has not reached the Supreme Court. Last month it reached a court of appeals.

The Seventh Circuit Just Narrowed Section 1466A

In United States v. Anderegg, No. 25-1354 (7th Cir. Aug. 25, 2026), the Seventh Circuit held § 1466A(b)(1) unconstitutional as applied to private possession of AI-generated material in the home. The court assumed the material was obscene and held that Stanley v. Georgia, 394 U.S. 557 (1969), still protects mere possession at home, and that the statute’s interstate commerce element did not change the answer, because Stanley is about where possession occurs rather than how the material arrived.

The production and distribution counts under § 1466A(a) survived and were not disturbed. Nothing in the opinion reaches material depicting an identifiable real child, and nothing in it protects distribution. Judge Lee, joined by Judge Kolar, wrote separately that the panel was bound by Free Speech Coalition and urged the Supreme Court to revisit it in light of generative AI. The useful reading is narrow and real: in the Seventh Circuit, a home-possession count under the second subsection now has a live constitutional defense that a production count does not.

The Cyberstalking Counts

The cyberstalking counts in the Strahler case (under 18 U.S.C. § 2261A) rest on an older federal statute that does not require any AI element. Section 2261A prohibits using an interactive computer service or electronic communications system to engage in a course of conduct that places a person in reasonable fear of death or serious bodily injury, or that causes substantial emotional distress, where the conduct involves multiple states or jurisdictions. It is the statute the Justice Department often uses as the criminal backbone for cyberstalking and harassment cases that cross state lines.

Those counts are not legally novel. What makes them relevant in a synthetic-media case is the way AI-generated material can supply the underlying course of conduct (i.e. the harassing communications, the threats, the publication of fabricated imagery directed at the victim, her family, her employer). The synthetic media is not, in that posture, the offense element itself. It is the instrumentality of an older offense.

The Defense Issues This New Landscape Surfaces

For criminal defense lawyers, the synthetic-media era raises a small set of issues that recur across these cases, regardless of which charging statute the government has chosen. Five are worth attention now.

• Authentication of the imagery itself. Under Federal Rule of Evidence 901 (and Texas Rule of Evidence 901), the proponent of a piece of evidence must produce evidence sufficient to support a finding that the item is what the proponent claims it is. For a digital image alleged to be an AI-generated forgery of a real person, the State must establish both that the image exists and that it is, in fact, AI-generated. For a digital image alleged to be authentic CSAM, the State must establish that it depicts a real minor. The two questions look superficially similar but require very different evidentiary showings, and the consequences of mistakes run in both directions.
• Chain of custody for generative-AI outputs. Synthetic media has a forensic property traditional evidence does not: the same input prompt produces different outputs each time it is run, and most consumer-facing generative-AI systems do not retain the precise model weights, parameters, or random seeds that produced any given output. Reconstructing the production process from the output alone is, in many cases, technically impossible. That asymmetry between the certainty of the final image and the uncertainty of how it was produced is a real evidentiary problem the State will need to address in some cases.
• Mens rea, and specifically knowledge of the AI nature of the content. The Take It Down Act’s criminal provisions are knowledge-based. The statute reaches knowing publication of non-consensual intimate visual depictions, including knowing publication of digital forgeries. A defendant who actually believed the material was authentic, or who actually believed the depicted person had consented, raises a knowledge defense that is conceptually clean and factually complicated. Section 1466A similarly requires knowing conduct. The contours of “knowing” in a synthetic-media context are not yet well developed in the case law.
• Statutory construction. The Take It Down Act is barely a year old. Its key terms (“intimate visual depiction,” “digital forgery,” “identifiable individual,” “publishing”) are defined in the Act, but the application of those defined terms to specific fact patterns will be litigated for years. Defense lawyers handling early cases under the Act will be doing first-impression statutory construction, and the early appellate decisions will shape how the statute is read for the next generation of cases.
• The constitutional questions. Ashcroft v. Free Speech Coalition is still the controlling Supreme Court authority on purely synthetic depictions of minors. Section 1466A relies on obscenity doctrine to avoid the Ashcroft problem, and lower courts have generally accepted that theory. But the doctrine has not been tested at the Supreme Court level in the modern generative-AI era, where the realism of synthetic imagery and the volume that can be produced are categorically different from anything in the case law. Constitutional challenges to specific applications of § 1466A are no longer merely likely. One succeeded in the Seventh Circuit in August 2026 on a home-possession count, and the reasoning invites more.

None of these defense issues is a shortcut around the underlying offense. None is a strategy for evading prosecution in cases where the conduct is what the State says it is. They are the ordinary work of testing the government’s proof on every element the statute requires which is the work the Sixth Amendment guarantees. That work is harder, not easier, in a synthetic-media case, because the technology is newer than the precedent.

What the Digital Forensics Community Is Working On

Synthetic-media identification is an active research area in academic computer science, federal law enforcement, and private digital-forensics practice. The tools developed in those settings are the tools that will, increasingly, be offered as expert evidence in criminal cases.

Several techniques are in use or under development:

• Provenance metadata and watermarking. Some generative-AI systems embed metadata or invisible watermarks in their outputs that identify the producing system. Industry standards organizations have proposed standardized provenance metadata frameworks. The existence (or absence) of expected provenance markers in a piece of evidence is a fact a forensic examiner can speak to.
• Statistical artifact analysis. AI-generated imagery often contains subtle statistical artifacts (ex. patterns in pixel-level noise, frequency-domain signatures, anatomical or geometric inconsistencies) that distinguish it from camera-captured imagery. These signatures vary by generation model and have been the subject of substantial published research. They are also a moving target as the technology improves.
• Reverse-image and source-image matching. Where a synthetic image has been derived from a real source image (a common pattern in face-morphing cases), the source can sometimes be identified through reverse-image-search tools and similarity-matching algorithms. The link between the synthetic output and the real source is itself a fact, and a forensically demonstrable link can be both inculpatory and (in the right case) exculpatory.
• Device and platform forensics. Many AI-generation cases are not actually built on analysis of the imagery itself. They are built on analysis of the defendant’s devices including installed applications, browser history, model files, prompt logs, generation histories, and the kinds of indicia of intentional use that have been the bread and butter of digital forensics for two decades. The Strahler case, per the Justice Department’s public statements, involved evidence of more than two dozen AI platforms installed on the defendant’s phone and more than 100 web-based models accessed from it.

Each of these is also a defense issue. Expert testimony based on emerging detection techniques is subject to admissibility scrutiny under Federal Rule of Evidence 702 (and Texas Rule of Evidence 702, with the Texas Court of Criminal Appeals’ Kelly framework). The validation of the technique, the error rate, the qualifications of the examiner, the chain from the device to the analysis, and the limits of the conclusion the examiner is willing to draw are all properly tested on cross-examination and, where appropriate, in pretrial motions. The authentication fight itself, and the Texas standard that governs it, is covered at length in our post on deepfake evidence and is not repeated here.

Texas Law and the State-Court Picture

Federal law is not the whole story. Texas has its own statutes that intersect with the synthetic-media space, and most prosecutions of conduct involving Texas victims, Texas defendants, or Texas-based platforms will involve some combination of state and federal exposure.

Texas moved hard in 2025, and the state picture is no longer the one this case was charged against. Four statutes matter. Penal Code § 21.16 is the closest Texas analog to the federal publication provision and it predates the federal Act. It is a state jail felony, and it covers conduct involving Texas victims even where the federal jurisdictional hook is missing. Its definition of visual material was amended effective September 1, 2025, to include a file in any digital format. One caution if you go looking for the case law: the Court of Criminal Appeals upheld subsection (b) against a First Amendment challenge in Ex parte Jones in 2021, but that opinion is unpublished, which means it cannot be cited as authority, and it construed the pre-2017 text rather than the statute as it now reads.

Penal Code § 21.165 is the Texas deepfake statute, and Senate Bill 441 rewrote it effective September 1, 2025. The heading now reads unlawful production or distribution of certain sexually explicit media, not videos. The old intent-to-deceive element is gone. A threat offense was added. Consent now requires a knowing, voluntary, plain-language written agreement, and a disclaimer that the media is not authentic is expressly not a defense. The base offense is a Class A misdemeanor, rising to a third-degree felony on a prior conviction under the section or where the person appearing to be depicted is younger than eighteen, and restitution for psychological, financial, or reputational harm is mandatory. Any research older than September 2025 is describing a different statute.

Penal Code § 43.26 was restructured the same day by Senate Bill 1621. The change that matters here is new subsection (a-2), which reaches a depiction of a computer-generated child, defined as one that appears to be under eighteen, was created using artificial intelligence or other software, and is virtually indistinguishable from an actual child to a reasonable person. There is no obscenity element in it. The same bill created a rebuttable presumption that a depiction is of an actual child rather than a computer-generated one, which puts the burden of raising the synthetic question on the defense and makes the forensic work described above load-bearing rather than optional.

The fourth statute is brand new and it is the one that maps most directly onto the federal case in this post. Senate Bill 20 created Penal Code § 43.235, effective September 1, 2025, covering possession, promotion, or production of obscene visual material appearing to depict a child, and it says in terms that it applies regardless of whether the depiction is an image of an actual child, a cartoon or animation, or an image created using an artificial intelligence application or other computer software. The same section makes it a separate offense to use an image of an actual child to train an artificial intelligence model to produce child pornography. The base offense is a state jail felony, rising with prior convictions under the chapter.

Notice the drafting choice, because it is the same one Congress made in 2003. Section 43.235 carries an obscenity element and § 43.26(a-2) does not. The obscenity element is what places § 43.235 inside Miller and outside the Free Speech Coalition problem. The computer-generated-child provision does not make that move, and it is the provision most likely to draw a constitutional challenge in Texas.

A practical warning on § 43.26. Two bills amended it with the same effective date, Senate Bill 1621 and House Bill 1778, and they conflict. Senate Bill 1621 was enacted later and should control, but the free online codes have not caught up. Do not quote punishment tiers from this section without pulling the current official text.

Texas also got its own notice-and-removal remedy. Senate Bill 441 added § 98B.0022 to the Civil Practice and Remedies Code, effective September 1, 2025, making the owner of a website or application liable to the depicted person for failing to remove artificial intimate visual material within seventy-two hours of a request and to make reasonable efforts to find and remove identical copies. A violation is a deceptive trade practice, and the Attorney General may sue repeat violators. It is narrower than the federal provision, which reaches authentic imagery as well, and slower, seventy-two hours against forty-eight.

For Texas defendants, the practical reality is that conduct of the kind alleged in the Strahler case would, in most circumstances, expose the defendant to both federal charges under the statutes discussed above and state charges under §§ 21.16, 21.165, 43.235, and 43.26. Which sovereign actually charges, and in which order, depends on the facts. Federal indictments tend to follow cross-jurisdictional conduct, large volumes of material, and use of interstate communications systems. State indictments tend to follow conduct with primarily local effects and identifiable Texas victims. The dual-sovereignty exposure is real and is, in many cases, the most important strategic feature of the case for the defense to understand at the front end.

The Takeaway

The Strahler case is the first conviction under a new federal statute, and the sentence is the part worth remembering. Fifteen years, and almost none of it from the new statute. It is an early data point on how synthetic-media prosecutions will be charged, defended, and proved, and it is not the last. The defense bar is going to be confronted with these cases in growing numbers over the coming years, and the lawyers handling them will need to be conversant in three things: the statutory architecture, federal and Texas, and which piece of it actually carries the exposure; the constitutional doctrine, which moved in the Seventh Circuit last month and which Texas rebuilt a year ago this month; and the digital-forensics methods that are increasingly the spine of the State’s evidentiary case.

None of that work is easy. None of it is a substitute for the careful, methodical, fact-specific defense any serious criminal case requires. But understanding the framework is the foundation. Cases of this kind are going to keep coming, and the early cases will set the precedent that governs later ones.

If you or someone you know is under investigation or facing charges involving synthetic media, AI-generated content, or the federal or Texas statutes discussed above, the right time to retain counsel familiar with both the digital-forensics evidence and the underlying statutory framework is at the front end of the case, not after the indictment lands.

Douglas E. Huff is a partner at Deandra Grant Law, an ACS-CHAL Forensic Lawyer-Scientist, past president of the Dallas Criminal Defense Lawyers Association, and a defense lawyer with substantial training in digital forensics. The firm handles criminal cases involving digital evidence, computer forensics, and online conduct throughout North and Central Texas. To discuss a pending investigation or charge in confidence, call (214) 225-7117.