First AI forgery indictment in the United States: New York congressional candidate arrested for campaign deepfakes

First indictment for AI forgery in the United States: New York candidate arrested for campaign deepfakes

🔎 A political candidate charged for the first time in the United States

On June 24, 2026, Jonathan Rinaldi, a former candidate for the State Assembly in Queens (New York), was arrested at his home and charged with forgery and criminal possession of a forged instrument. The reason: he used generative AI to create fake political endorsements, fabricated press articles, and a deepfake of his opponent Andrew Hevesi, then disseminated them on social media during his campaign.

This is a major legal precedent. Until now, political deepfake cases in the United States involved consultants or third parties. Here, it is the candidate himself who is facing criminal prosecution. The US justice system is entering a new phase: the direct criminal crackdown on the algorithmic manipulation of elections.

The case takes on a particular resonance in the context of the chip war: the United States authorizes the sale of H200 to 10 Chinese companies, but Beijing blocks deliveries. The same AI models that fuel geopolitical competition are ending up in the hands of a local candidate to deceive voters. AI is no longer a foreign policy issue. It has become a campaign tool.

The essentials

• Jonathan Rinaldi, a Queens State Assembly candidate, was arrested on June 24, 2026, and charged with forgery for using AI to fabricate fake endorsements and a deepfake of his opponent.
• This is the first time an American political candidate has faced direct criminal charges for the unfair use of AI in a campaign.
• New York recently tightened its legal framework with the AI and Social Media Protections Act (April 2026), while the EU AI Act imposes mandatory labeling starting August 2, 2026.
• This case sets a legal precedent and raises the central question of the First Amendment: where does political free speech stop when AI makes deception undetectable?

Recommended tools

The facts: what Rinaldi actually did

A falsification scheme spanning several months

The investigation by Queens District Attorney Melinda Katz revealed a falsification scheme stretching from October 2025 to June 2026. Rinaldi did not make an isolated mistake. He built a structured disinformation operation.

As early as October 2025, according to the Queens Eagle, Rinaldi doubled down on AI-generated videos and doctored photos to deceive voters in the 28th district. In April 2026, a few days after being accused of fraud in a separate context, he posted a deepfake of his opponent Andrew Hevesi. The audio and video were AI-generated but presented as authentic on his campaign channels.

On June 23, 2026, primary day, voters in the Forest Hills neighborhood were exposed to these falsified contents just before casting their ballots. The arrest took place the following day.

The specific charges

The Queens DA charged Rinaldi with two counts: forgery and criminal possession of a forged instrument. These charges do not fall under AI-specific laws, but rather existing offenses under New York criminal law adapted to a new technological context.

The prosecutor's strategy is clear: rather than invoking a deepfake law that could be challenged on First Amendment grounds, Katz is using general document forgery offenses. A fake AI-generated endorsement is treated as a forged document, in the same way as a forged check. This approach bypasses the constitutional hurdle.

New York's legal framework: a recently strengthened arsenal

The AI and Social Media Protections Act

New York is not starting from scratch. In April 2026, Governor Kathy Hochul signed the AI and Social Media Protections Act, a set of measures specifically targeting political deepfakes. This law prohibits the distribution of AI-generated images depicting candidates without explicit mention of their synthetic nature.

The context of this law is directly linked to the gubernatorial election campaign. In January 2026, the New York Times reported that the Cuomo campaign had released an AI-generated video showing candidate Zohran Mamdani eating rice with his hands — an image designed to stir up ethnic prejudice. This incident accelerated the adoption of the text.

New York Attorney General Letitia James also published a detailed guide on protecting voters from AI-driven electoral disinformation, listing possible remedies and warning signs to watch out for.

The common law approach: more robust than specific laws

The legal genius of the Rinaldi case is the use of the forgery offense rather than a specific AI law. According to Ballotpedia, the State of New York has accumulated several texts on deepfakes since 2019, but their constitutionality remained uncertain in the face of the First Amendment.

Forgery, on the other hand, is a recognized offense for centuries. The New York Supreme Court has already established that the falsification of documents for the purpose of deception is not protected by the First Amendment. By applying this framework to AI-generated content, the prosecutor creates a precedent that does not depend on the constitutional survival of a specific deepfake law.

This is a crucial distinction. Anti-deepfake laws passed in various states risk being struck down by federal courts. Forgery does not.

Federal precedent: from Louisiana to Queens

The Biden robocalls case (2024)

The Rinaldi case is not the first time AI has generated federal election-related prosecutions. In May 2024, Reuters reported the indictment of a Louisiana political consultant for creating robocalls imitating Joe Biden's voice. The goal: to deter voters from participating in the New Hampshire Democratic primary.

This consultant, Paul Carpenter, was indicted for conspiracy to commit wire fraud and voter suppression. But he was a third party, not a candidate. The difference in status changes everything in terms of liability and deterrence.

The Dao Yin case: wire fraud and fake AI donations

On the same day as Rinaldi's arrest, the DOJ announced charges against Dao Yin, a political operative from Flushing (Queens), for wire fraud. According to the DOJ press release, Yin allegedly used AI to generate fake donations and false campaign finance reports.

These two simultaneous cases in the same New York borough are probably not a coincidence. The Eastern District of New York (EDNY) is sending a clear signal: the era of impunity for the algorithmic manipulation of campaigns is over.

However, G7 Évian : Altman, Amodei et Hassabis réunis pour la première fois au sommet — et les États-Unis bloquent toute gouvernance contraignante shows that AI leaders are still hesitating over regulatory frameworks. On the ground, prosecutors are not wasting any time.

The First Amendment battle: freedom of speech vs electoral deception

Why the First Amendment is the real battleground

Rinaldi's defense will almost certainly rely on the First Amendment. The argument is predictable: false statements in politics are protected by the Constitution, as long as they do not cause immediate and specific harm (the Brandenburg v. Ohio, 1969 standard).

The Supreme Court has historically been very protective of political speech, even when false. In United States v. Alvarez (2012), the Court struck down a federal law criminalizing false claims of having received a military medal, on the grounds that the First Amendment protects even lies.

But forgery is different. The Court has always distinguished between false verbal statements (protected) and the falsification of documents (unprotected). A fake endorsement is not a false opinion: it is a counterfeit document. This is exactly the line the Queens DA has chosen to draw.

Do AI models make the Brandenburg standard obsolete?

The fundamental problem is this: the Brandenburg standard requires "imminent and likely" harm. But a deepfake circulated on social media can remain in circulation indefinitely, be seen by thousands of people, and be impossible to completely remove. The harm is no longer imminent: it is permanent.

Models like GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro make the creation of deepfakes undetectable to the naked eye. The NVIDIA Nemotron 3 Ultra 550B : le modèle open-source le plus puissant des États-Unis débarque au Computex illustrates this democratization of synthesis power. When anyone can generate a flawless false political statement, the American legal framework inherited from the 1960s becomes inadequate.

The European angle: the EU AI Act and mandatory labeling

August 2, 2026: a deadline that changes everything

While the United States is discovering criminal prosecution, Europe is moving forward through transparency. The EU AI Act imposes mandatory labeling of all AI-generated content starting August 2, 2026. Platforms will have to integrate mechanisms for detecting and marking deepfakes.

The approach is fundamentally different. Europe does not criminalize the creation of deepfakes (except in specific cases such as child pornography). It requires that the consumer be informed. The responsibility lies with the platforms, not the end user.

How effective against a determined candidate?

Would the European approach have prevented the Rinaldi affair? Not necessarily. A determined candidate can strip the marking metadata, use self-hosted open-source models, and publish on decentralized channels. Labeling is a safety net, not an absolute bulwark.

The combination of both approaches — prevention through labeling (Europe) and deterrence through criminal law (United States) — will likely emerge as the global standard. But for now, each jurisdiction is moving at its own pace, and malicious candidates are exploiting the gaps.

Comparison table of legal frameworks

Implications for the 2026 midterm elections

A real — but limited — deterrent effect

Rinaldi's arrest comes exactly four months before the November 2026 midterm elections. The timing is not coincidental. Federal and local prosecutors are sending a warning to all candidates: AI does not give you a blank check.

But deterrence has its limits. Rinaldi was a marginal candidate, nicknamed "The Sperminator" in New York political circles according to the NY Post, with no real hope of winning. A serious candidate with solid legal resources could have designed forgery schemes that are harder to trace, using intermediaries or self-hosted models like Kimi K2.6 or GLM-5.1.

The detection issue

Current deepfake detection models struggle against the most recent generations. An audio deepfake produced by GPT-5.5 or Claude Opus 4.7 is virtually undetectable without the original metadata. Detection therefore relies less on technical analysis than on human investigation: who published it, when, and with what sources?

This is precisely what made Rinaldi's arrest possible. It was not an algorithm that spotted him, but journalists from the Queens Eagle and citizens who reported inconsistencies. Technology alone will not be enough. Civic vigilance remains the first line of defense.

❌ Common mistakes

Mistake 1: Confusing false political statements with forgery

Many commentators have equated the Rinaldi case with simple "campaign lies". This is inaccurate. Forgery involves the creation of a false document or instrument — a fake press article, a signed fake endorsement. The distinction is crucial because verbal lies are protected by the First Amendment, forged documents are not.

Mistake 2: Thinking that anti-deepfake laws are enough

New York's AI and Social Media Protections Act is a good political signal, but its constitutional solidity is doubtful. The real precedent of the Rinaldi case is the application of classic criminal law to AI content. Lawmakers who believe that an "anti-deepfake" law will be enough are mistaken.

Mistake 3: Ignoring the role of open-source models

Analyses focus on commercial models (OpenAI, Anthropic, Google). But models like DeepSeek V4 Pro or Kimi K2.6, often self-hosted, offer comparable synthesis capabilities without any traceability mechanism. Regulating commercial models only covers part of the problem.

❓ Frequently Asked Questions

What are Rinaldi's exact charges?

Forgery and criminal possession of a forged instrument, under New York State Penal Law. These are felonies punishable by several years in prison. The DA chose these charges rather than a specific AI law to avoid First Amendment challenges.

Why is this a major precedent?

This is the first time an American political candidate has been criminally prosecuted for the unfair use of AI in a campaign. Previous cases (the 2024 Biden robocalls) involved third parties, not candidates directly involved.

Would the EU AI Act have prevented this case?

Not necessarily. Mandatory labeling (August 2026) requires transparency for AI content, but a malicious actor can bypass these measures by removing metadata or using non-compliant models. The European approach is preventive, not punitive.

Which AI model did Rinaldi use?

The investigation has not publicly identified the exact model. The falsified content included images, audio, and text, suggesting the combined use of several generation tools. Current models like GPT-5.5, Claude Opus 4.7 or Gemini 3.1 Pro make this type of falsification accessible to anyone.

What is he actually risking?

Penalties for forgery in New York State go up to 7 years in prison for the most serious cases. Possession of a forged instrument is also a felony. The prosecutor will likely seek an exemplary sentence to set a precedent.

✅ Conclusion

Jonathan Rinaldi's arrest marks the moment when the US justice system stopped treating political deepfakes as a technological curiosity and started treating them as what they are: criminal falsifications. The precedent has been set, and it does not rely on a fragile anti-AI law, but on a forgery offense that has been tried and tested for centuries. The 2026 midterm elections will be the first real-world test of this new legal reality.