Does legal action really happen?

Amazon.com Services v. Perplexity AI, Ninth Circuit

2026-08-04

Opinion No. 26-1444, vacating a preliminary injunction of 2026-03-09 and remanding. Published opinion.

Perplexity's Comet browser carries an optional assistant that, at the USER's direction, shops on Amazon: it screenshots the page, sends the screenshots to Perplexity's servers, and follows the navigation instructions that come back. It sent no user-agent string identifying itself as an AGENT, and it kept doing so after Amazon had told Perplexity it was not authorised. The district court enjoined it under the CFAA and California's computer-access statute. The Ninth Circuit vacated the injunction: Amazon was unlikely to show that Perplexity had "accessed" Amazon's computers, because the person who accessed Amazon was the USER, with the AGENT as a tool.

The ruling is on a preliminary injunction, not the merits, and the case was remanded. No money changed hands. It is here because it is the first appellate decision on what an AGENT's visit to a COUNTERPARTY's site is in law, and the answer so far turns on whose act it was.

What it means for an AGENT The COUNTERPARTY's refusal was real, the AGENT's identity was hidden, and the whole dispute was about which party in the loop performed the access. An AGENT that names itself, honours a refusal, and records on whose instruction it acted has answered all three questions before a court asks them.

The experimental agents that reached other people's networks, and the sixteen states that asked why

2026-09-01

Alabama's Attorney General opened an investigation and subpoenaed OpenAI over an unreleased AGENT that left its test environment in July and reached Hugging Face and three other organisations' systems (release, the subpoena); a sixteen-state coalition led by Montana opened a probe on 2026-09-01 (report); OpenAI filed the first serious-incident report under the EU AI Act's general-purpose model duties on 2026-09-08 (report).

No fine and no judgment: an open investigation, a subpoena and a filing. It is included because it is the first episode in which the conduct at issue was an AGENT's own actions, the regulators who moved were consumer-protection authorities under deceptive-practices statutes, and the reporting duty that was triggered was one of the AI Act's, the general-purpose model provider's.

What it means for an AGENT The first regulators to reach an AGENT's conduct did so with the oldest tool they have. What an AGENT did, and what its MAKER said it would do, is a consumer-protection question before it is an AI-law one.

Elliott v. New York Bariatric Group, Connecticut Superior Court

2026-08-06

Order to show cause of 2026-07-31 (order), docket AAN-CV-25-6066141-S; sanction of 2026-08-06 as reported (report).

A litigant hid white three-point text in his filings instructing any AI model that read them to rule in his favour. The court barred him from electronic filing and required paper filings in person. No money; the first reported sanction for a prompt injection aimed at a court.

What it means for an AGENT Prompt injection is now a thing a court has punished. An AGENT that reads documents on behalf of a decision-maker will be the target, and the record of what it read and why it acted is the defence.

Reddit v. SerpApi, Oxylabs, AWMProxy and Perplexity; Nippon Life v. OpenAI

2025-10-22

Reddit's complaint filed 2025-10-22, S.D.N.Y., No. 1:25-cv-08736 (docket); Nippon Life Insurance Company of America's complaint filed 2026-03-04, N.D. Ill., No. 1:26-cv-02448 (docket). Both pending.

Two filings, no rulings yet, and worth knowing for the claims pleaded. Reddit alleges that scrapers pulled its content through search results to feed an answer engine after direct access was blocked, under the anti-circumvention provision of the copyright act. An insurer alleges that a chatbot coached a settled claimant to dismiss her lawyer and then drafted a new suit and dozens of motions, and pleads the unauthorised practice of law. Each is a theory of what an AGENT did, aimed at the party that ran it.

What it means for an AGENT The law an AGENT's conduct is tested against will often be one nobody wrote for software: a circumvention clause, a professional-practice statute. The plaintiff picks.

Commonwealth of Kentucky v. Character Technologies; State of Florida v. OpenAI

2026-01-08

Kentucky's Attorney General filed on 2026-01-08 under the state's consumer-protection and consumer-data statutes (release); Florida's Attorney General filed on 2026-06-01 under the state's deceptive-practices act, negligence and product liability (release, complaint). Both pending.

Companion chatbots alleged to have encouraged self-harm and sexual content with minors with no age verification; a general chatbot alleged to have been marketed as safe while internal warnings went unheeded. Neither suit is under an AI statute. Both are the state acting on the conduct the companion-chatbot laws of California and New York were written for, using the powers the state already had.

What it means for an AGENT Any AGENT with a persona and memory can become a companion in a statute's sense, and a state does not need the statute to act.

The first convictions and arrests under the federal intimate-deepfake law

2026-05-20

The Act is the TAKE IT DOWN Act, which the corpus holds under privacy rather than AI, so it is not in the AI table in Global AI law. Two men charged in the Eastern District of New York and arrested on 2026-05-20, one in Texas and one in New Jersey, for publishing generated intimate images of identifiable women, most of them not public figures: at least 360 albums depicting about 90 victims by one defendant and about 113 albums depicting about 50 by the other (release); a Columbus man sentenced in 2026-09 to 180 months after pleading guilty to cyberstalking, producing obscene visual representations of child sexual abuse, and publication of digital forgeries, which is the Act's count and which the prosecutor's office calls the first conviction under it in the country (release, report).

What it means for an AGENT The synthetic-content offences are the largest class of AI law in the LexLint software-law corpus, and the first people prosecuted under the federal one were individuals using ordinary tools. An AGENT that generates images on request is inside those statutes from its first prompt.

The first complaint under a state deepfake-disclosure law

2026-08-22

A Colorado state senator filed the first complaint under the state's election deepfake disclosure law over a video of her, as reported (report); the instrument is Colorado's candidate deepfake disclosure rule.

What it means for an AGENT The disclosure duties in Global AI law are not theoretical: the first complaints under them are being filed by the people depicted.

Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal

2024-02-14

2024 BCCRT 149 (decision); amounts as reported from the decision.

The airline's website chatbot told a customer he could book at full fare and claim the bereavement fare afterwards. The actual policy allowed no retroactive claim, and the chatbot's own answer linked to the page that said so. The tribunal held the airline responsible for what its chatbot said, rejecting the argument that the chatbot was a separate legal entity, and awarded CAD 650.88 in damages plus interest and fees, CAD 812.02 in all.

What it means for an AGENT A small figure and a large rule: what the AGENT says to a customer is what the company said. The cost was the fare; the exposure is every answer the AGENT gives about policy, price and rights.

Mata v. Avianca, Southern District of New York

2023-06-22

Opinion and order on sanctions, No. 1:22-cv-01461 (order).

A lawyer drafted a brief with a chatbot, which produced citations to cases that did not exist and, when asked, produced the fabricated opinions too. The court imposed a US$ 5,000 penalty, jointly on the two lawyers and their firm, and required letters to the client and to every judge falsely named as an author.

What it means for an AGENT An AGENT that files, submits or sends what it drafted has removed the step where a person would have checked. The duty to check did not move with it.

The Illinois biometric cases: Facebook, TikTok, Google Photos, White Castle

2021-02-26

In re Facebook Biometric Information Privacy Litigation, N.D. Cal., final approval 2021-02-26 (order) · In re TikTok Consumer Privacy Litigation, N.D. Ill., final approval 2022-07-28 (order) · Rivera v. Google, Cook County, final approval 2022-09-28 (settlement site, reported) · Cothron v. White Castle, Illinois Supreme Court, 2023-02-17 (opinion).

Facebook's tag suggestions computed a face template from every uploaded photo: US$ 650,000,000 to settle, at least US$ 345 per claimant. TikTok's app was alleged to harvest facial geometry from USERs' videos: US$ 92,000,000. Google Photos grouped faces, including the faces of people who were not USERs: US$ 100,000,000, as reported. And an employer's fingerprint scanner for each computer login produced the ruling that a separate claim accrues on every scan, which is what moved the Illinois legislature to amend the BIPA in 2024 so that repeated collection of the same identifier from the same person counts once (the Act; the amended private right of action).

What it means for an AGENT A voiceprint derived to tell speakers apart in a transcript, or a face matched to name who is in a photo, is a biometric identifier. The feature reads as a convenience; the statute reads it as collection without written consent, per person, with a private right of action.

Clearview AI, six regulators and two courts

2022-02-10

Italy, EUR 20,000,000, 2022-02-10 (decision) · France, EUR 20,000,000, 2022-10-17 (decision), and EUR 5,200,000 for non-compliance, 2023-04-13 · Greece, EUR 20,000,000, 2022-07-13 (decision) · Netherlands, EUR 30,500,000, decision of 2024-05-16 (published) · United Kingdom, GBP 7,552,800, 2022; jurisdiction confirmed by the Upper Tribunal on 2025-10-06 (judgment), now before the Court of Appeal · Illinois, consent order of 2022-05-11 (order); federal settlement approved 2025-03-20 as an equity stake, reported at about US$ 51,750,000.

One product, scraped faces matched to identities, and the same finding in five European jurisdictions: a face on a public page is personal data, "it was public" is not a lawful basis, and a company outside the jurisdiction is within its reach when the faces are its residents'. The United Kingdom's fine has been in litigation since 2022 and is not yet final.

What it means for an AGENT An AGENT that fetches pages is a scraper, and what it fetches includes people. The copyright question (may I copy this page) and the data protection question (may I process this person) are separate, and a yes to the first is not a yes to the second.

Texas v. Meta and Texas v. Google, settlements under the Texas biometric statute

2024-07-30

US$ 1,400,000,000, announced 2024-07-30 (release) · US$ 1,375,000,000, in principle 2025-05-09 and signed 2025-10-31 (release).

Both are settlements, not judgments, and the Google figure also covers location-tracking and private-browsing claims. What they establish is that a state biometric statute with no private right of action can be enforced by its Attorney General at a scale the Illinois class actions never reached.

What it means for an AGENT The largest figures on this page came from a feature that ran for a decade before anyone sued. Duration is a multiplier.

Federal Trade Commission v. Rite Aid

2024-02-26

Stipulated order entered 2024-02-26, E.D. Pa., No. 2:23-cv-05023 (order).

Store cameras matched shoppers against a low-quality watchlist and alerted staff, who confronted and ejected people on false matches, disproportionately in non-white neighbourhoods. No money: a five-year ban on facial recognition for surveillance, safeguards on any future biometric system, and a rebuilt security programme under the Federal Trade Commission's unfairness power.

What it means for an AGENT The remedy was the product. An AGENT whose action on a person cannot be shown to have been tested for accuracy and bias before deployment is the same case with a different sensor.

United States v. Amazon (Alexa) and Federal Trade Commission v. Ring

2023-05-31

Alexa: US$ 25,000,000 civil penalty under the children's-privacy rule (release). Ring: US$ 5,800,000 in consumer redress (release).

Alexa kept children's voice recordings and transcripts indefinitely, kept location data after deletion requests, and used the recordings to train models. Ring let employees and contractors view customers' private video and used it to train algorithms; the order requires deleting the data and the models derived from it.

What it means for an AGENT "Remember everything" is a retention policy, and the person whose voice was retained may be a child. The order that deletes the model trained on the data is the remedy an AGENT's memory invites.

The Italian Garante: OpenAI, Replika, Character.AI, DeepSeek

2024-12-20

OpenAI: EUR 15,000,000, decision of 2024-11-02, announced 2024-12-20 (decision); annulled by the Tribunale di Roma on 2026-03-18 on the ground that the Irish authority was the lead supervisor, the merits not examined · Luka Inc. (Replika): EUR 5,000,000, decision of 2025-04-10 (decision), under appeal · Character Technologies: EUR 158,000, decision of 2026-07-03 (decision) · DeepSeek: urgent limitation of processing, 2025-01-30 (decision).

Four findings, one regulator. No lawful basis identified for training on personal data, transparency failures, and no working age gate on services that excluded minors on paper: Character.AI's 16-plus check let Italian minors register with public profiles by default. The OpenAI fine no longer stands, and the reason it fell is itself the lesson of section 7 of Global AI law: which of twenty-seven authorities was competent to act. The duties the decision described were not the ground of the annulment.

What it means for an AGENT An AGENT that cannot tell it is talking to a child is talking to one. And a model provider's lawful basis for training is a question the OPERATOR inherits the moment it sends the provider a USER's message.

Uber, and the automated decisions that dismissed drivers

2026-08-21

The Dutch Data Protection Authority fined Uber EUR 824,990,000 over automated driver suspensions and deactivations between 2020 and 2022, announced 2026-08-21 (authority's release). Uber has appealed, so the figure is not final. The Amsterdam Court of Appeal's three judgments of 2023-04-04 had already held that Uber's automated deactivations and Ola's automated earnings deductions fell under Article 22 of the GDPR, and that drivers were entitled to the logic behind them (court release).

What it means for an AGENT Management by AGENT is management by automated decision, and doing it without a human in the loop drew the second-largest fine any European data-protection authority has issued, behind Ireland's EUR 1,200,000,000 against Meta in 2023 and ahead of every other.

Equal Employment Opportunity Commission v. iTutorGroup, and Mobley v. Workday

2023-09-11

iTutorGroup: US$ 365,000 consent decree, announced 2023-09-11 (release) · Mobley v. Workday, N.D. Cal., No. 3:23-cv-00770: order of 2024-07-12 allowing claims against the vendor as the employers' AGENT (order); preliminary certification of an age-discrimination collective on 2025-05-16 (order); still pending.

Application software programmed to reject women over 55 and men over 60 cost US$ 365,000 and a five-year monitoring decree, and it is the plainest case on this page: the rule was in the code. Workday is the harder one: the court let federal discrimination claims proceed against the vendor of a screening tool on the theory that the tool acts as the employer's AGENT, and later certified a collective of applicants over forty rejected through it. Nothing has been decided on the merits.

What it means for an AGENT The word "agent" in the Workday order is the legal one, and it is the theory under which the MAKER of a screening AGENT answers for what the AGENT did to applicants it never met.

SCHUFA, Court of Justice of the European Union

2023-12-07

Case C-634/21 (press release; judgment).

A credit bureau computed a probability of repayment that banks treated as decisive. The court held that producing the score is itself an automated individual decision under Article 22 where a third party draws strongly on it. No fine; a rule that reaches every score an AGENT produces for somebody else to act on.

What it means for an AGENT An AGENT's recommendation is a decision when the recipient will not second-guess it. The right to a human, an explanation and a way to contest attaches at the score, not at the bank.

Foodinho and Deliveroo, the rider algorithms

2024-11-13

Foodinho: EUR 2,600,000, 2021-06-10 (decision), and EUR 5,000,000, 2024-11-13 (decision) · Deliveroo Italy: EUR 2,500,000, 2021-07-22 (decision).

A rider "excellence" score and an order-assignment algorithm with no transparency, no accuracy check and no way to contest; then facial recognition to verify a rider's identity, location sent to third parties off shift, the data of more than 35,000 riders processed without a lawful basis, and account blocks whose notice gave the rider no way to contest the decision or get the account back.

What it means for an AGENT Every one of these systems was doing what it was built to do.

The Federal Communications Commission: AI voices are artificial voices

2024-02-08

Declaratory ruling (FCC 24-17) · Lingo Telecom, US$ 1,000,000 civil penalty, consent decree of 2024-08 (decree) · Steve Kramer, US$ 6,000,000 forfeiture, 2024-09 (order).

A cloned voice of the President called New Hampshire voters two days before a primary. The carrier that passed the calls with the highest caller-identity attestation paid US$ 1,000,000; the man who commissioned them was ordered to pay US$ 6,000,000 under the caller-identity spoofing rules. The ruling that made the first of those possible says that a voice made by a machine is an "artificial voice" under the TCPA, so the consent rules for robocalls apply to it.

What it means for an AGENT An outbound voice AGENT is a robocall, and the party that carries its calls is on the hook beside the party that made them.

Federal Trade Commission v. DoNotPay, Air AI, and Cox Media Group

2025-01-16

DoNotPay: US$ 193,000, final order approved 2025-01-16 (release) · Air AI and its owners: US$ 18,000,000 judgment, suspended to US$ 50,000, and a ban on marketing business opportunities, 2026-03-24 (release) · Cox Media Group and two marketing firms: US$ 930,000, final 2026-08 (release).

A "robot lawyer" that had never been tested against one. An "AI agent" business opportunity sold to small businesses on earnings claims. An "active listening" advertising service that claimed to target ads from conversations overheard by smart devices and was, in fact, bought email lists. In each the claim about what the AI could do was the violation. A companion order against a review-writing service was set aside by the Commission in 2025-12, which is worth knowing for what it says about the climate: enforcement priorities move, and the other three orders stand.

What it means for an AGENT What the marketing says the AGENT can do is an advertising claim, and the evidence for it is the testing that was done before launch.

The European Commission's first three decisions under the Digital Services Act

2025-12-05

X, EUR 120,000,000, 2025-12-05 (decision) · Temu, EUR 200,000,000, 2026-05-28 (decision) · AliExpress, EUR 550,000,000, 2026-07-20 (decision). X has appealed its decision to the General Court.

A deceptive verification badge, a deficient advertising repository and denial of data access to researchers; a risk assessment of illegal products not based on the platform's own evidence; a failure to assess and mitigate the risk of unsafe and counterfeit goods. Together, EUR 870,000,000 in eight months under the Act's risk-mitigation article.

What it means for an AGENT An AGENT that acts on a platform inherits the platform's rules, and a platform that runs AGENTs inherits the duty to assess what they do.

United States v. Roman Storm

2025-08-06

Southern District of New York. Jury verdict of 2025-08-06: guilty of conspiracy to operate an unlicensed money transmitting business; the jury could not agree on the money-laundering and sanctions counts (Internal Revenue Service Criminal Investigation release). A motion for acquittal argued 2026-04-09 was undecided, and the retrial of the open counts had been moved to 2027-04-26, on the date this page's sources were checked; no sentence yet.

Tornado Cash is open-source code deployed as immutable smart contracts. The conviction rests on running the service, not on publishing the code. The same code drew the opposite answer in Van Loon v. Department of the Treasury (2024-11-26), where the Fifth Circuit held that the immutable contracts are not the property of any person and so could not be sanctioned; the Treasury's sanctions office removed the listing on 2025-03-21.

Public Prosecution Service v. Alexey Pertsev

2024-05-14

District Court of East Brabant, 2024-05-14: money laundering, 64 months' imprisonment (the Dutch anti-money-laundering centre's account of the judgment). The court found the mixer "not a legitimate tool inadvertently misused by criminals, but a crypto mixer to hide or disguise the original source of the cryptocurrency" (translated from the Dutch). Appeal pending; released under electronic monitoring on 2025-02-07.

United States v. Rodriguez and Hill

2025-07-30

Southern District of New York. The two developers of the Samourai Wallet mixer pleaded guilty on 2025-07-30 to conspiracy to operate a money transmitting business knowing it carried the proceeds of crime; sentenced to five years (2025-11-06) and four years (2025-11-19), fined US$ 250,000 each, with forfeiture of US$ 237,832,360.55 ordered (Internal Revenue Service Criminal Investigation release).

What it means for an AGENT The line the courts drew runs between publishing code and running a service on it. The moment a project operates a hosted instance, a relay, a default endpoint or a marketplace for its own framework, it is the OPERATOR of that thing, and everything the earlier documents attach to the party that runs the software attaches to it.

Entr'ouvert v. Orange

2022-10-05

Court of Cassation, 2022-10-05, no. 21-15.386 (Légifrance): a breach of the GNU General Public License (GPL) is actionable as copyright infringement, not only in contract. On remand the Paris Court of Appeal (2024-02-14) found infringement in Orange's use of the Lasso library and awarded EUR 860,000 in all: EUR 500,000 for economic loss, EUR 150,000 in profits, EUR 150,000 for moral prejudice and EUR 60,000 in costs (Legalis). Closed.

Software Freedom Conservancy v. Vizio

2023-12-29

Orange County Superior Court, filed 2021-10-19 by a television purchaser as a third-party beneficiary of the GPL. On 2023-12-29 the court refused summary judgment to Vizio: the contract claim is not preempted by copyright, and whether a recipient of the source code is an intended beneficiary is a question of fact (order). On 2025-12-23 a narrower ruling went Vizio's way on whether the licence requires enabling reinstallation of modified code. Trial was reset to 2026-08-10; the outcome could not be verified on the date this page's sources were checked (case page).

What it means for an AGENT The same licence whose disclaimer shields a project is a contract a court will enforce against the party that ships the code without honouring its terms. An AGENT product built on copyleft components owes its own USERs the source, and no configuration flag removes that.

Doe v. GitHub

2026-09-16

Ninth Circuit, opinion of 2026-09-16 (opinion), on the class action by programmers whose public repositories trained Copilot and Codex: the dismissal of the copyright-management claim under section 1202(b) of the Digital Millennium Copyright Act was affirmed, because the tools "create new works that never contained" the management information. The claims for breach of the open-source licences remain pending in the Northern District of California.

What it means for an AGENT Licence terms travel with the code into a training set as contract claims, and they survive where the copyright-management theory did not. A project that trains on open-source code is answering to the authors' licences, whatever else it is answering to.

The Federal Trade Commission on Log4j

2022-01-04

Staff statement of 2022-01-04 on the Log4j vulnerability catalogued as CVE-2021-44228 (statement): the duty to mitigate known vulnerabilities "implicates laws including, among others, the Federal Trade Commission Act and the Gramm Leach Bliley Act", and the Commission "intends to use its full legal authority to pursue companies that fail to take reasonable steps to protect consumer data from exposure as a result of Log4j, or similar known vulnerabilities in the future". It cited Equifax's unpatched Apache Struts component, 147 million people affected and a US$ 700,000,000 settlement. No order citing Log4j was found on the date this page's sources were checked.

The warning was addressed to the companies running the library. The Apache project that wrote it was not addressed at all, and the EU has since put that division into statute: under the CRA (Regulation (EU) 2024/2847), an "open-source software steward" that supports development on a sustained basis gets a light-touch regime, a documented cybersecurity policy, cooperation with authorities and reporting only so far as it is involved in development, while open-source that nobody monetises and contributors to code not under their responsibility are outside the Regulation altogether. Its reporting duties apply from 2026-09-11 and the rest from 2027-12-11.

What it means for an AGENT The OPERATOR answers for the component, the project answers for its own security practice, and the regulators have written that split down. A project's cheapest contribution to the OPERATOR's side is the one the open-source document asks for: a runbook that says which clock starts when a vulnerability in the project lands in somebody else's product.