LexLint

The law for AI, agents, crawlers, and scrapers

The case for caring

Is the legal risk real?

Law cited as of 19 August 2026

Lem, a small robot with a large LEM name plate across his front, stands on the road between two neighbouring building plots, each holding an identical half-built house. A permit case on the near plot's gate post holds a notice headed PERMIT. The identical case on the next plot's gate post is empty, its backing board bare, and a small red-bordered board wired across that plot's closed gate reads STOP WORK.
Lem filled in a form. His neighbour is having a quieter week.

The question is usually asked as though the answer were somewhere in the future, and that is the part which is wrong. Regulators are using powers they already have, companies are suing each other over training data and access, and courts are ordering changes to products rather than payments out of them. Here is the evidence, with the cases named.

§1 · It already started

The dates are behind you, not ahead of you.

Three commencement dates, all of them in the past:

  • 2 August 2025. The EU AI Act's obligations for providers of GPAI models took effect (Article 53). That was a year ago.
  • 2 August 2026. The AI Act's transparency obligations became enforceable (Article 50): a system that interacts with a person has to disclose that it is a machine, and synthetic output has to be marked.
  • 2 August 2026. California's AI Transparency Act became operative the same day (SB 942), with its own separate answer to the same question.

Of the instruments LexLint holds a commencement date for, 16 were already in force on the morning this page was built. The corpus behind that figure tracks 256 jurisdictions, and you can read it at lexlint.org/law.

And it moves backwards too

This is the half nobody plans for. A law you built against can be enjoined before it bites, or struck down after it has, or repealed and replaced with something differently shaped. These are tracked instruments that are no longer standing as enacted:

So a legal read done once is a photograph, not a position. That is the argument for a lint rather than a memo: the memo is correct on the day it is written, and the lint runs again on Tuesday.

§2 · Governments are enforcing

The powers are live, and they are being used.

European Commission and national market surveillance authorities · from 2 August 2026

AI Act Article 50, Regulation (EU) 2024/1689

Since 2 August 2026 the transparency obligations bind providers and deployers of chatbots, synthetic-media generators, emotion-recognition systems and deepfake tools, whether or not the system underneath is high risk. Non-compliance carries up to €15 million or 3 per cent of worldwide annual turnover, whichever is higher. Enforcement sits mainly with national market surveillance authorities. A transitional period runs to 2 December 2026, and it covers only the marking and detection duty for generative systems that were already on the market.

What it changes for your code If your app talks to a person, it says so. If it produces synthetic media, the output carries a machine-readable mark. Both are decisions about architecture, and both are far cheaper before the interface exists than after.

California's AI Transparency Act became operative on the same day and answers the same question differently: a covered provider has to publish a free detection tool, offer a visible disclosure option, and embed latent provenance in generated image, video and audio. Text output is out of scope. The penalty is $5,000 per violation, and each day in violation counts as a separate one. It bites at scale rather than on day one, since the definition of a covered provider starts at a million monthly users.

Enforcement, from the corpus

33 stories on file
Every enforcement story we hold →

§3 · Companies are suing each other

And the claim can be built out of a contract you never signed.

United States District Court, Northern District of California · 4:25-cv-10826-YGR

Google v. SerpApi

The claims are not copyright in the content. They are DMCA section 1201 anti-circumvention claims about getting around a bot detection system. The original complaint was dismissed on 20 July 2026 and amended on 10 August 2026 on a narrower theory, and the interesting part is what the narrower theory is built from: Reddit's licensing agreement with Google. Google pleads that Reddit directed it not to let third parties extract and commercialise the licensed content, and asked it to apply technical measures when scrapers did.

What it changes for your code The authority behind a barrier you route around may be an agreement between two other parties. You cannot read it, you are not bound by it, and it is still what turns a technical measure into a protected one.

The platform gets there before the court does

In March 2026 a district court granted Amazon a preliminary injunction against Perplexity's shopping assistant. On 4 August 2026 the Ninth Circuit vacated it, holding that the CFAA “remains principally an anti-hacking statute”, and that where an assistant acts under a user's direction it is the user, not the assistant's operator, who accesses the system.

Perplexity won. It won roughly five months after its product stopped working against the largest retail catalogue on the internet. Read the timeline rather than the holding: the cost landed at the injunction, not at the judgment, and no part of it was recovered by being right.

Your customer is the third enforcer

Article 50 puts obligations on deployers, not only on providers. An enterprise buyer running your system carries duties that depend entirely on what your system does and on what you can tell them about it. That is why the AI section of a security questionnaire exists, and it is why an unanswerable question there costs a quarter rather than a fine.

Litigation, from the corpus

51 stories on file
Every litigation story we hold →

§4 · The remedy is your product

Not a line item. A watermarking pipeline, a rebuild, and a licence.

Landgericht München I (Munich Regional Court I) · 42 O 763/25 · 31 July 2026

GEMA v. Suno

The court found that the works were retained in the model's parameters, and that full retention in model parameters falls outside the text-and-data-mining exception in section 44b of the German Copyright Act, which reaches only reproductions genuinely necessary for analysis. On outputs it held Suno responsible rather than its users, reasoning that the model's architecture and memorised training data, not the prompt, substantively determined what came out.

Then it took jurisdiction over the training itself. The injunction ordered Suno to stop copying the works “within the territory of the United States of America for the purpose of training an artificial intelligence (AI) model to generate music”.

What it changes for your code Where the training ran is not where the question gets decided. A TDM exception that covers you in one country may not cover the same pipeline in the next, and what fell outside it here was not the crawling but what the model kept.

What happened next is the part worth sitting with. Within eight days Suno announced audio watermarking and fingerprinting, integrated a third-party copyright detection system, capped how much subscribers can download, and signed a global licence with BMG that settles its prior use of that catalogue.

None of that is a payment. It is a provenance pipeline, a third-party integration, a product limitation and a commercial agreement, shipped against a deadline set by somebody else. That is what a remedy looks like in this area, and it is the reason “we will deal with it if it happens” prices the risk wrongly. What you are exposed to is not a number in next year's budget. It is a quarter of engineering you did not choose, on a schedule you do not control.

§5 · Which is why early is cheap

Every one of these is a design decision before it is a remedy.

Read back over the cases and notice what the obligations actually ask for. Does the app say it is a machine. Does its output carry a mark. Do you know where the training data came from. Does the crawler respect the barrier in front of it, and do you know what that barrier is protecting and for whom. While the code is a sketch each of those is an afternoon. After launch each one is a migration, a retrain, or a renegotiation, and Suno's eight days are what the compressed version looks like.

At design time

  • A disclosure is a string and a place to put it
  • Provenance marking is a library and a build step
  • Training-data records are a column you were going to add anyway
  • A jurisdiction you decide not to serve costs you nothing yet

After launch

  • A disclosure is an interface change across every surface
  • Provenance marking is a re-encode of everything already shipped
  • Training-data records are an archaeology project, if they exist
  • A jurisdiction you leave is revenue you already booked

LexLint exists to make the left-hand column happen. Your coding agent declares what the app does and where it runs, and the lint comes back with the obligations that attach, each with a citation and an as-of date. It is a lint, not a lawyer: it catches the basics early, while they are still cheap, and it says outright when it has no current data for a jurisdiction rather than passing you.

https://mcp.lexlint.org/mcp Connect your agent → See a full run →

Law cited as of 19 August 2026. Court decisions are summarised from published reports of the rulings; the instrument pages linked above carry their own citations and as-of dates. Everything LexLint tracks is at lexlint.org/law, and the coverage this page draws on is at lexlint.org/news. LexLint is not legal advice and is not a substitute for counsel in the jurisdictions you operate in.