Check what the regulation actually says

A line in a board paper says the EU AI Act "bans social scoring by governments." It came from a briefing written a while ago, and it's about to be copied into a risk register and a customer FAQ. Most of it is right. The part it leaves out is the part a private company most needed to know. This page shows how to check claims about laws and rules against the official text, how a result shows what applies from when, and how to keep a sourced record of what you checked.

Why claims about laws drift

Law keeps moving after the commentary is written. Summaries are drafted from proposals, then quoted from other summaries, and the details go stale while the prose still sounds authoritative.

The EU AI Act shows how. The European Commission's 2021 proposal did limit the social scoring ban to public authorities. In its December 2022 position, the Council extended the prohibition to private actors, and the final text of Article 5 doesn't restrict it to public bodies. Any briefing written from the proposal still carries the old version.

Dates move too. According to the Commission's AI Act page, the Act entered into force on 1 August 2024, its prohibitions have applied since 2 February 2025, and it became generally applicable on 2 August 2026. The AI Omnibus, in force since 27 July 2026, then set new dates for high-risk systems: 2 December 2027 for certain high-risk areas, and 2 August 2028 for AI built into products such as lifts or toys. An explainer from 2025 that gives August 2026 as the high-risk deadline is now wrong. Even the Commission's AI Act Service Desk page for Article 113 carries a notice that its text "has not yet been updated to reflect those amendments."

How a check handles a regulatory claim

For a single sentence, use Check a claim. For a wider question, such as what the AI Act now expects of companies that deploy high-risk systems, use Research a topic: a short brief on what's established, what's contested and what's still open, with every point linked. Our page on sourced research briefs shows what that looks like.

Either way, CiteJury splits the claim into parts and searches several independent search indexes for evidence for and against each one. It favors primary sources such as official pages, regulators and court records, reads PDFs as well as web pages, and skips social media. The best passages are sealed into one evidence pack. Claude, ChatGPT and Grok each read that pack and answer separately, every quote they cite is checked word for word against its source, and a final review writes the verdict and a safe way to say it.

Here's the AI Act claim going through. It ran on Deep, where the AIs also run their own searches while sources are gathered.

The AI Act claim, part by part

The claim was: "The EU AI Act entered into force on 1 August 2024 and bans social scoring by governments." CiteJury searched 145 results. The sources it found included the Commission's digital strategy pages, the AI Act Service Desk and the EU Publications Office, along with analysis on cambridge.org and from the law firm WilmerHale. EUR-Lex, the EU's official law site, couldn't be opened on this run, so the Commission's own AI Act pages carried the official side, and 28 sources went into the sealed pack.

The AIs split. Claude and Grok called it Correct, and ChatGPT called it Incorrect. The final review didn't paper over that: it returned Sources disagree and showed exactly where the sentence bends.

• It entered into force on 1 August 2024. Correct, from the Commission's own pages.
• It prohibits social scoring. Correct, under Article 5.
• The ban covers governments. Correct, but it doesn't stop there: it covers private companies too.
• The bans have applied since 2 February 2025. Needs care: that's right for the original bans, but the review noted that two bans added in 2026 apply later.

The safe wording is the sentence you'd want in the board paper: "The EU AI Act entered into force on 1 August 2024. Its Article 5 ban on social scoring applies to public authorities and private companies, and has applied since 2 February 2025." All three sources shown back it up, and two of them are the Commission's own. The check took about seven minutes.

It's a common shape for a drifting legal claim. Nothing in it is invented. The date is right and the ban is real, but a qualifier that matches the 2021 proposal rather than the final law has survived into the summary, and the summary is what people quote.

Checking a regulatory claim

1. Write the claim as it will be used. Name the jurisdiction and the instrument, and the date on which the claim is meant to be true.
2. Pick a tier. Use Deep for anything recently amended, where a wider search helps. Standard, at about 8 credits and usually one to three minutes, suits settled points. Quick, with one AI, is enough for a first look.
3. Read the parts. Who a rule covers and when it applies are often the parts that need care.
4. Weigh the sources by their rulings. An official text that Backs it up or Contradicts it counts for more than commentary marked Background.
5. Ask a follow-up. A question such as "Does the ban cover employers?" is answered only from the same sealed sources, for about 1 to 2 credits.
6. Open the official text for the provision the result points to, and look for any notice of later amendments.
7. Keep the record. Download the sources as Excel or CSV, group related checks into a project, or share the check's link with legal. Anyone with the link sees the check and its sources.

Habits for policy and compliance work

• Date every claim. "Applies" means little without "from when". Write "has applied since" or "will apply from".
• Separate the law from guidance. A regulation, the regulator's guidelines and a law firm's reading of both carry different weight. Say which one you're relying on.
• Check the scope words first. "Governments", "large companies", "consumers" and "EU-based" are small words that change who a rule reaches.
• Check proposals against final texts. Commentary written during negotiations can quote provisions that later changed.
• Check again before you reuse. A line that was right in last year's policy memo may not be right now.
• Watch your own sales copy. Compliance claims in marketing are claims too; see checking marketing claims.

What it won't do

CiteJury isn't legal advice. It tells you what the sources it found say about a claim, not whether a rule applies to your organization or what you should do about it. That's a question for counsel, and a sourced check gives them a faster starting point.

• It can't read pages behind paywalls or logins, such as subscription legal databases, and some sites block automated access. The result lists what it couldn't open.
• Scanned PDFs without a text layer can't be read, because there's no OCR.
• It checks text, not images or charts, so a timeline that exists only as a graphic won't be read.
• Very recent changes may have little coverage yet, which often means Can't confirm.
• A verdict is only as good as the sources found, and it can be wrong or incomplete. Every sentence links to its source so you can check it yourself.

Rules will keep changing faster than the summaries of them. Checking a claim against the official text each time you reuse it is slower than copying it forward, and much quicker than unwinding a decision built on last year's version.

Questions people ask