How AI Rationales Are Written

Some card pages include a short paragraph explaining why a card was banned or restricted. Where you see a note at the bottom of that paragraph, it means an AI language model helped write the first draft. This page explains, in plain terms, how that works.

Where the information comes from

The AI is never asked to explain a card from general knowledge of the game, and it isn’t connected to the internet while it works. It only reads the official banned and restricted announcements archived on this site that are relevant to that specific card, and is asked to summarize what those announcements actually say. If a card was never explicitly named but belonged to a group that was banned together, the announcement covering that group is what gets used.

In other words, the AI is acting as a summarizer of source material you could read yourself, not as an independent authority on Magic history.

How we steer the AI

Language models tend to ramble, hedge, or make things up when given loose instructions, so the AI is given a strict, narrow set of rules for every card:

Only use the announcements provided for this card. Don’t bring in outside knowledge.
Stay focused on the named card. Other cards are mentioned only if they directly affected its status.
Don’t restate the ban and restriction history — that’s already shown above the paragraph. Explain the reasoning behind it instead.
Don’t describe what the card does on the battlefield. Only its banning rationale matters.
Only reason about Standard, Pioneer, Modern, Legacy, Vintage, Commander, Pauper, and Extended. Other formats aren’t factored in.
Write one plain paragraph, in a neutral, encyclopedic tone. No filler, no hedging, no phrases like “this summary” or “here is,” and no pointing back at the source announcements themselves.
If the announcements don’t clearly explain the reasoning, skip the card entirely rather than guess.

The AI is also shown a handful of worked examples pairing an overly verbose draft with the tighter version we actually want, so it has something concrete to calibrate against beyond the rules above.

The goal is a consistent, factual paragraph that reads like the rest of the site, not something that sounds like it came from a chatbot.

Because of that last rule, coverage isn’t complete. A card without a rationale simply hasn’t had enough written about it in the archived announcements yet — it isn’t a sign that its ban was undocumented or arbitrary.

Human review

An AI-written paragraph is a draft, not a final word. A human editor reads it, checks it against the source announcements, and either approves it as-is or rewrites it. The note under the paragraph always tells you which model wrote the draft, when it was last updated, and whether it has been reviewed and approved yet.

Help us review

Rationales that haven’t been approved yet show a small prompt underneath the paragraph asking whether it looks accurate:

Yes or No casts a quick, anonymous vote. Votes help prioritize which drafts a human editor looks at first — ones flagged as inaccurate move to the front of the line.
Submit a Correction takes you to the contact form if you’d like to explain what’s wrong.
Don’t ask me again hides the prompt going forward. That preference is saved only in your browser and is never sent to us.

No account or personal information is involved in voting.

Questions or corrections

If a rationale looks wrong or missing context, please reach out and it will be looked into.