How we use AI

The SharkVault's guides are drafted by AI. Automated gates decide what is publishable, and people decide what publishes. This page says where that line sits, what those gates catch, what they miss, and what we cannot currently prove. If you paid for this work, you should know which parts a machine wrote.

What the AI does

AI drafts guide sections from a curated brief. It indexes and cross-references cheat codes across thousands of entries, which is work no one was ever going to do by hand. It also generates the cover art and backdrops you see on our own product pages, so the artwork is ours rather than a publisher's.

What people do

People curate the verified facts each draft is built from, correct sections by hand, and decide when a game opens to readers. No draft publishes itself. Corrections, disputes, and anything touching a code are handled by a person.

The three checks a section has to pass

Before a section can publish, it runs three gates. They exist because they each catch a different kind of lie.

1. Citation density. Anything the draft structures as a claim has to show where it came from.
2. Device codes. Every code-shaped token in the body must appear verbatim in our curated facts for that game. When we hold no verified facts for a section, it may print no codes at all.
3. Substance floor.A body too thin to be a section gets held, even when it is perfectly honest. “No documented easter eggs for this title” is a true sentence and it is not a section.

What those checks do not catch

Citation checking counts citations, not correctness. A draft can hang a real source on a paragraph and still be wrong inside it. We found this on Grand Theft Auto (PS1) in August 2026: a fabricated GameShark address carried a genuine citation and scored a pass.

The code gate was built in response to that, and it is deterministic rather than model-judged. It compares what the draft printed against what we verified, which makes an invented code structurally unpublishable. It cannot catch a code that is wrong but present in our facts file. That one is the curator's job, and curators are people.

Grounding a draft in verified facts reduces invention. It does not eliminate it. On that same title, a fully grounded regeneration still produced two player-name codes that were nowhere in the facts file and corrupted one address we had already verified. Every section needed correction by hand before it published. We are telling you this because the alternative is letting you discover it.

What we cannot currently prove to you

We record who approved every published section. For most of the corpus that record names a batch script rather than a person, because sections were approved in bulk runs that an operator started. The automated gates ran on all of them. A named editor signing off on that specific section did not, or did and left no trace, and from the record we cannot tell those two apart.

So we are not going to tell you a human read every line. Human review is real, and we have pulled and rewritten sections by hand. It is not evidenced section by section, and an assurance we cannot show you is worth nothing. We are adding per-section editor attribution, and once that number means something we will publish it here.

Codes are held to a stricter standard than prose

A cheat archive is only worth what its codes are worth. Ours live in a curated database rather than in model output, and they are copied out of verified sources rather than retyped by a model. Where two archives disagree, we record the conflict and publish neither, so that a later editor cannot rediscover a disputed code as a fact. When we are unsure, we leave it out. An empty section beats a confident wrong answer.

Controls on the AI itself

Hosted AI authoring sits behind an operator kill switch that halts every generation within about a second and needs no redeploy. Each editor also has a daily generation cap. Both exist so that a stuck loop or a compromised account cannot quietly run up spend or flood the review queue with drafts nobody asked for.

Where we do not use AI

We do not use AI to answer support mail, and we do not use it to decide what ships. The RetroShark is an early prototype and we describe it as a design intent rather than a working capability, because that is what it currently is.

When we get it wrong

Tell us. Email hello@gamesharklabs.com with the game and the section. Reports about codes go to the front of the queue. We would rather pull a section than leave a bad code standing.

Last updated: 2026-08-24.