I Forced Ai To Play Pokemon... Its Save Is Terrifying

Storyline (what happens in the video)

• The creator trains an AI to play Pokémon using real save-file data, aiming for the AI to produce a “normal-looking” Pokémon save—something that could plausibly be mistaken for a human player’s file.
• They run multiple attempts (five total), each time giving the AI more training examples (“Pokémon 101,” more game cartridges/regions, and behavioral constraints like “don’t scare me”).
• As the AI’s saves become more realistic on the surface, the creator keeps finding unsettling or illogical details—leading to increasingly creepy “this is too close to real” moments.
• At the end, three final saves are compared against two “real save” baselines. The twist: all three final saves were AI-generated.

Gameplay highlights / key observations from the AI attempts

Attempt 1 (FireRed-style test, minimal training)

• The AI generated highly abnormal Pokémon team/roster behavior—outlandish names and nonsensical traits.
• The creator describes the result as turning the game into “creepypasta.”

Attempt 2 (more training examples)

• Saves look more human at a glance:
  • Pokémon have “normal” real-like move sets and believable team composition.
• Still uncanny details remain, such as:
  • Suspicious items (e.g., large quantities of rare candies)
  • Odd naming choices

Attempt 3 (expanded knowledge + Hoenn region/Ruby usage)

• Reported as the closest to reality so far:
  • Many Pokémon appear believable and superficially match Gen 3 norms.
• Deeper checks reveal incorrect evolution/learnset details (e.g., evolution levels and move expectations that don’t match game data).
• The AI also uses the in-game mail system (word-based chat/letter constraints) in a way that terrifies the creator.

Attempt 4 (Sapphire + “Danny/Regirock” cave incident)

• The AI’s progress seemingly locks it into a specific in-cave encounter.
• The creator interprets the save as trapping them during Regirock content, including looping/locked-in behavior.
• A particularly eerie “save-data-only” interaction is shown/called out as extremely unsettling.

Attempt 5 (final smoothing and comparison challenge)

• The creator generates a final AI save to compare with two others (presented as human to a friend for blind testing).

Strategies / “how the AI was coached” (implicit method)

• Train the AI with examples:
  • Start with a single reference save, then expand to multiple saves/games.
• Provide “Pokémon 101” basics:
  • Teach expected Pokémon behaviors and what typical human progression looks like.
• Increase region coverage:
  • Add more cartridges/knowledge so the AI can model how players enter and explore new regions (specifically Hoenn).
• Add constraints:
  • Ask the AI not to scare the creator and to avoid obviously wrong artifacts.

Final blind comparison (how the creator tests “authenticity”)

• The creator downloads three saves and sends them to a friend (Colby) to shuffle identities, so the friend can’t tell which save is which.
• The friend plays them and reports observations:
  • One save: ~87 hours, 42 minutes; looks “normal enough,” with a plausible endgame team and items.
  • Second save: ~52 hours; extremely high money and lots of shinies/legendaries; suspicious item/value profile.
  • Third save: very short (~17 hours) with odd behavior:
    • Lots of potions
    • Suspiciously minimal/strange team progression
    • Relies on “refusing new moves” behavior (“Little Timmy method”).

Key tips / conclusions emphasized by the creator

• The AI can reach a stage where saves appear believable from the outside (team composition, progression pace, items).
• But believable doesn’t mean correct:
  • Detailed checks (movesets, evolution levels, expected learnsets, item quantities) still reveal inconsistencies.
• Even when a save seems normal, “human plausibility” can be faked using data patterns—and this can trick people.

Gamer(s) / sources featured (as named at the end)

• Colby (friend who played the shuffled saves)

• The creator references additional internal references/archetypes and past “player Danny” material, but only Colby is explicitly identified as a person used in the final test.