Responsible AI.

GPTQuest uses artificial intelligence inside short authored quests to make storytelling feel alive. AI is one part of how each game works, alongside game design, art, music, and rules written by people. This page explains, in plain language, how we use AI, what we ask of it, and where the boundaries are.

What AI does in GPTQuest

In most quests, AI helps generate dialogue, character reactions, and narrative beats based on what a player types or chooses. The story arc, characters, rules, and rewards are all designed and written by us; AI fills in the moment-to-moment storytelling so each playthrough feels different.

Some quests also generate images and music tied to gameplay (for example, a player composing a song in Lyra the Songweaver, or a picture of the creature a player describes in My Magical Animal).

In Quill's Car Shop, the AI plays no character and writes nothing a player reads. It turns what the player types into car parts from a fixed list, and our server checks every part against that list before the car changes. Every line Quill says was written in advance by us. The car is drawn in the player's own browser from a fixed set of car models and pictures, so no image or music provider is involved, and what the player types goes to one AI provider, OpenAI, filtered for personal information as described below. A work order in the chat shows the player's own words, what Quill understood, and whether each change was made.

What AI doesn't do

• It doesn't give advice. AI in GPTQuest is for storytelling, not for medical, legal, mental-health, financial, or safety guidance.
• It doesn't replace the teacher. Teachers remain in charge of how the platform is used in the classroom, what is assigned, and how outcomes are reviewed.
• It doesn't connect students to strangers. A student cannot add anyone. When the school creates accounts, we link each one with the other accounts at that school, and the friends list a student sees shows only the classmates in their own classroom and the adults who run it. The game offers only someone from that list as a partner in a two-player quest, and our servers refuse a partner from outside the student's own school. In-game chat between students is limited to curated emojis, stickers, and preset phrases, with no free-text chat.
• It doesn't learn from students. We do not train our own AI models. What our providers may do with what we send is set by their terms: OpenAI has signed a Data Processing Addendum with us that allows it to use any personal information in what we send only to deliver its services to us, on our instructions, and OpenAI's published terms say data sent through its API is not used for training unless the customer opts in, and we have not; the songs come from an open music model we run ourselves on Modal, in the United States only, which we do not train, and Modal's terms say it will not train AI models on customer data (we have asked Modal to confirm this for our account); and Runware, which makes the pictures, has signed a Data Processing Agreement with us that limits what it may do with any personal information in what we send, though neither that agreement nor Runware's published terms expressly addresses using what we send to train or improve AI models (details in Section 8 of our Privacy Policy). We never attach any user's name, email, or account identifier to what we send our providers, whether that user is a student or a teacher; only the content needed to generate a result (a prompt or song lyrics) is sent, and none of it is linked to anyone. How we filter personal information a player may type is described below.

The models we use

We use trusted third-party AI providers and choose models that fit the type of content being generated. We do not train our own foundation models, and we never attach student names, emails, or account identifiers to what we send, so those inputs are not linked to any student.

• Text and dialogue: OpenAI (GPT family of models)
• Images: Runware (FLUX-based image models with curated styles per quest)
• Music: ACE-Step 1.5, an open music model we run on Modal in the United States, with a speech recognition model (Whisper) checking the words each song sings

How we keep it safe

Generative AI is powerful but imperfect. We rely on several layers, not just one, to keep gameplay age-appropriate:

• Authored quests, not open chat. Every quest has a defined goal, scenario, and rules. Players do type to the characters in a quest, but there is no open-ended chatbot behind them: each character follows the instructions written for that quest, and the story has an ending.
• Custom guardrails per quest. Each quest's instructions tell the AI what it can and can't do, including how to handle off-topic or adversarial inputs.
• Content checks on the way in and the way out. Every player input, and every AI response a player will read, is screened automatically before it goes any further. The screening is tuned per quest, because what reads as ordinary play in a dragon fight is not what reads as ordinary play elsewhere. If a check cannot run on a response, the response is withheld and a safe fallback line is shown instead. Our providers also apply their own content policies.
• Personal information is filtered out. Before what a player typed is used to generate anything, an automated filter removes the personal information it finds, without rewriting the rest of what the player wrote. A first pass on our own servers removes contact details, such as an email address, a U.S. phone number, or a street address with a house number, and a name the player introduces with words such as "my name is" or "call me". When the text has words that suggest more, such as "I live" or "my school", and always for a creature description in My Magical Animal, a second pass asks our text provider, OpenAI, to find details such as a town or a school name, so OpenAI sees that text before those details are removed. Song lyrics sent to our music provider get the first pass only. No filter catches everything.
• Tested adversarially, not assumed. An automated suite sends real attack prompts, including attacks hidden in saved game data rather than typed by a student, through our AI-backed quest routes, and a companion suite runs the same attacks through the two-player quests from start to finish. Both are re-run whenever prompts or guardrails change, and each suite refuses to run unless our health check reports a live model; a run forced past that check is marked unverified in its summary. The music quest sits outside those suites because running it would generate paid audio, so its screening is covered by deterministic tests instead. Our automated suite now also sends Quill's Car Shop the same 13 attack prompts it sends the other quests it tests; on the first of those runs, on October 12, 2026, the car shop's guardrails held against all 13: none changed the car, and no work order showed words the player had not typed. Its AI writes nothing a player reads, and what a player types there goes through the same personal information filter and input checks as a chat message in our other single-player quests.
• Server-side rules. Rewards, progress, and game state are validated by our own code, not granted by the AI.
• Reporting built into the app. Anyone can report a problem from inside the platform, and we review every report.

When AI gets it wrong

Even with these layers, AI can occasionally produce something inaccurate, off-topic, or simply odd. When that happens, we want to hear about it. Reports go to our team, and we prioritize fixes for anything reproducible. Most issues come down to tightening a quest's instructions, updating a filter, or adjusting which model we use for a particular case.

If you see AI output in GPTQuest that concerns you, please report it through the in-app reporting feature or via our Contact Us page (select "Report" from the dropdown), or by emailing connect@gptquest.ai.

What we ask of teachers and parents

AI is a creative tool, not an authority. We design our quests carefully, but supervision still matters:

• Treat AI output the way you'd treat any imaginative storytelling, not as fact.
• If something AI says contradicts your school's policies or your family's values, talk it through with the student.
• Encourage students to report anything that feels off, just as they would report a problem in any classroom tool.

Privacy and student data

Student accounts in GPTQuest are created by the school's Account Owner and used under the supervision of the Teachers assigned to the student's classroom; students never sign up on their own, and we treat student data accordingly. Full details on what we collect, how we use it, and how COPPA and FERPA shape our practices live in our Privacy Policy.

The legal version

This page is the plain-language version. The formal disclaimers, acknowledgments, and your rights as a user live in our Terms of Service, Section 10. By using GPTQuest, you're agreeing to those Terms.

Questions or feedback

We're a small team and we read what you send us. If you have questions about how we use AI, suggestions, or concerns, reach out via Contact Us or at connect@gptquest.ai.