AI Ethics Policy

Back to the summary: How Studiorum uses AI

Studiorum uses AI to draft teaching materials, tutor students, and give a first read of student work, but teachers decide every grade, students are told when AI is involved, and student data is limited to what the work needs. This policy turns design choices already built into the app into commitments, and it names where the app still falls short.

In effect November 2, 2026. These commitments are Studiorum's policy now. An education-privacy attorney's review is pending, and the policy may change after it.

Purpose and limits

Anything Studiorum's AI can create, a teacher can also create by hand, from questions and activities to rubrics and feedback. The goal is teacher agency: AI supports a teacher's creative process and never takes it over.

Studiorum uses AI for five jobs, and a person stays in charge of each one.

• Teaching materials. AI drafts questions, activities, passages, and study content. Teachers review what they publish to their classes.
• A first read of student writing. AI applies the teacher's rubric and drafts a grade and feedback. The teacher decides the grade.
• Tutoring. The Magis tutor points students toward an answer without giving it away.
• Characters. AI plays characters in roleplay, debate, and chatbot activities.
• Support for teacher judgment. When a teacher asks, AI reviews a student's writing history or summarizes how a class is doing.

AI in Studiorum is never:

• the final word on a grade;
• a finding that a student cheated;
• the source of a score a student sees, since students see only grades a teacher approved;
• trained on student work;
• a counselor.

Teachers decide grades

Every AI judgment that affects a grade is a draft until a teacher signs off.

• Levels, not invented points. The AI judges rubric levels or required parts. The teacher's own conversion table turns those into points.
• Drafts don't count. A draft stays out of the student's score, the gradebook, mastery tracking, and the LMS.
• Bulk approval is allowed. Teachers approve row by row or accept many grades at once with Accept all. The teacher is accountable for every grade and chooses how closely to review. The feedback a student will receive is always visible to the teacher before approval.
• AI-decided credit is a draft too. Credit the AI decides during an activity, such as reaching a roleplay goal, waits for the teacher like any other AI grade.
• Teachers can correct the AI. A teacher can overrule the AI on one answer and apply that ruling across the class, with undo. Flags such as "Second opinion disagreed" never change a score on their own.

The release setting decides when an approved grade reaches the student.

What students are told

Students can tell when AI judged their work, wrote what they're reading, or is talking to them.

• Approved grades. When a teacher approves a grade the AI drafted, the student sees "Your teacher started from an AI draft of this grade and feedback and approved the final version." The teacher is the source of every grade and comment and decides how much AI help to use. The word is "approved," not "reviewed," because a teacher may approve in bulk.
• AI chats. Roleplay, debate, and chatbot conversations show a label saying the character is played by AI.
• Writing history. Before typed work, the activity says "Studiorum saves a history of how you write your typed answers, like when you paste text and how fast you type." It tells students their teacher can play that history back and watch each answer being written. It adds that the teacher may ask AI to review that history, but a review alone can't prove a student didn't write the work.
• AI-written passages. A passage the AI wrote is labeled as AI-written (see Sources and AI-written content).
• AI tutor and feedback. The Magis tutor labels its replies "AI tutor," and self-check feedback is labeled "AI feedback," so students know both come from AI.

AI that students use

Student-facing AI helps students think; it doesn't do their work or grade it.

• No answers. Magis asks one question at a time and doesn't give the answer. Asked to write an assignment, it helps the student think it through instead.
• No scores. Feedback to students never shows or hints at a score. On full-credit work it names strengths without calling the work "ready to submit." A code check backs up the prompt.
• No false praise. Magis can't call an answer correct unless the graded feedback supports it.
• Not during tests. Magis isn't offered on a live summative test or to guests.
• Claude only. Every student feature runs on Anthropic's Claude. Google's Gemini API terms require users to be 18 or older, and Studiorum students are 13 and up. Read-aloud for students uses the browser's built-in speech.
• Reading level. Students pick their own reading level in self-study. Teachers set it for content they assign.

Teacher tools, including the AI's first read of student work, run on Gemini. Studiorum refuses Gemini to students who come through a school's LMS. An account that signs itself up as a student can still open the teacher tools, because some of those accounts belong to teachers trying the student side first. A student who signs up on their own and opens the teacher tools uses Gemini, which violates Google's Gemini API terms of service and Studiorum's Terms of Service if the student is under 18.

Student safety in AI chat

Studiorum's AI is a study tool, not a counselor, and it points students in crisis to people who can help.

If a student mentions self-harm, suicidal thoughts, abuse, or danger, the AI steps out of tutoring or out of character. It replies briefly and with care, and encourages the student to talk to a trusted adult. The app then shows a fixed card:

• 988 Suicide & Crisis Lifeline: call or text 988.
• Crisis Text Line: text HOME to 741741.
• Call 911 if you're in immediate danger.

This rule runs ahead of every student chat, including chatbots a teacher writes, and a teacher's instructions can't remove it. Studiorum doesn't monitor these conversations, doesn't alert anyone, and never records that the card appeared.

What is saved matters here. Magis tutor conversations aren't saved anywhere; they exist only in the open browser tab. Roleplay, debate, chatbot, and conversation-practice chats in a teacher's activity are saved with the submission, because they are the assignment. A teacher grading that work can read everything the student wrote, including a disclosure.

AI characters

Characters played by AI follow three rules that an activity's instructions can't change.

• Labeled. Every roleplay, debate, and chatbot conversation shows that the character is played by AI.
• No living people. AI never plays a real person who is alive or may be alive. When an activity asks for one, the AI plays a fictional composite in the same role, with an invented name, and says so once.
• Hateful views, handled honestly. A historical figure who held racist, sexist, antisemitic, or other hateful views stays historically accurate. The character may argue for those views so students can practice answering them. It never uses slurs, and Studiorum never endorses the views.

These rules apply the same way in self-study as in teacher-assigned work. Claude may still decline an argument, and Studiorum doesn't try to override that.

Academic integrity

Studiorum gives teachers evidence for a conversation with a student, never a verdict.

• Writing history. Typed answers record the whole writing process, including pastes, typing pace, and revisions, and teachers can play it back from start to finish. Students are told before they start.
• Paste flags (a mark on any answer where 200 or more characters appeared at once) are plain code. Simple rules flag large pastes, not AI. Moving your own text around isn't flagged.
• AI review of writing history only when a teacher asks. A teacher can ask AI to review each answer's writing history for evidence of AI misuse or plagiarism. This review is not automatic and cannot be run by students.
• Evidence sets the ceiling. The hard evidence caps what the AI's review can conclude. The review treats dictation, typing tools for other languages, and assistive technology as legitimate. It also treats polished writing as no sign of cheating.
• Shared only by the teacher's choice. A student sees the review of an answer only if the teacher shares that answer's review. The student then sees the same caution teachers see: a review alone can't prove a student didn't write the work. No number is shown.
• No automatic penalties. No flag or review changes a grade by itself. The Terms bar teachers from using a signal as the sole basis for discipline. Teachers should use the information as an opportunity for discussion with the student according to their personal policies for academic integrity.

Fairness and accuracy

Studiorum measures how well AI grading matches expert readers and says plainly where it hasn't measured.

• Strictness fits the course. College Board-level strictness applies only to AP^® prompts. Other work is read fairly, and spelling and grammar cost points only when the rubric grades spelling and grammar.
• Partial credit. Short answers and rubric-level grades earn credit in quarter points, so a partly right answer isn't scored all or nothing. This is a variation from College Board grades which are often single rubric numbers, but give students better indication of opportunities for growth.
• Hidden names. Teachers can grade with student names hidden and answers shuffled.
• Published accuracy. For AP^® courses, AI grades are compared with College Board scored samples from exam years the prompts weren't tuned on. Studiorum publishes how often the AI lands within one point, with sample sizes, and labels each figure with the model and date.
• Model changes. Every model change is checked against a sample of those answers. A full re-measure follows a change of model family, and happens once a year. Choosing the model stays Studiorum's decision.
• What isn't measured. Non-AP^® grading has no comparison with human graders. There, the teacher's approval of each grade is the safeguard, and teachers can steer the AI. They can adjust the AI's instructions for one question, or set profile rules that shape every AI prompt they run.
• English learners. Studiorum tests whether AI grading lowers scores for features of English-learner writing that the rubric doesn't grade. A pilot on 30 AP^® History answers in October 2026 found no penalty. A full run across subjects isn't scheduled yet.

Sources and AI-written content

AI-written source material is always labeled, and the AI is told never to invent a source.

• Real sources first. When content needs a primary source, Studiorum looks for the real text, for example on Wikisource, before AI writes a stand-in.
• Labels. Every passage carries one of the labels in the table below. Self-study marks unverified sources the same way the teacher side does.
• Nothing invented. Generators are told never to invent a quotation, document, source, or catalogue number. Links the AI makes up are removed.
• Simplified text. When a teacher lowers a passage's reading level, the original stays word for word, and the simplified version is labeled.
• Fact-checking. Guided Project answers are checked against web search, and the findings go to the teacher.
• Images. Images come from Wikimedia Commons with the artist recorded, never from pages that rely on fair use.
• Teacher Published Activities. An activity a teacher publishes is the teacher's content, even when AI drafted it, because the teacher approved it. No AI generation tag is used in this case.

Student data and AI providers

Studiorum sends AI providers what a task needs and never a student's name or account details.

• Providers. Google Gemini runs teacher tools, and Anthropic Claude runs student features. Both work under commercial API terms that bar training on this data. Studiorum keeps copies of those terms on file and re-checks them every year.
• Keys stay on the server. The browser never holds a provider key.
• Ages 13 and up. Student features are for ages 13 and up. A neutral age screen checks the birth year and then discards it.
• No tracking. YouTube videos play in privacy-enhanced mode, and Studiorum runs no advertising or analytics code.

How each commitment is enforced

Some commitments are enforced on Studiorum's servers, some only in the app, and some by the Terms of Service. This table says which.

Going around the app violates the Terms of Service. That includes calling Studiorum's servers directly, switching the AI model, or reading data the app doesn't show. Students and parents keep their right to request education records, including AI notes, through support.

AI tools we use to build Studiorum

Studiorum is built and supported with AI coding tools, and student data gets the same care there as in the product.

• Only to fix a problem. Student data is opened only to fix a specific problem, and only as much as the fix needs.
• No names left behind. Names and identifying details stay out of notes, commit messages, code comments, test data, and AI tool memory. A problem is described as "one student's essay in an activity," never by name.
• No copies kept. Nothing is kept once the fix is done.
• Test data isn't student data. Grading tests use public College Board scored samples and invented cases, not Studiorum students' work.

Cost, environment, and accessibility

Students never pay for assigned work, AI's environmental cost is weighed in decisions, and accessibility claims wait for an audit.

Cost. Work a teacher assigns is billed to the teacher's plan, never to the student. Students who study on their own use their own weekly allowance, which is twice the standard free allowance and students can purchase personal plans for self-study.

Environment. Studiorum weighs the estimated environmental impact of AI alongside cost and accuracy when it chooses models and features. Neither Google nor Anthropic publishes per-call figures, so Studiorum models them along with its methodology. Published per-prompt figures are split into per-token rates and combined with grid carbon intensity. Results are low–high ranges for energy, carbon, and water, not measurements, and unverified factors are marked. Administrators of schools with a school-wide license see their own school's estimates and environmental impact data is available upon request.

Accessibility. Students can set reading level, have text read aloud by the browser, use a line reader, and change text size and contrast. The app respects reduced-motion settings. Integrity checks treat assistive technology as legitimate. An accessibility audit is under way, and Studiorum makes no conformance claim until an audit supports one.

Review and changes

This policy is reviewed every August and before any change that would make it untrue.

• Yearly. Each August, together with the security program review and the College Board course audit.
• Before material changes. Before a new model family, a new student-facing AI feature, a new kind of data sent to a provider, or a provider terms change.
• Legal review. An education-privacy attorney reviews the policy before Studiorum signs district contracts.
• Dated changes. Each revision carries its date, and teachers are told about material changes.
• Owner and contact. Brian Tittl, BYTSized LLC, is responsible for this policy. Questions go to support@studiorum.io.