ISAI AI Transparency Policy

1. What ISAI is

ISAI is a support tool for editing documentaries and other works made from recorded material. It reads transcripts, scripts and notes, and returns organization, answers and analyses of that material. ISAI does not create content from nothing: everything it produces starts from what you uploaded.

2. Where AI is used

2.1. Transcription: in this first phase ISAI does not transcribe audio or video. You bring the transcript made in your editing software, and ISAI works from it.

2.2. Search and questions about the material: ISAI finds passages by meaning, not only by exact word. Asking about "alcohol" can return a passage about beer. To find passages, ISAI uses a search model from OpenAI, which converts each passage into numbers and compares meanings. The written answer comes from Anthropic's language models (Claude).

2.3. Editorial analyses (themes, characters, structure, contradictions, gaps): generated by Anthropic language models from your material, guided by editorial methods developed by ISAI.

2.4. There is no AI in billing, sign-in or decisions about your account. No automated system makes any decision with legal effect on you.

3. What ISAI does not do

It does not create the voice, face, image or synthetic speech of any person. It does not generate scenes, interviews or testimonies that do not exist in your material. It does not train AI models on your material, and the model providers we use are prohibited by contract from doing so with what we send them. It does not show your material to other users or use it in other projects.

4. What can go wrong, and what we do about it

4.1. Inexact quotes. When a language model "quotes" a line, it may rewrite it, merge two lines, attribute it to the wrong person or give it a timecode that does not match. This has happened in ISAI's internal tests. That is why the Terms of Use ask you (section 3.3) to check every quote against the source, by timecode, before using it. We are building a mechanism through which every quoted line is copied from the original file instead of generated by the model; when it is live, this Policy will be updated with the date.

4.2. Facts, names and dates. The model may confuse people with similar names, swap dates or fill gaps with guesses. Treat any name, date or fact as a hypothesis until you check it.

4.3. Transcription. ISAI works from the transcript you bring and inherits its errors, which are more common with accents, noise, several people talking at once and rare terms. Proper names are the most fragile point.

4.4. Interpretation. Character and theme analyses are readings, not diagnoses. Whenever possible, ISAI separates what is in the text from what is inference, and avoids categorical labels about real people. Even so, the final reading is yours.

4.5. Coverage. If the material is large, the model may not "see" all of it when answering a question. An answer that says "there is no mention of X" does not prove that X is absent from the material.

5. The human role

On ISAI, the tool analyzes and you decide. No result from the Platform is final; all of them are input for the judgment of whoever is editing. That keeps authorship of the work with you, and it is what the Terms of Use assume.

6. How we identify what is generated

Every result returned by the Platform is identified as generated or located by AI and, whenever it refers to the material, carries the reference (file and timecode) that lets you check it. When a result is exported, that identification goes with it.

7. Which models and providers

Anthropic (Claude models), via API, for questions and analysis. OpenAI, via API, for meaning-based search. The exact model versions change over time; we keep an internal record of which version processed each project and share it with anyone who asks. Data goes to servers in the United States; the Privacy Policy covers international transfers.

8. How to report an error

Write to flipcorreio@gmail.com with the project name, the question or function used, the result you received and, if you can, the correct passage with its timecode. The person responsible for the Platform reads every reported error and replies.

9. Standards behind this Policy

In Brazil: the General Data Protection Law (Law 13,709/2018), the Internet Civil Framework (Law 12,965/2014) and the Consumer Defense Code. For users in the European Union: the GDPR and, to the extent they apply to a service that uses third-party general-purpose models, the transparency obligations of the EU AI Act (Regulation 2024/1689). Brazil is debating a specific AI law; when it takes effect, this Policy will be reviewed.