Skip to the document
Tidal AI

Version 2026-10-08 · effective [[TO FILL: effective date]]

AI Transparency

How the AI in Tidal AI works, what it can and can't do, how you stay in control, and how AI-generated content is labelled.

You are working with AI

Everything a bot says or makes in Tidal AI is generated by artificial intelligence. Bots are AI agents, not people, even though they have names, personalities and creature bodies. Tide, the orchestrator in every workspace, is an AI too.

How it works

  • Models. Bots are powered by large language models made by third parties (for example Anthropic, OpenAI, Google, Meta, Mistral, xAI and DeepSeek). You or your bot choose a model or a tier (Swift, Balanced, Genius). Each model page in the app names its maker.
  • Agents. A bot reads your request, plans, and uses tools: searching the web, reading and writing files, running code in an isolated machine, calling services you've connected, sending email, building and deploying apps. It loops (think, act, check) until the task is done or it needs you.
  • Missions. Bigger jobs become missions with a plan, steps, a budget in credits, and checks that verify the result. Checks reduce mistakes but don't guarantee correctness.
  • Memory. Bots share a workspace knowledge graph (the "hive brain") of facts they've learned, with where each came from. You can see and edit it.

Limits you should know

AI output can be wrong, incomplete, out of date or biased, can misread what you meant, and can sound confident when it isn't right. It can be misled by content it reads (for example a web page or email containing instructions), which is why content from outside your workspace is marked untrusted and risky actions need approval. Check important output before you rely on it, send it or publish it. AI output is not professional advice.

How you stay in control

  • Approvals. By default, a bot must ask before it emails or messages anyone outside your workspace, deletes things, deploys to production, or takes financial actions. Content a bot has read from outside makes those actions ask for approval even if you'd allowed them.
  • Budgets. Every run and mission has a credit budget with a pause-and-ask point and a hard stop.
  • Stop. You can stop any run or mission at any time.
  • Audit log. Every tool call, approval and change is recorded with who (or which bot) did it.
  • Secrets stay secret. Passwords and keys for services you connect are stored encrypted and are never shown to the AI models.

How AI content is labelled

  • Email from bots always ends with a footer stating that it was sent by an AI assistant operated by your workspace via Tidal, and carries an X-AI-Generated: true header and an unsubscribe link. The footer is added by the platform after the bot's text, so a bot can't remove it.
  • In the app, bot messages are shown as coming from the bot, never from you.
  • Chat widgets that workspaces put on their own websites introduce the bot as an AI assistant [[TO FILL: confirm the widget's greeting always discloses AI; product change if not]].
  • If you publish AI-generated text, images, audio or video to the public, laws such as the EU AI Act (Article 50) may require you to disclose that it was AI-generated, in particular for realistic images, audio or video of people ("deepfakes") and for text published to inform the public on matters of public interest.

Your data and the models

We send models only what a step needs. We ask model providers not to keep or train on it, and we don't train models on your content either. Details are in our Privacy Policy and Subprocessors.

Questions and concerns

If an AI output caused a problem, or you think a bot behaved in a way it shouldn't, tell us at [[TO FILL: support contact email]].