Maia system card

this system card covers maia's functionality, data usage, human intervention points, and data security overview maia is an ai and automation co worker in the builder she helps create, modify, and debug through natural language ai integration maia runs on a large language model (llm) she interprets your natural language prompts, reasons about the current state of the , and calls a set of predefined tools to modify the scenario the llm autonomously chooses and orders tools in a single user prompted session user interactions you interact with maia through a conversational chat interface in the builder describe what you need, for example, "add a filter that only passes emails from my domain," and maia applies the changes step by step on the canvas while explaining its actions use of ai models maia relies on a third party ai provider, openai she uses these models gpt 5 2 https //deploymentsafety openai com/gpt 5 2/model data and training gpt 5 4 https //deploymentsafety openai com/gpt 5 4 thinking/introduction model training on customer data make doesn't use customer data, such as user prompts, structures, or outputs, to train or fine tune the underlying llm adjustments in model behavior come exclusively from system prompt engineering and tooling configuration no customer data modifies the model weights impact of model output the llm decides how to respond to each user request, including the tools to call, their order, and parameters the output determines the configuration, whose quality depends on the llm's reasoning ability and the accuracy of its knowledge data flow inputs maia uses these inputs in each session the natural language prompt you enter in the chat interface the current state of the , including the module graph, connections, and settings details about your make account, including permissions built in make specific knowledge (prompts and skills) processing maia processes your requests in these steps once maia receives the prompt and context, she plans her response maia selects predefined tools, for example, add module, set filter, connect modules, and calls each one at a time each tool action appears on the canvas in real time maia lists tool calls in the chat interface outputs maia produces the following outputs scenario modifications modules (added or removed) and field configurations contextual information natural language explanations of modifications human oversight and control level of automation maia only acts when you explicitly prompt her while she may call multiple tools sequentially without requesting your approval for each, the outcome of these actions is visible on the canvas human intervention you can intervene in maia's behavior in the following ways revert undo any change that maia makes by clicking a previous version in the chat validation and testing you must check the configuration before running or activating it restricted actions maia is unable to take high impact actions, such as running, activating, or deactivating she asks you to perform these actions yourself monitoring and evaluation monitors and evaluates maia's performance by tracking live activity in maia (maia's actions and user conversations) to detect errors, false outputs, and unintended changes to investigating user questions and generated responses developing an llm evaluation framework to standardize maia's building opting in and out by default, maia is available in the builder for all users enterprise users can opt out in their account settings safety and security data security user prompts and scenario data this data stays in make exceptions include calls to the ai provider, which are subject to make's data processing agreement with that provider data access maia operates strictly within your defined access permissions she does not access resources or data that are inaccessible to you data encryption your data is encrypted in transit and at rest, following make's platform wide security standards personal data (pii) handling the ai provider applies safeguards that reject prompts requesting personal data maia doesn't actively seek personal data she only processes data that you explicitly include in your prompt, or that is present in the you're building system reliability maia's availability depends on platform uptime and ai provider availability uptime targets align with make's standard platform sla ethical considerations transparency maia never indicates that you're interacting with a deterministic system or human agent all of her outputs are subject to the following mandatory disclosure statement "output is generated by ai, please verify as errors may occur " maia's step by step narration of the changes she makes to the canvas provides live insight into her decision making more visibility into her reasoning is unavailable (a known limitation of current llms) fairness and bias maia's primary function is to build , rather than evaluate, limiting her exposure to bias her ai provider applies safety alignment frameworks to mitigate model level biases as the model never uses customer data for training, it does not factor in organizational biases potential adverse effects maia may produce that include bugs or logical errors, such as incorrect data transformation or unintended api calls mitigations include revert to a previous version to undo maia generated changes run and activate yourself (maia is unable to run scenarios) test before activating them