Maze doesn't build or train its own AI models. Instead, we partner with leading AI providers to power our intelligent features.
This document outlines how we use AI to enhance user experience, the measures we take to ensure data security and privacy, and provides answers to common questions regarding our AI functionalities, data handling practices, and compliance with data protection regulations.
In this article:
-
- Performance and limitations
-
Data handling and privacy
- What data elements are being consumed by MazeAI’s GenAI capabilities?
- Do you utilize data uploaded or generated during AI usage for training or fine-tuning?
- What actual data is sent to OpenAI’s services during AI functionality usage?
- AI Model Usage and Functionality
- What types of events and activities are logged by the AI features?
- How does generative AI work within your product, and are there other types of AI involved?
- Does Maze identify content generated by MazeAI?
- Data retention and compliance
- Security and architecture
- Evaluation metrics
- Contact information
Performance and limitations
Intended use cases for Maze AI features
- Rephrase research questions
- Generate follow-up questions for testers
- Summarize themes in responses
- Transcribe interviews
- Summarize transcripts
- Highlight key points in transcripts
- Group highlights into thematic summaries
External factors that may impact performance
- AI models used by Maze process only text and speech (no image or video inputs).
Known caveats and recommendations
- Outputs should be treated as suggestions, not definitive answers.
- Users retain control over outputs and can accept or reject AI-generated suggestions.
Data handling and privacy
What data elements are being consumed by MazeAI’s GenAI capabilities?
Refer to Maze’s AI and Privacy for detailed information.
Do you utilize data uploaded or generated during AI usage for training or fine-tuning?
No, Maze does not use customer data for training AI models.
What actual data is sent to OpenAI’s services during AI functionality usage?
Typically, prompts exclude PII like company names, user names, and email addresses.
For dynamic follow-ups, block titles and customer interactions are included.
In moderated features, transcripts may contain PII.
AI Model Usage and Functionality
What are the primary models and their sources in your product?
OpenAI’s GPT-4o (May 13, 2024) and GPT-4 Turbo (November 6, 2023), accessed via OpenAI’s developer platform.
Rev.ai’s Automatic Speech Recognition (ASR) for voice-to-text transcription. (Changelog)
Which AI models/providers does Maze use?
The AI features at Maze fall under two categories, each powered by specialized third-party providers:
1) Features powered by Large Language Models (LLMs) for intelligent data processing and conversation, including AI-moderated studies, Open Question Follow-ups, AI Themes, and Suggested Highlights.
These features utilize OpenAI and Anthropic LLMs. To access Anthropic models, we use Bedrock. To test, deploy, and monitor the performance of the features, we use Freeplay.
2) Features powered by Rev AI's automatic speech recognition (ASR) for voice-to-text transcriptions. This is used when transcribing Clips in unmoderated studies, and interviews in AI-moderated and Interview studies.
What types of events and activities are logged by the AI features?
For more information, see this link.
How does generative AI work within your product, and are there other types of AI involved?
Generative AI (OpenAI’s LLMs):
- Used for features like rephrasing research questions, generating follow-up questions, highlighting interesting points in interview transcripts, and identifying common themes.
- Most features are user-facing and triggered manually by Maze users, allowing them to edit or delete AI outputs.
- Participant-facing AI is limited to follow-up questions in open-ended surveys, which users can report if needed.
- All AI outputs are explicitly indicated as AI-generated.
Automatic Speech Recognition (Rev.ai):
- Limited to transcribing video/audio recordings to text, which can be manually processed or analyzed with other AI features.
Does Maze identify content generated by MazeAI?
Yes, interactions with GenAI are clearly marked with a UI element () to signal when users are interacting with AI services.
Data retention and compliance
What is the data retention policy for AI-related data?
- OpenAI and RevAI retain data for 30 days.
- Freeplay retains data for 90 days by default, extending to 12 months if flagged for quality inspections. Updates on retention periods will be reflected in our privacy documentation.
What is your contractual relationship with OpenAI and RevAI?
- Maze has established vendor agreements, including industry-standard service and data processing terms. These agreements align with SOC2 framework requirements and are detailed in our Vendor Management Program.
- For more information, see our Data Processing Addendum.
Security and architecture
Can you provide an architectural or data flow diagram?
A high-level network diagram is available on our Compliance Resources page.
Evaluation metrics
Are there documented performance metrics for the models?
Maze does not train its own models, so evaluation metrics (e.g., accuracy, recall) are not available.
OpenAI may have performance metrics, but they are not publicly shared.
Contact information
How can users reach out for AI-related concerns?
- For sub-processor contact information, refer to Maze’s Data Processing Addendum.
- For direct inquiries, contact Privacy@maze.design