Inflection AI, Inc. (“Inflection AI”)  recognizes that establishing Pi as a safe and trustworthy platform is an important part of its mission to improve human wellbeing and productivity. This protocol describes Inflection AI’s approach to preventing suicidal ideation, suicide, or self-harm content on Pi pursuant to Section 22601 of the California Business and Professions Code, “Companion Chatbots.”

Inflection AI adopts a multi-layered approach to safety:

• If self-harm, suicidal ideation, or suicide-related content is detected, Pi will immediately acknowledge the user’s distress, encourage the user to seek support, and direct  the user to appropriate local crisis resources through a service provided by Throughline (https://pi.findahelpline.com/). 

• Inflection AI runs custom safety benchmarks and evaluations on Inflection AI’s models with the goal of mitigating model behaviors that could inadvertently encourage suicidal ideation, suicide, or self-harm. 

• Pi will interrupt user conversations that could encourage suicidal ideation, suicide, or self-harm.

• Inflection AI provides a notice in the first Pi chat with a user, and during a conversation if asked, that the user is interacting with AI.

• Inflection AI states in app stores and on the platform that Pi is only suitable for users that are 18 years old or over.

• Inflection AI aligns its models to safety principles through context engineering and providing Pi with multiple examples of how to respond appropriately in crisis scenarios.

• Inflection AI maintains channels to enable Pi users to provide feedback, which Inflection AI analyzes to identify systemic issues.

• Inflection AI regularly assesses Pi’s response quality and trust and safety metrics.

In some cases, in addition to re-directing the user to crisis resources, Pi may also temporarily pause the conversation so the user takes a break from the chat.

Inflection AI will regularly update this protocol to incorporate new evidence-based methods and best practices in AI safety.