AI in Inkblots

Large language models (LLMs) are very useful, but they also make mistakes. You should always double-check the outputs of all tools based on LLMs, including the Inkblots Agent. Inkblots, LLC makes no claims or guarantees as to the accuracy or reliability of the output of the Inkblots Agent.

This article provides an overview of how Generative Artificial Intelligence models are used in the Inkblots application. It will be kept up-to-date with any changes to the underlying models used by Inkblots. This document serves as an addendum to our Terms of Service and Privacy Policy documents; if anything in here contradicts either of those, you should consider those to be authoritative.

How Inkblots Handles Your AI Usage Data

Inkblots, LLC retains the following pieces of information about each conversation you have with the Inkblots Agent:

  • Date and time

  • Number of turns

  • Tokens consumed

  • Number of tools called

  • Number of web searches performed

  • Any tool call errors that occurred

Full transcripts are stored so that you can view and continue previous conversations in the Inkblots app, but we do not view those transcripts without your explicit consent. You'll be directly asked to provide that consent when giving feedback or otherwise talking with us. We do not train any type of AI model on your data, and we don't work with AI service providers that do so either.

The Inkblots Agent

The Inkblots Agent is based on the reasoning large language model Claude Haiku 4.5 (system card PDF), served over Anthropic's API. According to their policies, Anthropic does not train models on inputs to their APIs, and all inputs and outputs are deleted within 30 days of receipt or generation.

Additional AI Models

In addition to the model powering the Inkblots Agent, the Inkblots application may use additional AI models for other purposes.

Currently, the only other model used is the open-weights language model GPT-OSS-20b (model card PDF). This model is used for inferring the short titles used to label conversations. We currently access this model via OpenRouter, so the actual model inference may be performed by a variety of providers. We've disabled logging on OpenRouter, and only allow OpenRouter to route these requests to one of the following providers, each of which doesn't retain any data and doesn't train on inputs:

  • Groq

  • Azure

  • Google Vertex

  • Baseten

  • Parasail

  • Fireworks

  • Amazon Bedrock

Was this helpful?