Introduction
At Pencil, your data belongs to you. Pencil does not use any user-provided data - including uploaded artifacts, entered prompts, or agent creation information - to develop or refine AI models outside of the user’s own workspace.
We maintain a strict 'no train' policy: the data you provide is never used to train or refine AI models for other customers. Our approach ensures your workspace data remains private, secure, and exclusively yours.
Below is a detailed breakdown of how we handle different types of inputs and how they are used within the platform.
Assets (Uploaded Files)
Assets are never used to train or fine-tune AI models.
You may share assets with AI models only for grounding and inferencing within your own workspace.
Grounding is information passed to the AI model as context and reference material to help improve the accuracy of the answer. In Pencil, grounding uses Retrieval-Augmented Generation (RAG), ensuring responses are based directly on the files you’ve provided.
Inferencing is the process of using an already 'trained' model to generate a response. This process does not enhance the knowledge of the model nor does it create a new model.
Prompts (Text Inputs)
Prompts are never used to train or fine-tune AI models.
Prompts are sent to the AI model only for inference - to generate a relevant response based on your specific query.
Prompts are almost always text input by the user, to provide additional context information related to grounding, images or videos the user may have uploaded.
Agents
An Agent is an LLM model tasked with instructions to perform a specific job and respond in a specific manner/style. Its user-defined customisation layer that sits on top of a pre-existing AI LLM.
Neither prompts nor uploaded assets are used to train or fine-tune AI 'Agent' models.
Such inputs can be used to ground the AI model powering your agent, but only within your workspace.
Grounding ensures the agent’s responses follow your instructions and reference your provided knowledge or files.
Data from Connected Ad Accounts
Ads (images/videos) pulled from connected ad accounts may be used to train dedicated performance prediction models. We only use the data from the customer’s ad account together with pooled data from other Pencil Standard customers to build a media prediction model that is specific to that customer.
These models are not shared across customers and remain unique to your workspace.
Custom Models (LORAs)
In addition to the above, Pencil also allows customers to create their own Custom Models using the platform’s fine-tuning option.
Customers can upload their own images into the platform to build a custom-trained model.
These custom models belong entirely to the customer and remain within their workspace.
Once created, they can be used to generate images using the custom-trained models.
No other training or fine-tuning (and especially nothing automatic) takes place inside the platform beyond this explicit customer-driven process.
Consistent Policy Across the Platform
Pencil's behaviour regarding data usage and model development remains consistent across different upload points or types within the platform, always adhering to the above stated principles and the 'no train' policy. In other words, Pencil's behaviour and policy adherence is consistent regardless of what and where you upload assets or input prompts.