Our approach to building CoWrite
Jennifer Funk avatar
Written by Jennifer Funk
Updated over a week ago

Overview

CoWrite is an AI writing assistant that helps you create text that frames, contextualizes, and/or highlights key insights in your data visualizations. It’s designed to quickly produce a first draft of text to review and refine before sharing with external audiences. The goal is to help save time while also better supporting your audience’s understanding of the data you share.

While we plan to expand CoWrite’s capabilities over time, it is not currently well suited for the following tasks: doing math to find insights in your visualization, generating alternative text for site accessibility, completing outputs in other languages, or answering questions about a particular data finding (e.g., why unemployment has increased or decreased).

Our approach to building CoWrite

  • We decided the best option for our customers and business is to build CoWrite (and upcoming features) by integrating with OpenAI.

  • CoWrite brings the power of ChatGPT to generate text with a single prompt directly into your data storytelling workflow in mySidewalk.

  • CoWrite leverages our meticulously sourced, cleaned, and structured data library to generate data-driven outputs via a technique called in-context learning. When you submit a CoWrite query, mySidewalk passes a summary of your data visualization to OpenAI to help facilitate data accuracy. By doing this, when you generate text with CoWrite, it is more likely to reference the data we provided than to manufacture its own, reducing the likelihood of hallucination.

  • In addition to facilitating accurate data stories, we use prompt engineering techniques to steer CoWrite toward generating outputs that meet our equity and inclusion standards. Examples of these include the use of plain language, inclusive language, appropriate reading levels, and a focus on stories that emphasize structures, not struggles.

  • Meeting these standards will be an ongoing process, given the existing limitations of language models — but we’re committed to it. Iterative improvement will happen based on customer feedback and the emergence of new research, techniques, and tools.

  • We’ve implemented safeguards like content moderation to prevent egregious misuse and (soon) evaluations to monitor performance on key criteria like readability and inclusive language.


Read our guiding principles for building responsibly with AI


FAQs

  • Is mySidewalk training its own language model?

    • No, the CoWrite feature uses one of OpenAI’s state-of-the-art instruction-tuned models and a combination of techniques like in-context learning and retrieval augmented generation to guide behavior for specific use cases.

  • Is mySidewalk or OpenAI storing data from CoWrite submissions?

    • mySidewalk stores inputs and outputs associated with the CoWrite feature. Occasionally, we will review this data to assess its overall performance and iterate on prompts to improve results. However, we are not using this data to train our own language model.

    • As of March 1st, 2023, OpenAI retains user data for 30 days, but no longer uses data sent via the API to improve their models. Here’s their data usage policy.

  • What does it mean that CoWrite is “built on data you trust”?

  • What does it mean that CoWrite has an inclusive and accessible voice?

    • CoWrite is designed to generate outputs that make data more accessible to people who aren’t data experts. We manage the quality of outputs through various prompt engineering techniques and tools. As we do so, we're especially focused on improving CoWrite's outputs on key criteria, including length, readability, the use of plain and inclusive language, factuality, and more.

  • What should I do if CoWrite generates inaccurate, offensive, biased, or otherwise concerning text?

    • While mySidewalk is monitoring outputs, we also ask that you share anything that concerns you with our team. Contact your customer success representative or email us at [email protected].

  • How can I provide feedback to mySidewalk about this feature?

    • There are lots of ways to connect with us! Generally, it’s whatever is most convenient and comfortable for you. That said, every time you use CoWrite to generate text, a link to a feedback form will appear below the text. Please use this form to share your thoughts.

  • How can I be sure to use this tool responsibly?

    • Make sure you use CoWrite in accordance with your organization’s guidelines around the use of AI.

    • Manually review each AI-generated result for accuracy, appropriateness, and compliance with your organization’s guidance on acceptable use.

    • Let audiences for your data stories know if the content was generated by AI.

    • Inform mySidewalk of AI-generated results that include inaccuracies, biased ideas or language, or otherwise concern you for any reason.

Did this answer your question?