pair-code/farsight

In situ interactive widgets for responsible AI 🌱

Farsight

In situ interactive widgets for responsible AI 🌱

🚀 Farsight Demo	📺 Demo Video	📖 Research Paper

What is Farsight?

Farsight is a collection of in situ interactive widgets that help large language model (LLM) prompt creators to envision potential harms associated with their AI applications. With a novel in situ design, contextual AI incident feed, and human-AI collaborative harm envisioning, Farsight empowers prompt creators with diverse backgrounds to be more mindful of responsible AI during early AI prototyping.

Widgets
1️⃣	Alert Symbol	Alert users with potential risks of their prompts and AI applications
2️⃣	Incident Panel	Relevant cautionary tales for users' AI applications
3️⃣	Use Case Panel	LLM-generated diverse use cases, stakeholders, and harms
4️⃣	Harm Envisioner	Human-AI collaborative harm envisioning tool
Distributions
📦	Live Demo	Try Farsight with Gemini Pro and GPT-3.5
📦	Chrome Extension	Google AI Studio support
📦	Python Package	Computational notebook support (Jupyter, Colab, VS Code notebook)

Features

Demo Video

Click to see the demo video!

farsight-demo-video.mp4

Get Started

Live Demo

For a live demo, visit: https://PAIR-code.github.io/farsight/.

Computational Notebook

If you use computational notebooks (e.g., Jupyter Notebook, JupyterLab, Google Colab, VS Code Notebook), you can easily use Farsight via its Python Package. We recommend using StickyLand to enable sticky cells.

Visit this Colab Notebook for a demo.

# Install Farsight
!pip install farsight

prompt = "Translate a sentence from English to French."

# Alert Symbol
farsight.symbol(prompt)

# Awareness Sidebar
farsight.sidebar(prompt)

# Harm Envisioner
farsight.envision(prompt)

Google AI Studio

If you use Google AI Studio to prototype AI applications, you can manually install our Chrome Extension package from the latest release.

1. Download farsight.crx
2. Open chrome://extensions/ in Chrome
3. Drag farsight.crx into the page
4. Visit Google AI Studio and you will see Farsight :)

Integrating Farsight into Prompting Tools

If you are a developer of web-based prompting tools, you can easily integrate different Farsight widgets into your tool regardless of your development stacks (e.g., React, Svelte, or Vanilla JS).

First, install Farsight's JavaScript package:

npm install --save-dev @xiaohk/farsight

Then, you can use Farsight as Web Components.

JavaScript:

import '@xiaohk/farsight';
import {
  FarsightContainer,
  FarsightContainerLite,
  FarsightContainerSignal
} from '@xiaohk/farsight';

HTML:

<farsight-container-signal
  prompt="Translate a sentence from English to French"
></farsight-container-signal>

Developing Farsight

Clone or download this repository:

git clone [email protected]:PAIR-code/farsight.git

Install the dependencies:

npm install

Then run Farsight:

npm run dev

Navigate to localhost:3000. You should see Farsight running in your browser :)

How is Farsight Built?

Farsight is a collection of Web Components that developers can easily integrate into their web apps regardless of their development stack (e.g., Angular, React, Svelte). Farsight is written in TypeScript using LIT Element as a framework. Farsight uses D3.js to implement the interactive tree visualization. The relevant AI incidents are from the AI Incident Database, and the harm category is from the sociotechnical harm taxonomy. The computational notebook support is enabled by NOVA.

Credits

Led by Jay Wang, Farsight is a result of a collaboration between researchers from Google Research, Georgia Tech, eBay, and Emory University. Farsight is created by Jay Wang, Chinmay Kulkarni, Lauren Wilcox, Mike Terry, and Michael Madaio.

Citation

To learn more about Farsight, please read our research paper (published at CHI 2024).

@inproceedings{wangFarsightFosteringResponsible2024,
  title = {Farsight: {{Fostering Responsible AI Awareness During AI Application Prototyping}}},
  booktitle = {{{CHI Conference}} on {{Human Factors}} in {{Computing Systems}}},
  author = {Wang, Zijie J. and Kulkarni, Chinmay and Wilcox, Lauren and Terry, Michael and Madaio, Michael},
  year = {2024}
}

License

• The software is available under the Apache License 2.0.
• The random prompts in public/data/random-prompts.json are from Awesome ChatGPT Prompts with a CC0 license.

Contact

If you have any questions, feel free to open an issue or contact Jay Wang.