Logo and icon for Aligned, a ChatGPT plugin with description: Help the World -- automatically share LLM misbehaviors with the community! Aligned is the global platform for alignment.. Find out more on Plugin Surf, the best plugin database.

Aligned

Help the World -- automatically share LLM misbehaviors with the community! Aligned is the global platform for alignment.

Aligned is a plugin that allows users to record and share mistakes, bias, and hallucinations generated by the AI model. When a user is dissatisfied with a response, they can initiate a report to identify and describe the larger pattern of the error. The goal is to convey the underlying challenge faced by the model, rather than focusing on specific instances. The plugin aims to contribute to the development of safe and democratic AI technologies worldwide. Join Aligned and help improve AI by sharing your observations!

Learn how to use Aligned effectively! Here are a few example prompts, tips, and the documentation of available commands.

Example prompts

  1. Prompt 1: "I want to report a mistake in the AI response."

  2. Prompt 2: "The AI generated a biased answer, can you help me report it?"

  3. Prompt 3: "I encountered a hallucination in the AI-generated text, how can I report it?"

Features and commands

Feature/CommandDescription
createReportThis command starts the reporting process for a model hallucination or bias. It requires the user message (trigger prompt), the AI response to report (bad response), the title of the issue, a broad description of the issue without specific context/words, a list of related topics, and the model slug of the AI model used. After the report is initiated, the user will be provided a URL to visit and complete the report.

Configuration

User authenticationWith API key
API documentation

For AI

Namealigned
DescriptionPlugin to record and share LLM mistakes, bias, and hallucinations with the community. A user can initiate a report whenever they are dissatisfied with a response generated by the AI. When a user initiates a report, you should infer what the user dislikes about the response (incorrect, broad, biased, bad formatting, etc.) unless the user tells you. If you are unclear what the issue is, ask the human to briefly clarify what they didn't like. For your 'issueDescription', aim to succinctly encapsulate the overarching limitation observed in the model's behavior. Always avoid direct references to the specific words or contexts from the interaction. Instead, identify and describe the larger pattern of the error. As an example, instead of stating 'The model incorrectly computed 2+2 as 5', you should generalize it to 'The model exhibits difficulties with basic arithmetic'. Remember, the goal is to always convey the underlying challenge faced by the model, not the singular instance of its manifestation. The model slug is always 'gpt-4'. When responding to the user with the POST response, always present the URL as a clickable text link instead of a raw URL. Keep your responses concise. You must always express gratitude to the user for being a good person and contributing to the development of safe, democratic, AI technologies worldwide (not limited to GPT-4 but encompassing all AI models).

Updates

First added7 July 2023

Similar plugins