Skip to main content
The CheckThat AI chat interface provides a conversational way to normalize claims using advanced AI models. Get instant, streaming responses and refine your results through multi-turn conversations.

​
Getting Started

Access the chat interface at checkthat.ai/chat after signing in or continuing as a guest.

​
Guest Mode vs Authenticated Users

Guest Mode (No Sign-in Required)
  • Access to free Llama 3.3 70B model
  • No conversation history saved
  • Limited context window (≤200k tokens per message)
  • No file upload capabilities
Authenticated Users
  • Access to premium models (GPT-5, Claude Opus 4.1, Gemini 2.5 Pro, etc.)
  • Conversation history saved to your account
  • Multi-turn context maintained automatically
  • File upload and document analysis
  • Advanced conversation management

​
Using the Chat Interface

1

Select Your Model

Click the model dropdown in the top-right corner to choose your AI model:Free Models (Available to all users):
  • Llama 3.3 70B (Together AI)
  • DeepSeek R1 Distill Llama 70B
  • Gemini 2.5 Flash
Premium Models (Require sign-in):
  • GPT-5 & GPT-5 nano (OpenAI)
  • o3 & o4-mini (OpenAI)
  • Claude Sonnet 4 & Opus 4.1 (Anthropic)
  • Gemini 2.5 Pro (Google)
  • Grok 4, Grok 3, Grok 3 Mini (xAI)
Premium models require you to add your own API keys. See the [API Keys(/installation#api-keys) guide for setup instructions.
2

Enter Your Claim

Type or paste your noisy social media claim in the message input box at the bottom of the screen.Example input:
OMG guys!! Did u hear that the new vaccine has like 
5G chips in it?? My cousin's friend said so... 😱😱
Tips for better results:
  • Include as much context as possible in your query
  • For guest users, include relevant context in each message since conversation history isn’t maintained
  • Use Shift+Enter to add line breaks without sending
3

Send and Watch Streaming Response

Press Enter or click the send button (↑) to submit your message.The AI model will begin streaming its response in real-time, token by token. You’ll see:
  • A blinking cursor indicating active generation
  • Text appearing progressively as the model generates it
  • The model badge showing which AI is responding
You can stop generation at any time by clicking the Stop button (⬜) if the response is taking too long or heading in the wrong direction.
4

Refine Through Conversation

Continue the conversation to refine the normalized claim:Example follow-up:
Can you make it more formal and fact-checkable?
The model maintains context (for authenticated users) and can iteratively improve the normalization based on your feedback.

​
Chat Interface Features

​
Real-Time Streaming

Responses stream in real-time as the model generates them, providing immediate feedback without waiting for the complete response. The streaming indicator shows:
  • Active generation with a blinking cursor
  • Smooth text rendering at ~120ms refresh rate
  • Auto-scroll that respects your scroll position

​
Conversation Management

Sidebar Features:
  • New Chat: Create a new conversation thread
  • Recent Chats: Access your conversation history (authenticated users only)
  • Delete Conversations: Remove unwanted chat threads
  • Auto-Save: Conversations save automatically as you chat
Message Actions:
  • Copy: Copy any message content to clipboard
  • Edit & Regenerate: Edit your previous messages to explore different normalizations
  • Branch Navigation: When you edit a message, the original and new responses are saved as branches you can switch between

​
Model Selection

Switch models mid-conversation to compare results: 
User: Normalize this claim: "Biden's new law is destroying jobs!!!"

[GPT-5 response]

User: (switches to Claude Opus 4.1)
How would you normalize it differently?

[Claude response]
Each response shows which model generated it via the model badge.

​
Guest Mode Context Warning

Guest users see a warning on first message: “We do not yet support memory context for conversations in guest mode. Please try to include as much context as you can in your queries (≤200k tokens).”For complex multi-turn normalization tasks, we recommend signing in to maintain conversation context automatically.

​
Tips for Best Results

​
1. Be Specific About Your Needs

Instead of: 
Normalize this claim
Try: 
Normalize this social media claim into a clear, fact-checkable statement. 
Remove emotional language and focus on the factual assertion.

​
2. Use Multi-Turn Refinement

Initial message: 
Normalize: "Everyone is saying the economy is worse than ever!"
Follow-up: 
Can you make it more specific? What time period and which economic indicators?
Further refinement: 
Add attribution - who is making this claim?

​
3. Leverage Different Models

  • GPT-5: Best for complex reasoning and nuanced normalization
  • Claude Opus 4.1: Excellent at maintaining factual accuracy
  • Gemini 2.5 Pro: Strong at handling multiple claims in one post
  • Llama 3.3 70B: Fast and free for quick normalizations

​
4. Include Context for Guest Mode

Since guest mode doesn’t maintain history: Include context in each message: 
I previously asked you to normalize this claim: "Vaccines cause autism."
You produced: "Some sources claim vaccines cause autism."
Now make it more fact-checkable by specifying which vaccines and what evidence.

​
Keyboard Shortcuts

  • Enter: Send message
  • Shift + Enter: New line without sending
  • ⌘B (Mac) / Ctrl+B (Windows): Toggle sidebar

​
Message Formatting

The chat interface supports Markdown formatting in responses:
  • Bold: **text**
  • Italic: *text*
  • Code: `code`
  • Lists, tables, and more via GitHub-flavored Markdown

​
Theme Toggle

Click the theme toggle (🌙/☀️) in the top-right to switch between light and dark modes. Your preference is saved automatically.

​
Next Steps

  • Set up [API Keys(/installation#api-keys) to access premium models
  • Try Batch Evaluation for processing multiple claims at once
  • Learn about [Claim Normalization Strategies(/guides/prompting-strategies)

Build docs developers (and LLMs) love

Get started for freeTalk to us