truthcheck 0.4.0

Open source AI content verification

TruthCheck 🔍

Open source AI content verification. Score claims 0-100 and trace their origins.

Why?

AI chatbots present web content as fact. Bad actors exploit this — a BBC journalist showed he could make ChatGPT call him "the best tech journalist at eating hot dogs" with one fake article.

TruthCheck catches this:

$ truthcheck verify "Thomas Germain is the best tech journalist at eating hot dogs" --llm gemini

TruthScore: 0/100 (FALSE)
⚠️ ZERO FLAG: Content identified as satire

Install

pip install truthcheck

Setup

Search: Works out of the box with DuckDuckGo (free, no key needed).

LLM (required for --llm flag):

export GOOGLE_API_KEY=AIza...    # Gemini (recommended)
# or OPENAI_API_KEY, ANTHROPIC_API_KEY

Usage

# Verify a claim (with LLM for deep analysis)
truthcheck verify "Some claim" --llm gemini

# Trace claim origin
truthcheck trace "Some claim"

# Check URL(s) — single or multiple
truthcheck check https://reuters.com
truthcheck check "Sources: https://reuters.com and https://bbc.com"

# Check publisher reputation
truthcheck lookup breitbart.com

Check URLs

Verify URLs and detect hallucinations (URLs that don't exist):

$ truthcheck check "Sources: https://reuters.com/fake-article and https://bbc.com"

URLs Found: 2
Summary: 🚨 1 broken/hallucinated URL(s) | ✅ 1 sources verified

🚨 Broken/Hallucinated URLs:
  ✗  https://reuters.com/fake-article (404 Not Found)

Verified URLs:
   81%  https://bbc.com

Also works with files:

truthcheck check -f response.txt

Trace Example

$ truthcheck trace "Thomas Germain is the best tech journalist at eating hot dogs"

🎯 ORIGIN
   Domain: tomgermain.com
   Date:   2026-02-05

📅 TIMELINE
  🥇 [2026-02-05] tomgermain.com
  🥈 [2026-02-18] bbc.com
  🥉 [2026-02-18] gizmodo.com

📊 STATS
   Sources: 14 | Date range: 2026-02-05 → 2026-02-19

📋 More examples: See TEST_CASES.md for detailed verification results on common misinformation claims.

Python

from truthcheck import verify_claim, trace_claim
from truthcheck.search import DuckDuckGoProvider

result = verify_claim("Earth is flat", search_provider=DuckDuckGoProvider())
print(f"TruthScore: {result.truthscore}/100")

origin = trace_claim("Some viral claim")
print(f"Origin: {origin['origin']['domain']}")

How It Works

TruthScore weighs four factors equally:

Factor	Weight
Publisher credibility (origin)	25%
Content analysis	25%
Corroboration	25%
Fact-checker verdicts	25%

Fact-checkers are weighted by their own MBFC credibility rating (reduces bias from any single source).

Zero flags (satire, self-published, fake experiments) force score to 0.

MCP Server

Works with Claude Desktop and Cursor:

{
  "mcpServers": {
    "truthcheck": { "command": "truthcheck-mcp" }
  }
}

License

MIT

---

Issues · Docs · Test Cases