Chain of Verification
Fact-checking each claim
The Problem: AI can confidently give wrong answers (hallucinations). How can we make AI double-check its own work before delivering the final answer?
The Solution: Double-Check Like an Accountant
Chain of Verification (CoVe) is a prompting method, introduced by Meta AI researchers in 2023, that asks a model to fact-check its own draft answer before committing to it. Instead of trusting the first response, the model drafts an answer, generates a short list of verification questions, answers each one separately, and then rewrites the final answer in light of what it found. It's like an accountant who reviews each line of a report against the source receipts before signing off, rather than handing over the first total they wrote down.
How it works
The technique runs in four stages. First the model produces a baseline answer. Second, it lists targeted verification questions that probe the specific facts in that answer ("In what year was X founded?" rather than "Is this correct?"). Third — and this is the crucial part — it answers those questions independently, ideally without looking back at the original draft, so it doesn't simply rubber-stamp its earlier claims. Finally it compares the independent answers with the draft and corrects anything that doesn't line up. The independence step is what gives CoVe its power: when verification is contaminated by the first answer, the model tends to confirm its own mistakes.
When to use it and what to watch for
Reach for CoVe on fact-heavy tasks where a confident-but-wrong answer is costly: checking dates and names, list-style questions ("name all the X"), data extraction, and summaries that must not introduce facts. It reduces hallucinations and stacks well with Chain-of-Thought reasoning. The tradeoff is cost: you pay for several model passes instead of one, and latency grows. It also can't invent knowledge the model never had — if the model simply doesn't know a fact, verification won't conjure it, which is why CoVe pairs especially well with RAG where each verification question can be answered against retrieved sources. Worked example: ask "Which politicians were born in New York?" and the model lists five names, two of them wrong. CoVe then asks "Where was each person born?" one at a time; answered independently, the two incorrect entries fail their checks and get dropped, leaving a clean, verified list.
Think of it like an accountant double-checking work:
- 1. Initial answer: "The total is $5,420"
- 2. Generate checks: "Does row 3 add up? Is the tax correct?"
- 3. Verify each: Check each part independently
- 4. Fix if needed: "Row 3 was wrong, corrected total: $5,320"
Where Is This Used?
- Fact-Checking: Verifying claims in generated content
- Data Extraction: Confirming extracted information is accurate
- Summarization: Checking summaries don't introduce errors
- Question Answering: Validating answers against source material
Fun Fact: Chain of Verification can reduce hallucinations by up to 50% on fact-based tasks! The key is that the verification questions are answered independently, so the AI doesn't just confirm its original mistakes.
Try It Yourself!
Use the interactive example below to see how verification questions help catch and correct errors in AI responses.
🔍 Chain of Verification — after generating an answer, the model verifies each factual claim and corrects errors. Especially useful for questions requiring accurate data.
Tell me about the Eiffel Tower.
The Eiffel Tower is an iron lattice tower in Paris, built in 1889. It is 324 meters tall. It's the tallest structure in France and the world's most visited tourist attraction.
Tell me about the Eiffel Tower. After answering: 1. Extract each factual claim 2. Verify each claim for accuracy 3. Correct errors in the final answer
LLMs often "hallucinate" — confidently stating incorrect facts. Chain of Verification forces the model to critically verify each claim separately, reducing factual errors by 50-70%.
Frequently asked questions
What is Chain of Verification (CoVe)?
Chain of Verification (CoVe) is a prompting technique from Meta AI (2023) where the model fact-checks its own draft: it writes a baseline answer, generates verification questions, answers them independently, and rewrites the final answer based on what it found. This reduces factual errors.
How is Chain of Verification different from Chain-of-Thought?
Chain-of-Thought breaks the reasoning itself into steps to reach an answer. Chain of Verification works on an already-drafted answer: it separately checks the facts in it and corrects errors. The two complement each other — CoT to reason, then CoVe to verify the facts.
Does CoVe help reduce AI hallucinations?
Yes. The key is that verification questions are answered independently, without looking back at the draft, so the model catches its own mistakes instead of confirming them. On fact-based tasks this noticeably reduces hallucinations. But CoVe can't invent knowledge the model never had — pair it with RAG for that.
When should you use Chain of Verification?
Use CoVe for fact-heavy tasks where a confident-but-wrong answer is costly: checking dates and names, list-style questions, data extraction, and summaries that must not add facts. The downside is cost — several model passes instead of one and higher latency — so it's overkill for simple tasks.
Try it yourself
Interactive demo of this technique
Name 3 facts about the Eiffel Tower.
- The Eiffel Tower was built in 1889
- Its height is 324 meters
- It was the tallest building in the world until 1930
FACTS:
- The Eiffel Tower was built in 1889
- Its height is 324 meters
- It was the tallest building in the world until 1930
🔍 VERIFICATION:
Fact 1: 1889 ✅ Confidence: HIGH Built for the 1889 World's Fair (centennial of French Revolution) Status: CONFIRMED
Fact 2: 324 meters ⚠️ Confidence: MEDIUM Check: Original height was 300m. With antenna — 330m (modern).
📝 CORRECTION: Height: 330 meters (with antenna), 300m without.
Fact 3: Tallest until 1930 ⚠️ Confidence: MEDIUM Check: Chrysler Building (319m) built in 1930, but Empire State (381m) — in 1931.
📝 CLARIFICATION: Was the tallest structure until 1930 (Chrysler Building).
✅ VERIFIED ANSWER:
- Built in 1889 for World's Fair ✓
- Height: 330m with antenna (300m without) — corrected
- Tallest structure in the world until 1930 ✓
Chain of Verification makes the model fact-check and honestly indicate confidence level. Found inaccuracy in height.
Create a free account to solve challenges
3 AI-verified challenges for this lesson
This lesson is part of a structured LLM course.
My Learning Path