How We Extract & Classify Every Factual Claim — The Claim Extraction Engine

Estimated Read Time: 10 minutes | Category: Technology Deep-Dive

The Anatomy of AI Hallucination

Studies consistently show that 30% of AI-generated claims are incorrect—fabricated statistics, misattributed quotes, invented citations. The problem isn't that AI lies intentionally; it's that AI generates fluent text that sounds authoritative whether or not it's true.

At PromptReports.ai, we solve this with radical transparency: every factual claim in your report is extracted, classified, and verified independently. This blog post explains Step 5 of our 11-step pipeline: the Claim Extraction Engine (CEE).

Pipeline Step 5: Claim Extraction Engine (CEE)

After the research phase produces a synthesized draft report, the CEE takes over. Its job: parse natural language prose into atomic, independently verifiable claims.

What is an "Atomic Claim"?

An atomic claim is the smallest unit of verifiable information. Consider this paragraph:

"Tesla delivered 1.8 million vehicles in 2023, a 38% increase from the previous year, making it the world's largest EV manufacturer by volume."

The CEE extracts three atomic claims:

1. "Tesla delivered 1.8 million vehicles in 2023" (Statistical)

2. "This represents a 38% increase from 2022" (Comparative)

3. "Tesla is the world's largest EV manufacturer by volume" (Attributive)

Each claim can be verified independently—and might have a different verification score.

The Seven Claim Types

The CEE classifies every claim into one of seven types, because different claim types require different verification approaches:

1. Statistical: Numerical data, measurements, quantities. Example: "Revenue grew 23% YoY". Verification: CVM (code-based verification)

2. Attributive: Assigning properties or roles to entities. Example: "Google is a search engine company". Verification: Multi-source corroboration

3. Causal: Asserting cause-and-effect relationships. Example: "The rate hike caused the market crash". Verification: Temporal + expert consensus

4. Comparative: Ranking or comparing entities. Example: "Apple is more profitable than Microsoft". Verification: Direct data comparison

5. Temporal: Time-based assertions. Example: "The law took effect in January 2024". Verification: Official record lookup

6. Existential: Asserting something exists or occurred. Example: "A vaccine exists for this disease". Verification: Authority source verification

7. Analytical: Interpretations, predictions, opinions. Example: "The market will likely decline". Verification: Multi-source agreement check

Verification Priority Assignment

Not all claims matter equally. The CEE assigns each claim a Verification Priority based on:

CRITICAL: Central to the report's thesis, used in executive summary, or cited by other claims. Receives full CGA + CVM if quantitative.

HIGH: Supports key arguments, from sections likely to be read. Receives full CGA.

MEDIUM: Supporting detail, contextual information. Receives standard CGA.

LOW: Background context, widely known facts. Receives quick verification pass.

Up Next

In the next post, we'll follow these extracted claims into the verification pipeline—specifically the Citation Resolution Service (CRS) and the three-stage Content Grounding Analyzer (CGA).

PromptReports.ai is a Verified Intelligence Platform that delivers AI-powered analyst reports with claim-level source verification. Generate your first verified report →