Opinion vs Fact in Journalism: How AI Learns to Tell the Difference

Why the Opinion/Fact Distinction Is Fundamental

The ability to distinguish between factual claims and opinions is foundational to both journalism and AI fact-checking. A factual claim is an assertion that can, in principle, be verified as true or false against external evidence — "The unemployment rate in India is 7.8%" is a factual claim. An opinion is a subjective assessment or value judgment that cannot be objectively verified as true or false — "The government has failed on economic policy" is an opinion, even if held by many economists.

Fact-checking only applies to factual claims. Attempting to fact-check an opinion is a category error — opinions are not true or false, they are more or less well-supported, coherent, and informed. A fact-checker who renders verdicts on opinions rather than facts is practising advocacy, not journalism. Similarly, an AI fact-checking system that assigns "True" or "False" verdicts to opinions will mislead users rather than inform them.

How AI Models Are Trained for Claim Classification

AI systems classify statements as facts or opinions using claim detection models — NLP classifiers trained on large annotated datasets of text labelled by human experts as factual, opinion, or mixed. Training datasets include claims from PolitiFact, FactCheck.org, ClaimBuster, and FEVER (a fact extraction and verification dataset from University of Edinburgh with 185,000 annotated claims).

Claim detection models use several linguistic cues: factual claims tend to contain specific names, dates, locations, statistics, and quantifiable assertions; opinion statements tend to use subjective evaluative adjectives ("excellent," "disastrous," "unfair"), hedging language ("I believe," "in my view"), and comparative value judgments. Modern transformer-based models (BERT, RoBERTa, DeBERTa fine-tuned on claim detection tasks) achieve 85–92% accuracy on standardised benchmarks.

Mixed Claims: The Hardest Cases

The most challenging cases for both human and AI fact-checkers are mixed claims — statements that combine verifiable facts with opinion framing in ways designed to give false impressions. "Since taking office, unemployment has risen by 2 percentage points, proving the government has failed" contains a verifiable factual component (the unemployment figure) and an opinion (it proves failure). Fact-checkers must verify the factual component while noting that the conclusion drawn is interpretive.

Omniscient AI's classification system distinguishes between three verdict categories for exactly this reason: "Factual and Verified," "Factual but Disputed," and "Primarily Opinion/Unverifiable." This prevents the system from generating false precision by labelling complex opinion-fact hybrids as simply "True" or "False."