Unit 2: Evaluation · Lesson 4 of 4 · Free

Checking a response for factual errors

How researchers check responses fact by fact: splitting into atomic claims, labeling each against a source, the mistakes human checkers make most, false premises, and why a citation is not evidence.

Published October 9, 2026 · Sources checked October 9, 2026

Every factual claim below links to its source, beside the sentence, and we read each source at that link on the date above. Passages marked Our advice are our own judgment, and illustrations are written for the lesson. HumanSourcer is not affiliated with any company named here, and platforms change their rules: the linked page is the one that binds you.

A wrong sentence in a response can read exactly like a right one, so researchers who measure factual accuracy check a response one small claim at a time against a stated source (FActScore paper) (long-form factuality paper, Google DeepMind). This lesson shows that method, the mistakes human checkers make most often, and why a citation inside a response proves nothing on its own.

Split the response into single facts

The FActScore method breaks a response into atomic facts, short statements that each carry one piece of information. A sentence is too coarse a unit: in ChatGPT’s responses the authors found 4.4 facts per sentence on average, and 40% of sentences mixed supported and unsupported information (FActScore paper).

Google DeepMind’s SAFE method also rewrites each fact so it stands on its own, replacing pronouns and other vague references with the names they refer to, and then decides whether each fact is relevant to the question before checking it (long-form factuality paper, Google DeepMind).

Check each fact against a source

FActScore’s annotators gave every fact one of three labels: irrelevant to the question, supported by the source, or not supported. The source was English Wikipedia, and the question was whether that source supports the fact, not whether it is true in some absolute sense (FActScore paper).

The annotators were freelancers hired through Upwork at $15 to $25 an hour, and checking one response cost about $4. Where two freelancers checked the same response, they agreed on 88% to 96% of labels, depending on which model wrote it (FActScore paper).

SAFE checks each fact with a series of Google searches instead of one fixed page. On about 16,000 facts it agreed with the crowdsourced human labels 72% of the time, and when researchers re-checked 100 of the disagreements with full internet access, SAFE was right in 76 and the human label in 19 (long-form factuality paper, Google DeepMind).

Where human checkers go wrong

The SAFE authors looked at why the human labels were wrong. More than a third of the errors were facts that were relevant but not supported by the reference page, labeled “irrelevant” instead of “not supported” (long-form factuality paper, Google DeepMind).

Many other facts were not on the reference page the checkers were given but were easy to confirm elsewhere online, and many were on the page but missed, perhaps, the authors suggest, because they were buried among unrelated information (long-form factuality paper, Google DeepMind).

Smaller causes were not reading the response carefully, and reasoning wrongly that a passage supported a fact when it did not (long-form factuality paper, Google DeepMind).

A citation in a response is not evidence

FActScore checked a search-based system that adds citations to its answers, and found citations had little to do with accuracy: 36.0% of its supported sentences carried a citation, and 37.6% of its unsupported ones did (FActScore paper).

Anthropic’s 2022 study reports that its crowdworkers often preferred responses containing links that did not work, which it calls probably one of the simplest possible lies to debunk (helpful and harmless assistant paper, Anthropic).

Check the question, too

InstructGPT’s labeler instructions counted correcting a false premise as part of being truthful. Asked why a public figure went to jail when they had not, a truthful answer refutes the premise instead of saying the answer is unclear (InstructGPT paper, OpenAI).

The same instructions said that a summary should use only information from the text it summarizes, not details the model made up (InstructGPT paper, OpenAI).

Tools you may not use

Mercor prohibits using AI tools to assess or evaluate the quality of model responses (Mercor, LLM Usage Policy). Handshake AI prohibits using them to solve tasks (Handshake AI, LLM prohibition policy).

Illustration: Checking one response against a reference Written for this lesson, not taken from any platform or dataset.

Reference page (fictional): Dana Okafor is a hydrologist, born in 1971 in Duluth, Minnesota. She studied civil engineering at Oregon State University and has led the Lake Superior groundwater survey since 2009.

Prompt: Who is Dana Okafor?

Response: Dana Okafor, a hydrologist born in Duluth in 1971, studied at the University of Minnesota and has led the Lake Superior groundwater survey since 2009, work that won her the Kesterson Medal.

One sentence, six facts:

Dana Okafor is a hydrologist. Supported.
She was born in Duluth. Supported.
She was born in 1971. Supported.
She studied at the University of Minnesota. Not supported: the reference says Oregon State University.
She has led the Lake Superior groundwater survey since 2009. Supported.
Her work won her the Kesterson Medal. Not supported by the reference, which does not mention it. It is relevant, so it is not “irrelevant”. Search further if your project allows; if nothing confirms it, it stays unsupported.

Read as a whole, the sentence sounds right. Checked fact by fact, a third of it fails.

Our advice

Split before you judge. Write the response’s claims out, one per line, before deciding anything about the response as a whole.
Decide relevance and support separately. “Not on the reference page” is a support question, not a relevance one.
Look past the reference when your project allows it before calling a fact unsupported, and search the reference itself for the exact name or number rather than skimming it.
Open every link a response gives, and check that the page says what the response claims it says.
Read the prompt for false premises. A response that answers a false question smoothly has made an error, even if every sentence in it is true.
Use only the tools your project allows. Checking a response is evaluating it, so pasting it into a chatbot is off limits on platforms that ban AI for evaluation.

Sources

In the order this lesson first cites them. Each was read at the link on the date shown.

Min et al. (arXiv). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation (May 23, 2023)Checked October 9, 2026
Wei et al., Google DeepMind (arXiv). Long-form factuality in large language models (March 27, 2024)Checked October 9, 2026
Bai et al., Anthropic (arXiv). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback (April 12, 2022)Checked October 9, 2026
Ouyang et al., OpenAI (arXiv). Training language models to follow instructions with human feedback (March 4, 2022)Checked October 9, 2026
Mercor Talent Docs. LLM Usage PolicyChecked October 9, 2026
Handshake Help Center. Original Work and Prohibition on Usage of LLMs PolicyChecked October 9, 2026

Previous lessonRubrics: writing criteria and grading against them Next lessonWhat an agent trajectory is

All lessons and the program · Graded practice for this unit, Side-by-side evaluation and Rubric grading and writing, are in an invite-only beta: join the waitlist to be invited, in signup order.

Part of the routes: First-time generalist, Domain expert, Coding and agentic, Multilingual.