When a claim needs a citation (and when it really doesn't)

Citation is one of the most over-explained and under-understood parts of academic writing. Students arrive in college having been told either to cite everything or to cite only direct quotes; both are wrong. The right rule is more useful and more compact: a claim needs a citation when a careful reader could reasonably ask, “How do you know that?” and you can't answer with personal observation, your own argument, or facts the reader already accepts.

That's the whole rule. Everything else is examples.

The six classes of claim

DraftGuard's Citation Coach sorts every input into one of six classes. The classification matters because the right answer to “does this need a citation?” is genuinely different in each bucket.

Common knowledge is the lowest-friction class. A claim that any educated reader of your essay would accept without proof. Water freezes at 0°C. The Second World War ended in 1945. Shakespeare wrote Hamlet. No citation. Your reader already knows. Citing common knowledge is mildly insulting; it suggests you think they don't.

The catch with common knowledge is that it is field-specific. The boiling point of water is common knowledge to anyone past tenth grade. The boiling point of liquid nitrogen is common knowledge to a chemistry undergrad and not to a literature undergrad. When in doubt, cite. The cost of a needless citation is one footnote; the cost of a needed-but-missing citation is your credibility.

Empirical claims are the most common citation trigger and the most often missed. An empirical claim is a statement of fact that depends on data, study, or measurement. Fish stocks have declined sixty percent since 1970. Students who sleep less than six hours a night perform 23 percent worse on standardized tests. The unemployment rate in Singapore in 2025 was 1.9 percent. Every one of these requires a real source. The source type matters: peer-reviewed studies for behavior claims, government datasets for economic claims, primary research for scientific claims.

AI loves empirical claims because they sound authoritative. AI also loves to make them up. “Studies have shown that students who use AI tools complete assignments 22 percent faster” is a statistically reasonable-sounding sentence that may or may not correspond to any actual study. DraftGuard refuses to fill in the citation; we leave a placeholder and prompt you for the real one. If you don't have a real one, the right move is to remove the statistic, not to keep it.

Expert claims attribute a view to a person or organization. Marx argued that capital concentrates over time. The IPCC reports that warming above 1.5°C raises sea levels roughly 30 cm. These need citation to the named source's actual work — not to a Wikipedia summary, not to a popular blog post about the source. The expert claim is the most embarrassing kind to fake because the actual work is usually findable.

Contested claims are statements where reasonable experts disagree. AI will replace 30 percent of jobs by 2030. Free will is incompatible with neuroscience. Increasing the minimum wage reduces employment. These require both a citation and a hedge. Cite a study that supports your version, and acknowledge the dispute. The fix isn't to remove the claim — your essay can argue a contested position. The fix is to argue, not assert.

Opinions are your own arguments and interpretations. I think reflective writing is harder than expository writing. The novel's central irony is that the narrator never learns. No citation required. Frame the claim as opinion — “I argue,” “I think,” “In my view” — so the reader knows it's an argument, not a report.

Personal anecdotes are claims about your own life. My third-year writing seminar had nine students. I have been studying Mandarin since I was eleven. No citation, but they have to be real. AI loves to invent personal anecdotes because they make application essays sound voiced. We won't help you do this. If you put a personal detail in your essay, it has to be true. If you can't think of one, leave a placeholder and come back when you can.

The working test

Here is a procedure that fits on a sticky note. For every sentence in your draft that asserts something:

1. Could a careful reader say, “How do you know that?” If no, you are in common knowledge or opinion territory. Move on.
2. If yes, can you answer with your own observation? Your own argument? Your own data? If yes, frame it that way (“In my experience,” “I observed,” “My data shows”) and move on.
3. If no, you are making an empirical, expert, or contested claim. You need a citation. Find the actual source, or remove the claim.

That's it. Three questions, sixty seconds per sentence. Faster once you have done it for a few drafts.

Why we don't generate citations for you

There is an entire class of products that will fabricate plausible-looking citations on demand. They will give you an author surname, a year, a journal title, a page range. Some of the citations are real. Most are not. The model has no way to know which is which because the model is producing the same kind of plausible string that fooled the average sentence in its training corpus into existing.

DraftGuard refuses this category of help. The Citation Coach classifies your claim and recommends the kind of source that would support it. We never produce a specific reference. If you don't have a real source, the right answer is to soften the claim, frame it as opinion, or cut it. Submitting a fake citation is the kind of move that ends careers. We will not be the accomplice.

Citations are not gatekeeping. They are how readers verify that you are not just confident but right. Used well, they make your argument stronger. Used badly, they suggest you don't know which of your own sentences you are willing to defend. The Citation Coach is built to help you tell them apart.