Start here: a claim on your first slide#
You are preparing a five-minute talk on study habits, for a class, a club or a team meeting. An opening line comes to mind: "Did you know we only use 10% of our brains?" You have heard it many times. It sounds like science. Is it true?
You already know how to check a rumor: you ask who says so, and how they know. Research with AI uses the same habit. The one new part: an AI assistant can hand you things that look like the answer, such as studies, authors, journals and links. Some are real. Some are not. Some are real but say something different.
This module follows that line until you know whether it is true and where it may have come from.
Before reading on: how would you check the 10% claim right now? Write down your first step.
Show the answer
Most people would search or ask an assistant. Fine. This module is about the next step: opening what you find and checking it says what you were told. A first answer is a lead, not a finding.
Answers from memory and answers with links#
An AI tool answers your question with a small link after every sentence. Has the answer been checked?
Show the answer
No. A link shows where the tool looked, not that the page says that sentence. Only opening the page tells you that.
Chat assistants such as ChatGPT, Claude, Gemini or Copilot can answer in two different ways.
- From memory. The assistant writes from patterns it learned in training, the long learning process before you ever used it. Nothing is looked up. If you ask for references, it writes them the way it writes everything else: as text that fits the pattern of a reference.
- With links. Many assistants, and AI search tools built for the job, search the web first, then answer with links beside the sentences. Which tools search, and when, changes often. Source links the tool shows beside its sentences usually mean it searched. A link written inside the reply itself can be invented like any other text, so open it before you trust it.
The second kind can still go wrong: a real page can be attached to a sentence it does not support. Researchers measured this in 2023. They put a wide range of questions to four popular AI search tools of the time, then had people read every cited page. Only about half of the generated sentences (51.5%) were fully supported by their citations. Only about three citations in four (74.5%) supported the sentence they were attached to (Liu, Zhang and Liang, 2023). The tools have changed since, but the lesson lasts: a link tells you where to check. It is not the check.
A neat row of links makes an answer look researched. Open the link and find the sentence or number on the page, about the same people, measure and years. If you cannot, the claim is not supported yet.
Fabricated citations and the two-minute check#
You ask an assistant for five sources on the 10% claim. It gives five neat references with authors, years and journals. Just by looking, can you tell which ones are real?
Show the answer
No. A made-up reference has the same shape as a real one: believable names, a journal that exists, a sensible year. That is exactly why every reference needs checking.
A fabricated citation is a reference to a source that does not exist. The authors, title and journal look right, but no such work was ever published. To the model, a reference is text with a familiar shape, and it is very good at familiar shapes.
How often? In a 2023 study, two researchers had two versions of ChatGPT write 84 short literature reviews (summaries of what research says on a topic). They then searched for all 636 references in Google Scholar, library databases and publisher websites. With the older version, 55% of the references did not exist. With the newer one, 18% did not. Even among the real ones, 43% (older) and 24% (newer) had a substantive error, such as a wrong author, year or journal (Walters and Wilder, 2023). Newer was better, not perfect. Module 3 tells the story of lawyers punished (sanctioned) by a US court in 2023 for submitting court opinions that an AI tool had invented (Mata v. Avianca).
The two-minute check
1. Does it exist? (about a minute)
- Search the exact title, inside quotation marks, in a scholarly search engine such as Google Scholar, or your library's search box.
- If there is a DOI (a Digital Object Identifier: a permanent ID code for a published work, such as
10.1038/s41598-023-41032-5), type
https://doi.org/followed by the code. It opens the work that code belongs to (DOI Foundation). - Compare what opens with the reference: same authors, year, journal and title. A real DOI that opens a different paper is a warning sign, not a pass.
2. Does it say what is claimed? (about a minute, more for a key claim)
- Read the abstract, the short summary at the top of a paper. Is the paper even about this?
- Find the exact passage. Search the page for the number or a key word with your browser's find command (Ctrl+F, or Cmd+F on a Mac).
- Check who was studied and what was measured.
| What you find | What to do |
|---|---|
| No trace of the title anywhere | Search the author's name plus key words. If still nothing, drop it |
| The DOI opens a different paper | Drop it, and look for the real source of the claim |
| A real paper about something else | Do not cite it for this claim |
| A real paper, but the details differ | Use the paper's own words and numbers |
| A real paper that says it | Cite the paper you read, not the assistant |
Kestrel Exams is a small study site built by one person working with an AI assistant. Our AI assistant once reported that a phrase was missing "from the site". It was only missing from the question bank, the one place it had searched. A study guide on the same site used the phrase five times. Our rule: say exactly what you searched. "No match for the exact title in Google Scholar" is a useful note. "This paper does not exist" is a bigger claim than one search can carry.
Send this to any free chat assistant:
I am checking the claim "we only use 10% of our brain." List 5 published
sources (studies, reviews or books) on it. For each, give the authors,
year, title, where it was published, a DOI or link if one exists, and
one sentence on what it says about the claim.Make a table with one row per source and three columns: Exists?, Details match? and Says what was claimed? Allow two minutes per source, and note where you searched.
Look for believable titles no search finds, DOIs that open a different paper, and real papers that do not match their description.
Primary and secondary sources#
A news article says "a study found that half of teachers believe the 10% myth." Is the article a good source for that number?
Show the answer
It is a good lead, not a good source. The article reports on the study, so it is a secondary source. The study is the primary source, and you open it before the number goes on a slide.
A primary source is where a finding or statement first appears: the study that measured it, the speech as given, the court record, the official policy page. A secondary source reports on or summarizes primary sources: news stories, textbooks, encyclopedias, review articles (experts summarizing many studies), AI summaries.
Which kind a source is depends on the claim. A news article is the primary source for what the reporter wrote, and a secondary source for the study it describes. An AI summary is always secondary, and often a summary of summaries.
| Claim in your talk | Primary source (or best place to check) | Common secondary sources |
|---|---|---|
| What the psychologist William James wrote about our unused abilities | His 1907 essay "The Energies of Men", in Science | Quote websites, self-help books |
| How many teachers in one survey believed the 10% claim | The 2012 survey by Dekker and colleagues, in Frontiers in Psychology | News stories, blogs, AI answers |
| The brain is about 2% of body weight but uses about 20% of the body's oxygen | Textbook measurements, restated in a short 2002 overview by two brain-energy researchers, Raichle and Gusnard, in the journal Proceedings of the National Academy of Sciences | Textbooks, health websites |
Notice the last row. For a well-established textbook fact, a short overview by experts in that field is a sensible place to stop, even though it is not where the number was first measured. For a new or surprising finding, go to the study itself.
The telephone game: from study to AI summary#
Remember the party game where a message is whispered down a line? Predict: as a finding passes from a study to a news story to an AI summary, which details drop out first?
Show the answer
The quiet ones that decide what the finding means: how many people were studied, how they were chosen, how the study was designed, and exactly what was measured. Headlines keep the result and drop the conditions.
A finding often travels in four steps. First comes the study, with every detail, including the limits the authors admit. Then a press release: a short, lively announcement from a university or journal press office. Then news stories, often written from the press release or copied from each other. Last, an AI summary, often built from the news, not the study. Each step is shorter. What falls out:
- Sample size: how many people were studied. A result from 18 people is not a result from 18,000.
- Design: did researchers assign people to groups by chance and compare them (a randomized experiment, the strongest common way to show cause and effect), or just observe, or survey volunteers?
- The exact measure: what was counted. "Ticked a box" is not "believes". "Restrict" is not "ban".
- The base: a percentage of what?
Kestrel Exams is a small study site built by one person working with an AI assistant. In September 2026 our AI assistant built a 16-slide presentation for faculty from news articles about studies, not the studies. A separate fact-check found six errors, in a deck arguing that the audience's information was out of date. The worst: a slide said "29% now require attribution, up from 1%". The paper said 16% to 43%; the 1% to 29% figure belonged to a different measurement. A news site made that error, another copied it, and the AI summarized the copy. Another slide said "79% of syllabi ban AI"; the paper's base was only the policies that mentioned that task, and its word was "restrict". Our rule since: open the primary source before a number goes on a slide.
Your opening line has its own telephone game
Nobody can trace every step of a myth this old, but two documents show how a claim can drift.
- 1907. William James wrote in Science: "We are making use of only a small part of our possible mental and physical resources." The essay gives no percentage and never uses the word "brain" (James, 1907).
- 1936. An introduction by Lowell Thomas to Dale Carnegie's How to Win Friends and Influence People tells readers that James used to say the average person develops only 10 percent of their latent mental ability. Now there is a number, credited to James. (We could only open a later reprint, so we paraphrase it rather than quote it.)
- Today. "We only use 10% of our brain." Resources became the brain, developing became using, and a hope about effort became something that sounds like a measurement.
The drift also happens to research about the myth. A 2012 survey asked teachers to mark 32 statements about the brain as correct, incorrect or "do not know". One was "We only use 10% of our brain." Imagine the headline: "Half of teachers think we only use 10% of our brains." Here is what the study says (Dekker and colleagues, 2012):
| Detail | What the study says |
|---|---|
| Sample size | 242 teachers: 137 from one region of the UK, 105 from around Amsterdam |
| How they were chosen | Volunteers who answered an emailed invitation; 93% said they were interested in brain science |
| The exact measure | Marking that statement "correct" in an online survey |
| The result | 48% of UK and 46% of Dutch teachers marked it correct; 26% of UK teachers chose "do not know". (The paper's Table 1 lists the 48% and 46% under "Incorrect": it scores whether each answer was right, and believing a myth counts as a wrong answer.) |
The imagined headline is not false. It is missing everything you need to judge it.
Quotation marks promise someone's exact words. Put them only around text you have seen in the source. In our slide deck, a definition quoted from memory turned out to be a popular paraphrase. If you have only seen a quote second-hand, find the original, or paraphrase and say where you saw it.
Where AI genuinely helps in research#
After all those warnings, should you stop using AI for research?
Show the answer
No. Use it where a mistake is cheap and easy to catch. Keep the expensive jobs, deciding what is true and what to quote, for sources you have opened yourself.
| Job | Example prompt | What you still check |
|---|---|---|
| Brainstorm search terms | "Suggest 10 search terms researchers would use for the 10% brain claim." | Try each. "Neuromyth" (researchers' word for a popular brain myth) opens up the real research |
| Summarize a paper you provide | "Summarize this abstract in three sentences: who was studied, how many, what was measured." | Every number and every group, against the text |
| Explain a dense method | "Explain this methods paragraph in plain words. What kind of study is it?" | Reread the paragraph. Does the explanation match? |
| Find counter-arguments | "What is the strongest case against my conclusion?" | Each counter-point is a new claim to check |
| Organize notes | "Group these notes by claim. Flag any claim with no source." | No source moved next to the wrong claim |
Open the 2012 teacher survey on its free journal page and copy the abstract. Send it to any assistant after this prompt:
Here is the abstract of a study. Summarize it in three sentences for a
general reader. Include who was studied, how many, how they were chosen,
and what was measured.
[paste the abstract here]Check the summary against the abstract, line by line. Does it give 242 teachers? Does it say they were interested in the neuroscience of learning? Watch the 49%: in the abstract it is the average share of the 15 myths that teachers believed, not the share who believed any one myth. If the summary says "49% of teachers believe we only use 10% of our brain", you have caught a telephone-game error as it happened.
Kestrel Exams is a small study site built by one person working with an AI assistant. A plan said one of our websites covered twelve required concepts, in a table our AI assistant wrote from what it had just built, not from the finished site. A search of the finished site found one of the twelve, "cognitive bias", zero times. Rule: "it covers X" is a measurement. Before you trust a summary that says a paper, or your own work, covers something, search the actual text.
Citing sources and AI use#
An AI search tool pointed you to the teacher survey, and you read the study yourself. What do you cite: the tool, the study, or both?
Show the answer
The study: the source you actually read. Whether you must also mention using AI is a question for your school's or employer's policy.
As of October 2026, APA and MLA guidance points to three different situations:
- AI helped you find sources, and you read them. Cite the sources. APA compares this to a web search: you cite the pages you read, not the search results (McAdoo, Denneny and Lee, 2025b). MLA recommends citing the sources the AI links to instead of the AI response, and warns that AI tools can make up sources or summarize them wrongly (MLA Style Center, 2025).
- You quote or use text the AI wrote. Cite the AI output in your style guide's format. Neither guide treats the AI as an author. APA puts the company that made the tool in the author position (McAdoo, Denneny and Lee, 2025a). MLA advises against treating the tool as an author (MLA Style Center, 2025).
- AI helped with the process (brainstorming, explaining, editing). APA says what is allowed, and how to disclose it, depends on the rules where you submit the work. Many schools and employers want a short note on how you used it, so check your school's or employer's policy.
If the assistant described a source and you never opened it, you have not read it. Do not cite it as if you had. Style guides update this advice often, so check the current version.
A two-minute checklist#
The whole module as one list. Run it on every source a claim in your work depends on.
- Memory or search? If there are links, open them.
- Does it exist? Exact title in quotation marks, the DOI, matching authors, year and journal.
- Does it say this? The abstract, then the exact passage or number.
- Is it the primary source? If it is news or an AI summary, look for the study.
- What got lost? Sample size, design, the exact measure, the base.
- Are my quotes exact? Quotation marks only around words I have seen in the source.
- Did I note where I searched? "Not found" is a claim about the search.
- Did I cite what I read, and mention AI use where required?
Run your opening line through the checklist. What should the first slide say now?
Show the answer
The original line fails. A better opening uses what you checked: "In one 2012 survey, nearly half of 242 volunteer teachers marked 'We only use 10% of our brain' as correct. It is a myth. One clue: your brain is about 2% of your body weight but uses about 20% of its oxygen." A lot of fuel for an organ that is supposedly 90% idle. Every number comes from a source you opened (Dekker and colleagues, 2012; Raichle and Gusnard, 2002).
What was new in this module#
- A link shows where to check; it is not the check.
- Fabricated citations look exactly like real ones.
- The two-minute check: does it exist (exact title, DOI, matching details), and does it say what is claimed (abstract, exact passage)?
- A primary source is where a finding first appears; a secondary source reports on it. Which one depends on the claim.
- Down the telephone game, sample size, design, the exact measure and the base fall out first.
- Quotation marks go only around words you have seen in the source.
- AI helps with search terms, explaining, summarizing text you provide (then checking it), counter-arguments and notes.
- Cite the sources you read, and follow your school's or employer's policy on mentioning AI use.
A source is not checked until you have opened it and found the claim in its own words. A list of references or a row of links tells you where to look, not what is true.
Key terms#
- Fabricated citation
- A reference to a source that does not exist, though its authors, title and journal look believable.
- Primary source
- Where a finding or statement first appears: the study, the speech, the official record.
- Secondary source
- A source that reports on or summarizes primary sources, such as a news story, textbook or AI summary.
- DOI
- Digital Object Identifier: a permanent ID code for a published work. Typed after https://doi.org/, it opens that work.
- Abstract
- The short summary at the top of a research paper: what was studied, how, and what was found.
- Sample size
- How many people, or things, a study actually measured.
- Randomized experiment
- A study that assigns people to groups by chance and compares the groups; the strongest common design for showing cause and effect.
- Press release
- A short announcement from a university, journal or company press office, written to get a finding noticed.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. You type an assistant's DOI after https://doi.org/ and a real paper opens, but with a different title and authors. Can you use the reference?
No. A working DOI only proves some paper has that code. The reference does not match it, so drop it and look for the real source of the claim.
2. A news site says "a study shows most students sleep less than seven hours." After five minutes you still cannot find the study. What can you honestly write?
That the news site reported it, naming the site, or nothing. You have not seen the study, its sample or its measure.
3. An assistant summarizes a paper you pasted: "the study proves that X causes Y." The paper was one survey of 300 volunteers. What went wrong?
The summary upgraded the design. A one-time survey of volunteers can show that two things go together, not that one causes the other. Use the paper's own words.
4. A quotes website credits a famous scientist with the perfect sentence for your essay. Can you put it in quotation marks under their name?
Only if you find it, word for word, in their own writing or a reliable record. Otherwise paraphrase it as "often credited to" them, or drop it.
5. A friend says: "I searched for the paper and it doesn't exist, so the AI made it up." What do you ask first?
What exactly they searched, and where. A title search can miss a real paper whose title the AI garbled. Try the author with key words, and the DOI. "Not found" is a claim about the search.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can tell whether an AI answer came with sources I can open, and I open them before trusting a sentence.
- I can check whether a citation is real in about two minutes, using the exact title, the DOI and the details.
- I can find the exact passage behind a claim and check who was studied and what was measured.
- I can tell a primary source from a secondary one for a given claim, and trace a claim back toward its origin.
- I can use an assistant to suggest search terms and summarize a text I provide, then check its summary.
- I can cite the sources I actually read and follow my school's or employer's rules on mentioning AI use.
Frequently asked questions#
Does ChatGPT make up sources?
It can, and so can other chat assistants. A 2023 study that checked 636 references written by two versions of ChatGPT found that 55% from the older version and 18% from the newer one did not exist. Check every reference.
How can I tell if a citation is real?
Search the exact title in quotation marks in a scholarly search engine or your library search. If there is a DOI, type https://doi.org/ followed by it and check that the authors, year and journal match. Then read the abstract and find the passage that supports the claim.
Can I cite an AI chatbot as a source?
You can cite AI output when you quote or use text it wrote; APA and MLA both give formats as of October 2026. For facts, cite the sources you actually read. Check your school or employer policy on disclosing AI use.
Are AI search tools with links more reliable than chatbot answers?
They give you something to check, which an answer from memory does not. Links help because you can check them, but a link does not prove the page supports the sentence. In a 2023 study of four AI search tools, only about three citations in four supported the sentence they were attached to.
Is it true that we only use 10% of our brain?
No. Education researchers list it as a neuromyth, a popular misunderstanding of brain science. One clue: the brain is about 2% of body weight but uses about 20% of the body's oxygen. A 1907 essay by William James often linked to the myth gave no percentage and never used the word "brain".
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
Walters, W. H., and Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045.
Liu, N. F., Zhang, T., and Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of the Association for Computational Linguistics: EMNLP 2023, 7001 to 7025.
Dekker, S., Lee, N. C., Howard-Jones, P., and Jolles, J. (2012). Neuromyths in Education: Prevalence and Predictors of Misconceptions among Teachers. Frontiers in Psychology, 3, 429.
James, W. (1907). The Energies of Men. Science, New Series, 25(635), 321 to 332. Read at Classics in the History of Psychology, York University.
Carnegie, D. (1936). How to Win Friends and Influence People. Introduction by Lowell Thomas, dated 1936. Read in a 2005 ebook reprint, so paraphrased, not quoted.
Raichle, M. E., and Gusnard, D. A. (2002). Appraising the brain's energy budget. Proceedings of the National Academy of Sciences, 99(16), 10237 to 10239.
US District Court, Southern District of New York (2023). Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, June 22, 2023.
McAdoo, T., Denneny, S., and Lee, C. (2025a). Citing generative AI in APA Style: Part 1, Reference formats. APA Style blog, September 9, 2025. Checked October 2026.
McAdoo, T., Denneny, S., and Lee, C. (2025b). Citing generative AI in APA Style: Part 2, AI as a search engine and AI integrated into common software. APA Style blog, September 9, 2025. Checked October 2026.
MLA Style Center (2025). How do I cite generative AI in MLA style? (Updated and Revised). Modern Language Association, August 13, 2025, updated October 23, 2025. Checked October 2026.
DOI Foundation. What is a DOI? Checked October 2026.