Research and verification

Use AI to find and read faster without losing rigor. About 1 h, intermediate level.

You'll be able to

• Choose the right kind of AI research tool for finding, reading or mapping sources.
• Verify a citation by locating the source and confirming it supports the specific claim.
• Use lateral reading and cross-checking to judge claims and sources you don't know.

Finding is faster; checking is still yours

AI research tools fall into a few groups, and each does a different job:

• Search-first answers (Perplexity, and the web or Deep Research modes of general assistants) read web results and write a summary with links.
• Paper search and extraction (Elicit, SciSpace, Consensus) search academic papers and pull out methods, samples and findings into tables.
• Citation maps (ResearchRabbit, Semantic Scholar) show what cites what, so you find earlier, later and related work. ResearchRabbit generates no text, so it can't invent claims.
• Citation context (Scite) shows whether later papers support or dispute a result.
• Grounded reading (NotebookLM) answers only from sources you upload.

All of them cut the time spent finding and skimming. None of them removes your responsibility for what you cite. Your job shifts from finding sources to verifying them.

How AI research goes wrong

The failures are predictable, which makes them catchable:

• Invented citations. Plausible authors, a real-sounding journal, a year, sometimes a DOI, for a paper that doesn't exist. In 2023, lawyers in a New York federal case, Mata v. Avianca, were sanctioned after filing a brief citing cases ChatGPT had made up.
• Real paper, wrong claim. The source exists but doesn't say what the summary says, or says it with caveats that were dropped.
• Blended claims. A sentence combines two sources into a conclusion neither makes.
• Flattened nuance. “May be associated with, in one small sample” becomes “causes”.
• Extraction errors. A table cell reports the wrong sample size or the control group's result.
• Stale or retracted work presented as current.

General chat assistants without search are the most likely to invent references. Tools that search real papers rarely invent them outright, but they still misstate what papers found.

Check a citation in four steps

For every source you plan to rely on:

1. Find it independently. Search the title in USC Libraries or Google Scholar, or resolve the DOI at doi.org. Don't trust the link the AI gave you; confirm authors, year and venue match.
2. Open the full text. An abstract isn't enough to support a specific claim. If it's paywalled, USC Libraries can usually get it.
3. Find the passage. Locate the sentence, table or figure that supports the claim. If you can't, the claim is unsupported, however plausible.
4. Check the fit. Does the paper's population, method and hedging match how you're using it? Is it retracted or heavily disputed? Scite and publisher pages can show this.

Then cite the paper you read, never the AI tool's summary of it.

Read laterally

When you meet an unfamiliar website, report or statistic, the instinct is to read it closely and judge whether it seems credible. Professional fact-checkers do something different: they leave the page almost immediately and open new tabs to see what other sources say about it. This is called lateral reading.

Applied to AI output:

• Who is behind the source? Search the organization or author. Is it a research group, an advocacy site, a company selling something?
• Do independent sources agree? Look for the same claim from a second, unrelated source, ideally the original data or study.
• Where did the number come from? Trace statistics back to where they were first published, not to articles repeating them.
• Ask a second tool. Run the key claim through a different tool or a library database. Agreement isn't proof, but disagreement is a flag.

Lateral reading is quick. It often takes a minute and saves you from building on a bad source.

Keep a verification log

A short log turns checking into a habit and gives you evidence of your process. For each important claim, record:

• the claim, in your words
• the source the AI gave
• whether the source exists
• the passage that supports it, with page or section
• a verdict: supported, weak (partly supported, overstated or missing caveats) or wrong (unsupported or contradicted)

Prioritize. You can't trace every sentence of a long report, so start with the claims your argument depends on, plus anything specific: names, numbers, dates and quotes, which are where errors cluster.

The log also makes disclosure easy. Journals, instructors and supervisors increasingly ask how AI was used; “I used Elicit to find candidates and verified each cited claim against the full text” is a clear, honest answer.

Tracing one claim from a research report

Suppose a Deep Research report on study habits says: “A 2019 review found that spaced practice doubles long-term retention.”

Step 1: find it. You search the review's title in USC Libraries. It exists, and the authors and year match.

Step 2: open it. You read the results and discussion, not just the abstract.

Step 3: find the passage. The review reports that spacing improved retention across many studies, but gives no single figure and notes the effect varied by task and delay. “Doubles” appears nowhere.

Step 4: check the fit. Most studies used university students and word lists. Your paper is about workplace training.

Verdict: weak. The direction is supported; the size is invented and the population differs.

Your rewrite: “A 2019 review found spaced practice generally improved retention, though effects varied by task; most studies involved university students.” The claim is smaller, and now it's true.

Try it

A verification table for five key claims from an AI research report, each traced to a primary source and marked supported, weak or wrong, with the supporting passage noted.