Source Checking & Citations

Perplexity

Cites numbered sources for every claim and links to the original pages. In 2025 accuracy tests, Perplexity tied every claim to a source for about 78% of complex research questions (vs. 62% for ChatGPT). Still: the citations link to real web pages, but the AI sometimes misattributes — claiming a source says something it does not. Always click and read the cited source for any claim you will use.

Consensus

Restricts all references to published, peer-reviewed papers. The safest tool for academic citations — every source it cites is at least a real, published paper. Still requires you to verify that the paper actually supports the specific claim being made, not just that the paper exists.

NotebookLM

Works only from documents you upload — so it cannot cite a paper that does not exist in your notebook. Every answer cites the exact passage from your source. The safest citation approach for working with documents you already have, since it cannot hallucinate from training data.

Uncertainty-flagging prompts

• "Answer my question, but flag any specific fact, statistic, or citation you are not highly confident about. Use phrases like 'I'm uncertain about this' or 'you should verify this' where appropriate."
• "After answering, tell me: which specific claims in your response would be hardest to verify, and where should I look to check them?"
• "If you are about to cite a specific paper, statistic, or quote, tell me how confident you are that it exists exactly as you describe it, on a scale of low / medium / high."
• "For this response: do not include any citation you are not highly confident actually exists. If unsure, describe the general concept without a specific citation rather than guessing."

For quick fact-checking (2 minutes)

1. Identify the specific claim you want to verify
2. Open Perplexity — search for the claim directly (not the AI-generated answer, the claim itself)
3. Check 2–3 independent sources cited by Perplexity that confirm the claim
4. If Perplexity cannot find confirmation: the claim is likely inaccurate or unverifiable

For academic citations (5–10 minutes)

1. Take the citation to Google Scholar — search by title and author
2. If not found: discard the citation entirely
3. If found: access the paper and use Ctrl+F to search for the specific claim
4. If the claim is not in the paper: discard, note the paper exists but does not say what was claimed
5. If found: note the exact page and quote — you now have a real, usable citation

For statistics and data (5–15 minutes)

1. Search the statistic + likely source type ("[stat] site:gov" or "[stat] research report")
2. Find the primary report — not a blog or news article that cites it
3. Locate the exact figure in the primary document
4. Check methodology: sample size, date, geography, what was actually measured
5. If you cannot trace it to a primary source: do not use the figure

Using AI to help verify AI (the right approach)

1. Use Perplexity (cited) or Consensus (academic papers only) rather than plain ChatGPT for research
2. Ask a different AI assistant to check the same claim independently — disagreement is a signal
3. Ask: "What is the most credible primary source I should check for [claim]?" — then go check that source yourself
4. Upload the actual document to NotebookLM and ask questions — it cannot hallucinate beyond what you gave it

Well-established conceptual explanations

AI is very reliable when explaining widely-taught concepts — how photosynthesis works, what a derivative is, how TCP/IP functions, what the French Revolution was. These are covered extensively in training data and the general explanation is stable. You still may want to verify fine details, but the conceptual framework is usually sound.

Widely known, frequently documented facts

Facts that appear in thousands of sources — major historical dates, well-known scientific constants, the capitals of countries, the authors of famous works — are very unlikely to be wrong. The more extensively documented a fact is, the more reliably AI gets it right.

Reasoning and logic (not facts)

AI's reasoning about a set of facts you provide is generally more reliable than its recall of external facts. If you give the AI accurate data and ask it to analyse, compare, structure, or reason about it — as distinct from asking it to recall facts from memory — the output tends to be much more dependable.

Hallucination rates have improved — but not disappeared

The best models in 2025–2026 have meaningfully lower hallucination rates than earlier generations. Gemini 2.0 Flash achieved a hallucination rate as low as 0.7% in some benchmarks in April 2025. But lower hallucination rate does not mean zero — and even 0.7% means roughly 1 in 140 claims is wrong, stated with confidence. At scale, that remains a significant problem.

NeurIPS 2025: fake citations in elite AI research (January 2026)

The discovery that 100+ hallucinated citations slipped through peer review at the world's top AI conference was a significant moment in 2026. GPTZero coined the term "vibe citing" for the pattern: AI derives plausible-sounding citations from combinations of real papers, producing references that look accurate to casual inspection but fall apart under close examination. The submission volume to NeurIPS grew 220% from 2020 to 2025, straining review processes and making this problem harder to catch.

C2PA watermarking becoming standard for AI content

The Coalition for Content Provenance and Authenticity (C2PA) standard is being adopted by OpenAI, Google, Adobe, and Microsoft to embed invisible metadata in AI-generated images, audio, and video indicating their AI origin. This helps receivers of content verify whether it was AI-generated, though watermarks can be stripped and detection is not foolproof.

AI fact-checking tools emerging

Dedicated AI hallucination detection tools — like GPTZero's Hallucination Detector — can automatically scan AI-generated text for unsupported claims and hallucinated sources. MIT's ContextCite research tool tracks how AI attributes information to sources, making it possible to trace errors back to where the AI went wrong. These tools are still early but signal a growing infrastructure for AI output verification.