Generative AI models produce text based on statistical pattern matching. Models often produce accurate statements, but they do not understand accuracy or factuality. You are accountable for the accuracy of your work. Use these techniques to verify AI output before use.
AI-Based Validation: Chain-of-Verification (CoVe)
CoVe is frequently implemented as part of applications that integrate generative AI, but you can adapt it to your daily workflow. In the CoVe framework, after its initial output, the AI:
- Analyzes its own draft and formulates fact-checking questions for each claim it makes.
- Checks answers to those questions without referencing any of its previous output—so that it does not just repeat prior errors.
- Compares its initial response against the results of the fact-checking questions and revises by correcting inaccuracies and removing unverified assertions.
To adapt CoVe to your workflow, start a fresh chat instance and ask the chatbot to factcheck its prior output:
Attached is a copy of a report I am writing for the Provost’s office. Go through the report claim by claim and perform a systematic verification: identify each factual, numerical, and contextual claim in the document and organize them into two categories: inaccurate claims (that are contradicted by external evidence or logical errors); and unverifiable / internal claims (that you cannot substantiate with the information available to you). When you are done, 1) research and correct all inaccurate claims; 2) output a list of unverifiable / internal claims that I still need to check; and 3) output a revision of the report with all the corrections you have been able to make.
Check Your Work Yourself: The SIFT Framework
Originally developed for digital literacy by Mike Caulfield, the SIFT framework is a systematic way to evaluate AI-generated claims, facts, and references:
- Stop: Pause before accepting or distributing AI output. Remind yourself that AI models are designed to generate plausible-sounding text, not verified facts. Never rely on an unverified first draft.
- Investigate: Check specific authors, datasets, and citations in AI-generated output. Verify that the named experts exist, that the cited papers are real, and that the credentials match the subject matter.
- Find Better Coverage: Even when AI citations are accurate, they may not be the most credible sources. Use trusted, independent databases like those found on the library website to find the best source you can on the topic. Or scan multiple sources to see what the expert consensus actually is.
- Trace Claims to the Original Context: For each claim in the output, locate the original primary source—the book, article, dataset, etc. that that the supposed fact comes from—and confirm that the AI output did not take quotes out of context or misrepresent the findings.
To get the best out of AI-generated output—and to ensure that you can stand confidently stand behind your work—use both of these methods. AI can extend your capabilities, but it cannot replace your critical reading capabilities or your expertise.
Before using generative AI in your work, please consult the College policy on administrative use, as well as the data classification guidelines.