AI Use & Verification

Which AI tools I used

I used an AI assistant (Hermes) throughout the project — to pull and summarize data from the NAEP API, write the Python analysis and charts, draft and critique explanations, find and verify sources, and challenge my own interpretations.

How AI helped

• Data access and analysis: wrote and ran the code that queried the NAEP API, computed summaries, and generated the charts.
• Interpretation: helped phrase findings in plain language for an education audience.
• Critique: pushed back on overclaiming — for example, it argued me down from a causal claim to a descriptive one.
• Research: surfaced credible sources and checked DOIs against Crossref.

What I verified myself

• Every number. I re-queried the NAEP API directly and confirmed the scores matched; I did not take the AI's figures on faith.
• Every source. I checked the Kuhfeld et al. DOI against Crossref, confirmed the NCES figures on nationsreportcard.gov, and verified the Recovery Scorecard's wording on the original report.

Errors, limitations, and overclaims I noticed and corrected

• An AI paraphrase overstated the Recovery Scorecard's finding as "a full grade level behind" when the report says "nearly half a grade level" — I corrected it.
• The AI initially suggested benchmarking Texas against similar-demographic states; I declined, because for this descriptive project the raw national-vs-Texas gap is itself the finding.
• The AI flagged that NAEP is biennial and sample-based, which shaped how carefully I phrased the before/after claims.

What I changed based on my own judgment

• I kept the project descriptive even when it would have been easy to drift toward causal language.
• I kept the demographic context front and center in the comparison, rather than statistically adjusting it away.

What I learned about using AI responsibly

AI is a strong assistant and a poor authority. It accelerates the work, but every number, source, and claim still has to be checked by a person. The discipline of "describe, don't overclaim" came from my own judgment — the AI could support it, but I had to hold the line.