Using AI for report editing: How to handle citation verification

Conclusion first: Using AI tools like ChatGPT to organize materials and edit reports saves time, yes, but you absolutely cannot skip the final manual verification step. Treat AI as an intern; output must be reviewed.

I just saw a news item where Australia's age verification technology test report, costing AUD 3.48 million (approx. RMB 16.629 million at current exchange rates), was found by The Guardian to contain multiple citation errors. The author admitted using ChatGPT for editing but denied the errors were caused by AI "hallucinations." This is quite typical. Recently, I had my finance team use AI to organize policy documents and took the opportunity to streamline the process. From a financial perspective, this is like auditing an unaudited statement: no matter how fast the books are kept, you cannot skip the confirmation letters.

Day 1: Hand over materials to AI, but build the "cage"

The most common mistake beginners make is dumping news links or PDFs into the chat box and saying "summarize this." AI happily provides a seemingly organized draft, but it may contain content "filled in" from its training data.

My approach is to set boundaries. Open ChatGPT, paste the raw materials, and input this prompt:

1. Use ONLY the information provided in my materials; do not add any background knowledge;

2. After each key fact, note its location in the original material in parentheses (which paragraph, which sentence);

3. If relevant information is missing from the materials, write "Not Provided" directly; do not speculate.

In my testing, adding these rules reduced AI hallucinations/fabrications by 70-80%. But it still makes mistakes. For example, when organizing the news chain about the Australian ban, it automatically merged "96% of teens aged 10-15 use social media" and "70% encountered harmful content" into one conclusion. It looks smooth, but these two figures come from surveys by different institutions with different methodologies.

Day 3: Verify against the original text item by item; this is the lifeline of the process

The report author in the news said the citation errors weren't due to AI "hallucinations." Honestly, from a result perspective, it makes no difference. Whether AI invented a non-existent source or spliced two sentences incorrectly, it ends up as an error in your report. In finance, we call this a "questionable book entry." The only audit method is sending confirmation letters and manually checking against the original source.

The steps aren't complex:

1. Randomly pick 3 items from the AI-provided citation list, copy keywords to a search engine to find the original text;

2. If found, verify if numbers, dates, and subjects match;

3. If not found, mark it as "Suspected Fabrication" and reject the entire list for redo.

Last week, while processing an industry report, AI gave a source claiming it was statistics published by a certain institution last year, with a very specific title. I searched everywhere, but the document didn't exist. Later, I discovered it had mistaken a quote in another report for the title of a standalone report. So, for any citations produced by AI, I now default to treating them as "Unverified" until checked one by one.

Incidentally, discussions on Reddit about the effectiveness of this ban are interesting. Some say the ban failed; data shows teen usage dropped 35%-50% in the first three months; others point out the author didn't read the original report and that the statistical periods overlapped. With materials from both sides presented, AI can help list arguments from both, but determining who is right requires going back to the original report. AI cannot replace this action for you.

One Week Later: Make "Manual Review" a standard step, not an optional one

Once accustomed, I established a division of labor for myself, which I suggest you try:

• Can be delegated to AI: Organizing timelines, extracting key numbers, merging expressions from different sources, drafting frameworks;
• Must be monitored by humans: Citation verification, judgment of policy impacts, wording of final conclusions.

The logic is simple: AI-generated citations are like "unaudited accounts receivable." No matter how pretty the books look, someone must send confirmation letters to verify them. Otherwise, when scrutinized in formal settings, errors will definitely be found. Fixing it then costs far more than spending an extra ten minutes initially.

After learning this three-day plus one-week process, the next step is to try it on a report you actually need to use. Run the full verification suite and compare it with the version AI generated directly. Then you'll understand why I say this step cannot be skipped.