Imagine sending an AI-written weekly report, then freezing when your manager asks where its main number came from.
Before sending the report, check whether you can explain its main claim and show where the number came from.
A tight deadline can make you skip the checks
Suppose you need a summary of customer complaints before a meeting. AI gives you tidy themes, and you copy them into the slides.
Then someone asks whether the angry messages came from new customers or repeat buyers. You never checked, and the summary did not say.
The risk starts when you present a draft as work you understand.
A 2025 Microsoft and Carnegie Mellon survey found that workers with more confidence in AI for a task reported less critical thinking.
This survey does not prove that AI makes people less smart.
Before that meeting, you could read the complaints behind each theme and mark any grouping the source data does not support.
Give AI the facts before asking it to rewrite
Say a delivery is late and you need to email a customer. You know the tracking status, but your first draft sounds defensive.
AI can offer a calmer draft that you can adjust to fit the situation.
Give it the confirmed status and ask it to leave unknown dates blank. Read the result for promises you did not authorize.
A smooth apology becomes a problem if it adds a refund or promises delivery tomorrow.
We would keep the original facts beside the draft and remove anything the company cannot deliver.
Ask it to explain any phrase you do not know before sending it. You still need to know what the email promises.
A polished report can cite work that never existed
In October 2025, Deloitte Australia faced scrutiny over a government report citing research that did not exist and a fake quote from a court judgment.
AP reported that the revised report disclosed AI use and Deloitte agreed to repay the final instalment of its contract.
An agreement to repay does not confirm payment or reveal how each error entered the report.
The department's October statement addressed corrections to references and footnotes. A formal layout had not made the sources accurate.
You could face a smaller version while preparing a purchase proposal. An impressive paper title gives your recommendation weight until someone tries to find it.
Open the cited work before using it. Check that the author and title exist, then read the paper to see whether it supports your claim.
Asking the same bot again is not an independent check
In the 2023 Mata v. Avianca sanctions ruling, the court described filings that cited nonexistent cases generated by ChatGPT.
The record also shows lawyer Steven Schwartz asking ChatGPT whether the cases were real. It assured him that they were available through legal databases.
The court questioned his changing account of when he asked.
A further reply from the same tool did not prove that its earlier sources existed.
If you ask whether a product specification is correct, another confident answer still needs a document from the maker you can open and read.
Another model may flag a mistake, but matching answers do not prove the claim. Use an independent source, not a vote between chat windows.
Correct sums can still tell the wrong story
Consider a hypothetical sales report where orders rise from 80 to 100. If an AI draft calls that 20 percent growth, its baseline is wrong.
The increase is 20 orders, divided by the original 80, which gives 25 percent. Dividing by the final 100 instead gives 20 percent.
Now reverse the movement. A fall from 100 to 80 is a 20 percent decline, so the same counts need different percentages.
Use a calculator to check the arithmetic, then check whether the comparison makes sense.
A correct rate can still mislead if you compare a full month with only part of another.
Ask which records and dates produced the number. You still need to choose the comparison, even when a spreadsheet or calculator performs the arithmetic correctly.
The model can fill a gap with something that sounds right
If you ask AI for your rental notice period without sharing the agreement, it lacks the document needed to check that claim.
A language model generates text using learned patterns and the supplied context. Familiar wording can fit the request while getting your actual terms wrong.
The NIST AI risk guide describes this risk as the production of:
confidently stated but erroneous or false content
Search and document retrieval can provide evidence, but a system can still select the wrong passage or misstate what it says.
In our notes search test, the requested collision figure meant how many random IDs give a 50 percent chance of a repeat. The model gave the IDs' bit count instead.
For the rental example, provide the relevant clause and check the summary against it. Leave terms unresolved when the document does not supply them.
The answer you want can steer the reply
Imagine asking AI to prove that a new checkout page caused higher sales.
A helpful-sounding reply may build the case you requested while overlooking a holiday promotion or a change in traffic.
Research on sycophancy found that assistants can favor agreement with user beliefs over truth. Feedback that rewards agreement can help cause this behavior.
That does not mean every model always agrees. It does mean a pleasing explanation is a poor test of whether your assumption holds.
Ask what evidence would distinguish the page change from the promotion. You can request competing explanations without pretending the chatbot settles which one is true.
We would keep an untested explanation marked as untested. A report that admits a gap gives you something concrete to investigate.
A finished task does not prove you learned the skill
If AI writes a formula you cannot explain, you may finish today's spreadsheet and get stuck when tomorrow's columns change.
You can use it as a tutor instead. Ask for a small worked example, predict the result yourself, then compare your reasoning with the explanation.
In our local model comparison, a generated date parser scored six out of eight by rejecting every test input.
Its two valid dates both failed. All six invalid dates passed because the tests required an error, so the total concealed code we could not use.
You would miss that by copying the headline score. Reading each result shows you what the code does.
For your spreadsheet, change one input and explain why the result should change. If you cannot, ask about that step before relying on the formula.
Check the work before you put your name on it
Return to the weekly report. Before sending it, keep the underlying records open and use this review to find what still needs checking.
- Give AI the relevant material without sharing data your workplace does not permit.
- Separate confirmed facts from suggestions and missing information.
- Open important references and check whether they support the claim.
- Recalculate numbers and confirm the baseline, units and dates.
- Explain the main conclusion yourself, including what the evidence cannot establish.
- Check any promises or actions before the draft leaves your hands.
If your report uses an API response, our JSON formatter can help you read its fields. It cannot confirm that the values are true.
Our AI search fact-checking guide covers checking linked evidence. We would prefer a clear unknown over a made-up detail that makes the report look complete.
Pick one AI-assisted task today and verify its main claim before you send it.