Guide
How managers can detect AI in workplace reports
Reports, proposals and summaries are now routinely drafted with AI tools, and managers are being asked whether a given document was written by the person who submitted it. This guide sets out when checking is worth doing, which signals matter, how to use a detector without creating an unfair case, and how to raise the question with an employee.
Start with policy: should you check at all?
Whether AI use in a report is a problem depends on your organization's rules, not on a detector. If your policy allows AI drafting with disclosure, the real question is whether use was disclosed and whether the facts are right. If your policy prohibits AI for certain documents, such as client deliverables or regulatory filings, checking authorship becomes relevant.
Without a written policy, an employee cannot know what is expected, and any check is hard to defend. The NIST AI Risk Management Framework, released on January 26, 2023 and "intended for voluntary use," is a common starting point for organizations writing AI governance rules. A policy should say which uses are allowed, when AI use must be disclosed, and what happens when someone suspects misuse.
Signals worth noticing in a report
None of these signals proves AI use. They are reasons to read more closely. The strength column is AFIP's judgment of how much each signal tells you on its own.
| Signal | Why it matters | Strength on its own |
|---|---|---|
| Missing details only the author would know | Real work leaves traces: project names, numbers, decisions, people | Moderate |
| Sources, figures or quotes that do not exist | Generated text can invent plausible references; these are checkable facts | Strong for accuracy, not for authorship |
| A sharp change from the author's usual writing | Style shifts can reflect AI drafting, but also editing help or a template | Weak to moderate |
| Generic claims that would fit any company | Common in unedited AI drafts, and also in rushed human writing | Weak |
| A high score from one AI detector | Detectors misclassify human writing, especially short or non-native English text | Weak; a lead only |
Running an AI detector responsibly
If you use a detector, treat its output as one input. Independent testing of 14 detection tools found them "neither accurate nor reliable" (Weber-Wulff et al., 2023), and OpenAI withdrew its own classifier in 2023 "due to its low rate of accuracy" after warning it "should not be used as a primary decision-making tool" (OpenAI, 2023).
- Check confidentiality first. Reports often contain client, financial or personal data. Do not paste a confidential document into an online tool your organization has not approved.
- Scan substantial text. Detectors are weakest on short passages, so scores on a paragraph or an email tell you little.
- Use a second detector. A second tool that uses a different method makes a single anomalous score easier to spot.
- Keep a record. Note the tool, date, text checked and result, so any later review can see exactly what was done.
AFIP's guide on why AI detectors get it wrong explains the error patterns in detail, and the guide to telling whether text was written by AI sets out a five-step cross-check.
A decision path for managers
- Does your written policy restrict AI use for this kind of document? If no, treat any concern as a quality or accuracy issue and give normal feedback.
- If yes, is the concern about accuracy or authorship? If accuracy, verify the facts and sources first; errors matter whoever wrote them.
- If authorship, what evidence do you have beyond a score? If only a score, stop at a conversation; do not open a formal process.
- If there is more evidence, such as invented sources or no trace of the underlying work, have the conversation below and ask how the report was produced.
- If concerns remain, follow your organization's formal process and involve HR, keeping the record you made.
Talking to the employee: a neutral script
The aim of the conversation is to understand how the report was produced, not to win an argument about a score. Opening without accusation keeps the facts clear and protects you if the concern turns out to be wrong.
- "I'd like to understand how this report came together. Can you walk me through how you put it together?"
- "Where did the figures in section two come from? I'd like to check them with you."
- "Our policy asks us to disclose AI drafting for client documents. Did you use any AI tools on this one?"
- "Thanks for explaining. Going forward, please note any AI assistance at the top of the document."
What not to do
- Do not discipline anyone on a detector score alone.
- Do not run only one employee's work through a detector when others doing the same work are not checked.
- Do not paste confidential material into unapproved online tools.
- Do not assume formal or polished writing means AI; many people write that way.
Hiring teams facing the same question with job applications can use AFIP's guide to screening AI-drafted resumes and cover letters.
An AI detector score is a reason to ask questions, never a finding about an employee.
For related verification methods, see AFIP's guides on why AI detectors get it wrong and telling whether text was written by AI, or browse all AFIP verification guides.
Frequently asked questions
Can my employer tell if I used AI?
Sometimes, but not reliably from a detector alone. Employers notice missing project details, invented sources and sudden changes in writing style, and some run AI detectors, which make errors in both directions. If your workplace allows AI with disclosure, disclosing your use is the simplest way to avoid misunderstandings.
Is it legal to run employee writing through an AI detector?
It depends on where you are, your employment contracts and your privacy obligations, especially if the documents contain personal or client data. Many organizations restrict which online tools can process company documents. This is not legal advice; check with your legal or privacy team before introducing routine AI checks.
How accurate are AI detectors on business writing?
No widely cited independent benchmark focuses on business reports. Published tests on other kinds of text found detectors neither accurate nor reliable, with errors rising on short and edited text. Treat any score on a report as a lead to verify against the facts and the author's account.
Should a company ban AI for reports?
Not necessarily. Many organizations allow AI drafting with disclosure and human review of facts, which is easier to enforce than a ban that relies on unreliable detection. Whatever you choose, write it down, explain it to staff, and say how suspected misuse will be handled.