How to Humanize Performance Reviews and Difficult Feedback After ChatGPT
ChatGPT review drafts are balanced and empty. Keep specific incidents. Human Writes is a voice pass, not a substitute for the conversation. Do not invent examples.
Managers already use ChatGPT for the hard paragraph: the review, the PIP opener, the "we need to talk" note. Writing is a common workplace use of these tools. The draft comes back even, professional, and unusable. Every competency is "an opportunity." Nobody did anything on a Tuesday.
Human Writes is the voice pass after you supply the incidents. It is not a substitute for the conversation. If you would not say the sentence in a 1:1, do not put it in the file.
Not a weekly status: Humanize Work Reports. Not an inbox ping: Checklist for Writing Human-Like Emails. If the review will live in a personnel file, treat privacy like any other sensitive document (ethical AI use): do not paste names, medical details, or unreleased HR cases into a consumer chatbot your company has not approved.
What AI review drafts get wrong
| They produce | The person hears |
|---|---|
| Balanced praise/criticism in every section | You did not look at the year |
| "Takes initiative" with no example | Template |
| Softened ratings that contradict the box you ticked | Gaslight |
| Hedged asks ("might consider perhaps") | No decision |
| Identical voice for every report | Mail merge |
Difficult-feedback drafts fail the same way: they skip the incident and jump to "moving forward." The person cannot tell what to change.
Lock examples before any rewrite
Write these by hand. Do not let the model fill blanks:
- Incident (week, project, what happened)
- Impact (who waited, what slipped, what improved)
- Pattern or one-off (be honest)
- Ask (one behavior, observable, dated)
- Support (what you will do, not a slogan)
If (1) is empty, you are not ready to humanize. You are ready to look at the calendar.
Company policy first. Some employers restrict personnel text in consumer AI. Use the approved tenant when required. Redact surnames if you must draft in a tool that is not cleared for HR.
Where Human Writes helps
| Pass | Use it? | Why |
|---|---|---|
| Rating boxes / compensation language | No | Legal and HR own the numbers |
| Competency slogans | Delete, replace with (1)–(4) | Models love "synergy" |
| Narrative paragraphs | Yes, one pass | This is where AI sounds like every other review |
| The live conversation | Never | You still have to talk |
Paste stiff narrative into Human Writes once. Put back the week and the proper nouns. Then schedule the meeting. The document does not deliver the news.
Before and after
AI draft:
Jordan is a valued team member who consistently demonstrates a strong work ethic. Moving forward, there is an opportunity to be more proactive in stakeholder communication and to continue leveraging strengths.
After your incidents + one humanize pass:
In May, the client saw the deck before I did. That happened twice. I need a draft in Slack 24 hours before external sends. I will block Thursdays for a 15-minute review until September. The Q2 launch work was strong; this note is about the send path, not the craft.
Same form. Second version can survive "when did this happen?"
Difficult feedback is not a softer email
The email checklist still applies: one specific, one ask, read aloud. Reviews add three rules:
- No surprise in writing that you skipped in person. If the first time they hear "needs improvement" is the PDF, you failed the job.
- Do not invent a positive to "balance" a real problem. Fake praise trains people to ignore you.
- Do not let the model write the PIP. Structure, maybe. Facts and next steps, you.
What not to do
- Do not humanize until the examples are yours.
- Do not run five passes until every sentence is the same length.
- Do not paste the whole team's reviews into one prompt "for consistency." You will get one voice and mixed-up incidents.
- Do not treat a detector score as the point. The person in the chair is the point.
Workflow
- Fill the five example lines without a model.
- Draft structure in ChatGPT with those lines pasted in and a ban on invented incidents.
- Cut every sentence that could apply to anyone on the team.
- One Human Writes pass on remaining stiff prose.
- Read aloud. If you would wince in the 1:1, rewrite.
- Hold the meeting. Then file the document.
Bottom line
Reviews fail when they could belong to anyone. Human Writes fixes rhythm after the Tuesday is on the page. It does not have the conversation for you.
Paste the stiff narrative on Human Writes when the incidents are locked.