Which tool to choose
The number-one risk of a review written with AI is that it sounds fake or slips in details you never verified. Choosing the tool serves precisely to reduce this.
If you're writing a long, detailed review (a book, a course, a week-long stay, an appliance used for months): use Claude in the free version. It has a wide context window, designed to handle long documents: it lets you paste notes, photos of receipts and a raw list of pros and cons, and have them all reorganized together. The free plan includes file uploads and writing features, with a daily message cap that's reached sooner during peak hours. If you hit it, you wait for the reset (a few hours) or split the work into two sessions.
If you write short, frequent reviews (Amazon, Google Maps, TripAdvisor reviews of a few lines): free ChatGPT is fine. The free plan gives you a certain number of messages on the best model in a window of a few hours, then falls back to a lighter version until the reset. For short texts this cap won't bother you, and the free version also searches the web and reads images or files — useful if you want it to check the exact name of a model or product.
For this task there's no need to pay: both free versions hold up. The paid plan (about $20 a month) is justified only if you write reviews at an industrial volume and hit the limits every day.
How to do it
The principle that underpins everything: the AI doesn't know what you experienced. If you don't tell it, it invents it. So the real work is gathering your real facts before opening the chat.
Write down your real experience off the cuff, in disorder. Don't worry about the form. Note: what you bought/used, for how long, three things you genuinely liked, three that disappointed you, a specific detail (a number, an episode, a date). This material is your anti-invention reservoir.
Decide your verdict before writing. Would you recommend it? To whom yes and to whom no? If you're not clear on this, the review comes out limp. Honesty starts from a position, not from neutrality at all costs.
Open Claude or ChatGPT (from a browser or from the app, the path is the same: you paste the prompt and your notes into the message box).
Paste this prompt followed by your raw notes:
You are an assistant who helps me write an honest, balanced review.
Strict rules:
- Use only the facts I provide below. Don't add
features, numbers, strengths or flaws that I didn't write.
- If a useful piece of information is missing, don't invent it: flag it for me at the end
with the note "To verify".
- Keep a neutral, personal tone, in first person. No advertising
tones, no inflated enthusiasm, no gratuitous trashing.
Structure the review like this:
1. Context in one line (what I used, for how long, to do what)
2. What works (the concrete strengths I gave you)
3. What doesn't work (the concrete flaws I gave you)
4. Who it's suitable for and who it isn't
5. Final verdict in two lines
Length: about 200 words. Here are my notes:
[here you paste your notes from step 1]
Read the draft and hunt down every sentence you don't recognize as yours. It's the step that makes the difference between a real review and a generated one. If you find a detail you hadn't provided, delete it or ask the AI to remove it.
Feedback: a well-done review makes you think "yes, that's exactly what I wanted to say but written better". If you think "I'd never have written it like that" or "I never said this", go back to the previous step.
Ask for one final balance check with a follow-up message:
Reread the review and tell me: does the balance between strengths and flaws reflect
what I told you, or does it lean too far one way? If it's unbalanced,
rewrite it restoring the right weight, always without adding new facts.
Concrete example
Marta bought a 250-euro robot vacuum and has been using it for two months. She wants to leave a review on Amazon but when she tries it comes out either as a list of praise like an ad or two annoyed lines.
She opens ChatGPT (short review, free is fine) and pastes her raw notes: "used 2 months, 80 sqm home with a cat / cleans hair well, battery lasts the whole home, easy app / it gets stuck under the low sofa almost every time, the dust bin is small I empty it every 2 days, noisy on powerful mode / I'd recommend it to people with pets but not to those with low furniture". She adds the prompt above.
The AI gives back a 190-word review: context (cat, 80 sqm, two months), three concrete strengths, three concrete flaws, and the closing "suitable for people with pets, less so for those with low overhanging furniture". At the end it flags: "To verify: the exact model and the bin capacity in liters, if you want to cite it". Marta checks the box, adds "0.3-liter bin" and publishes. Total time: seven minutes, and every word corresponds to what she experienced.
When it does NOT work (and how to fix it)
If the AI invents features you never mentioned
It happens especially with famous products, whose spec sheets the model "knows". The fix is already in the prompt (the rule "use only the facts I provide"), but if it slips through anyway, reply: "You added this sentence, I didn't tell you it. Remove it." Never trust numbers, percentages or feature names you didn't provide: the AI generates plausible but wrong information, and on a reviews platform that exposes you to disputes.
If the review sounds fake or "AI-like"
The problem is usually an excess of adjectives and three-part phrases ("comfortable, practical and elegant"). Ask: "Rewrite it more sparely, remove the decorative adjectives and leave only facts and direct experience. I want it to sound like a real person, not a product sheet." A very specific, personal detail (an episode, a time of day, a small funny flaw) is what makes a text human.
If you have to review something you haven't finished using
Don't ask the AI to "complete" the verdict: you'd be writing falsehood. Declare it in the review itself. Add to the prompt: "Specify that it's an impression after X days of use and not a final verdict." An honest review that admits its own limits is worth more than a complete but inflated one.
If the product or service only disappointed you (or only thrilled you)
Balanced doesn't necessarily mean fifty-fifty. If a product is terrible, an honest review is negative. But still make the effort to find the one thing that worked and the one situation in which it might be fine for someone else: it makes the verdict credible instead of spiteful. Ask the AI: "Even though my verdict is negative, help me point out the one case in which it might still be useful to someone."
A tip from someone who actually uses it
Don't have the AI write the title or the first sentence. That's where the artificial hand shows the most, and it's the part the reader looks at first. Write the opening yourself in your own words — even crooked — and leave the central body to the AI, where order matters more than style. A review that opens with your real voice and continues in an orderly way beats any perfect but anonymous text.
Frequently asked questions
Do platforms like Amazon or Google ban reviews written with AI?
The policies ban false reviews or ones not based on real experience, not the use of a writing tool. If the review recounts a genuine experience of yours and you check every statement in it, you're within the rules. The problem arises when the AI invents: at that point it's no longer your experience. To be safe, avoid publishing the AI's text word for word and rewrite at least the opening.
Do I have to say I used AI to write it?
For a consumer review there's no obligation, just as you don't declare you used the spell checker. What counts is that the content is true. It's different if you review for work or are bound by specific editorial rules: in that case, follow those.
How do I avoid seeming biased if the product was gifted or sponsored to me?
Here transparency is mandatory, not optional: declare it openly in the review ("product received free of charge"). Then ask the AI to give explicit weight to the flaws: "I received this product for free, I want to be extra-critical so as not to seem biased. Give fair space to the problems I flagged for you." A declared freebie with real flaws is more credible than praise with no disclaimer.
Isn't it dishonest to have a machine write a verdict that should be mine?
It's the right objection to raise, but it confuses two different things. The verdict stays yours: you're the one who decides whether to recommend, which strengths and flaws matter, what you experienced. The AI doesn't express opinions in your place, it lines up yours. It becomes dishonest only when you delegate the experience too — that is, when you make it invent things you didn't live. As long as you bring the facts and check every sentence, you're using a writing tool, not a substitute for your judgment.