Reviewing AI-written captions: a checklist before anything is approved

AI writing tools have made the first draft of a social post almost free. That changes where the work is. Writing used to be the slow part and review the quick one; now a writer can produce five variations before the kettle boils, and the person approving them has five times as much to read. The risk is not that AI drafts are bad. Many are perfectly good. The risk is that they are plausible, and a plausible caption with a wrong date, an unsupported claim or the wrong tone slides through a tired review. This checklist is for whoever signs posts off, whether they are the writer checking their own AI-assisted draft or a client approving it.

Why AI drafts need a different kind of review

When a person writes a caption from scratch, mistakes tend to be visible: a clumsy sentence, a typo, an awkward joke. AI-assisted drafts fail more quietly. They are fluent, well structured and confident, which makes the reader trust them. The common problems are of a different kind:

• Filled-in details. Asked for a post about an event, a model may supply a plausible time, a price or a “limited places” line that nobody mentioned.
• Generic claims. “Industry-leading”, “the best in town”, “guaranteed results”: phrases that sound like marketing and may not be true or allowed.
• Borrowed voice. A draft that sounds like every other brand rather than this one.
• Misplaced confidence about networks. A caption with a clickable-looking link written for a network where captions cannot carry links, or hashtags stacked where they do not belong.

Good tools reduce these problems; none removes them. Review is still the step that makes a post safe to publish.

The checklist

Work through it in order. The early checks are the ones where a mistake costs the most.

1. Facts: is every detail in the brief?

Compare the caption with the brief, line by line. Every date, time, price, product name, location, opening hour and offer condition should appear in the brief or in a source the writer can point to. If a detail is in the caption and not in the brief, treat it as invented until proven otherwise.

• Dates and times, including the day of the week (does 13 November really fall on a Friday?)
• Prices, discounts and their conditions
• Names of products, people, partners and places, spelled correctly
• Links: do they open, and do they go to the right page?

2. Claims: could anyone object?

Look for anything that promises a result, compares the client with a competitor, or touches health, money, safety or the law. These are the sentences regulators and competitors read.

• Superlatives (“the best”, “the only”, “number one”): can the client prove it?
• Guarantees and outcomes (“you will save”, “results in a week”)
• Health, financial or environmental claims
• Comparisons with named or recognisable competitors

When in doubt, cut it. A plainer caption that is true beats a punchier one that needs a lawyer.

3. Disclosures and permissions

• Paid partnerships, gifted products and affiliate links are disclosed in the way the network and local rules expect
• People who are tagged or pictured agreed to it
• The client owns or has licensed every image, video and music track

4. Voice: does it sound like this brand?

Read the caption aloud. Would the client say it this way? Check it against the brand’s own rules: words they use, words they avoid, how formal they are, whether they use emoji. AI drafts often drift towards a cheerful, generic tone. If three posts in a row open with an exclamation, the calendar will read like a template.

5. Fit for each network

The same caption rarely suits every network. For each one the post is going to, check:

• Length: it fits the network’s limit and its preview (the first line or two). The character counter checks seven networks at once.
• Links: Instagram and TikTok captions do not carry clickable links; say where the link is instead.
• Hashtags: a few relevant ones, placed where that network expects them.
• Media: the caption describes or complements the image or video that is actually attached.

6. Accessibility

• Every image has alt text describing what is in it
• Multi-word hashtags are capitalised (#SmallBusinessSaturday, not #smallbusinesssaturday) so screen readers can read them
• Emoji are not used in place of words, and not in long runs
• Videos have captions, and text inside images is short and legible

7. Timing and context

Finally, read the post against the calendar and the news. A cheerful promotion scheduled the morning after a local tragedy, or a post about a sale on the day the shop is closed, is a mistake no draft can foresee. This is the check only a person can do.

Make the review fast without making it shallow

A long checklist only works if it is quick to apply. A few habits help:

• Write better briefs. Most invented details come from gaps in the brief. If the brief has the date, the price, the link and the one thing that matters, there is less for the AI to fill in.
• Batch reviews. Reviewing a week of posts in one sitting is faster than ten interruptions, and repeated phrases are easier to spot side by side.
• Keep feedback on the post. A note attached to the post (“the offer ends Sunday, not Saturday”) is seen by the writer at the moment they edit, and leaves a record of what was approved.
• Use a printed list once. Give approvers our approval checklist at the start, so their attention goes to the checks only they can make.

Who is responsible?

The person or team that publishes the post is, whichever tool wrote the first draft. Agree inside the agency, and with each client, that AI-assisted drafts go through the same approval as any other post, and that nobody publishes AI output unread. Many clients will ask whether you use AI at all; an honest answer, with a description of your review process, is usually reassuring.

Doing this in PostRoost

PostRoost’s AI writing assistant drafts from a brief in the client’s brand voice, with a version per selected network within its character limit, and offers hashtags, three alternative tones, a more engaging rewrite, grammar fixes and a shortened version. It is instructed never to invent facts, prices, dates or claims that are not in the brief, but like any AI it can be wrong, so suggestions are only inserted when you choose them.

Its risk check lists possible reputational, legal, cultural or accessibility issues without changing your text, a useful second pair of eyes for steps 2, 6 and 7 above, but not a replacement for them. With approval workflows switched on for a client (Team plan and above), AI-assisted posts wait for the same sign-off as any other, and reviewers can request changes with a note. The Free plan includes 5 assistant requests a week; paid plans are unlimited.