How-To Guide

How do you approve AI-generated social posts?

How teams approve AI-generated social posts: the checks, the approval record, and who signs off before anything publishes.

9 min read

By the Superdirector editorial team. Platform and regulator pages were read on 17 November 2026.

At a glance

How do you approve AI-generated social posts?

Approve an AI-generated post with three checks and one record before it publishes. The copy: every claim, number, name and price in text a model wrote is re-found in a source, and one named person signs it. The depiction: when an image or video shows a real person, place or event, or a realistic scene that did not happen, a person confirms what it shows and that the likeness has permission. The voice: a cloned or generated voice needs written permission before it is used. Then set the platform's AI disclosure at upload. The record stores the exact version shown, who checked which element, when, the prompt and output, the label setting and the outcome. Meta, TikTok and YouTube each require a disclosure setting and none of their pages requires human review, so the review is the brand's own policy. The EU AI Act is the exception that makes it matter legally: AI text on matters of public interest is exempt from the labelling duty only with human review or editorial control.

  • Three checks carry the risk: the words a model wrote, the depiction and its permission, and a generated or cloned voice.
  • A platform AI label is not a check. Meta says it may apply penalties when people fail to use its disclosure tool, and TikTok may remove or restrict unlabelled realistic content.
  • Version drift is the failure that matters. Kontentino describes version three approved and version four scheduled because approval lived in email and content lived in the tool.
Three panels showing what a person checks on an AI post: the copy the model wrote, where a confident claim or price has no source; the image or video, where a realistic depiction of a real person, place or event needs verification and permission; and a voice, where a clone needs written permission first. Each panel names the failure, the check and the record field, and a footer says the platform AI label is a separate setting.

The three checks an AI post needs, drawn from the approval requirements described by Hootsuite, kontentino and Planable and the disclosure triggers on YouTube's own help page, read 17 November 2026. The record fields are the brand's, not the platform's.

Superdirector

Make it work for your brand

Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.

The short answer

Approve an AI-generated post by checking three things before it publishes, and by writing down who checked them. First the copy: every claim, number, name and price in text a model wrote gets re-found in a source, and one named person signs it. Second the depiction: when the image or video shows a real person, place or event, or a realistic scene that did not happen, a person confirms what it shows and that any likeness has permission. Third the voice: a cloned or generated voice needs written permission before it is used. Then set the platform's own AI disclosure at upload and keep that choice in the record.

An approval is finished only when the record exists. Store the exact version that was shown, which element each person checked, when the decision landed, the prompt and output the draft came from, the label setting chosen, and the outcome. That record is what lets you answer a client, a platform or a regulator months later, and the same review is what lifts one EU labelling duty described below.

Two things make this different from approving a human's draft. A model produces fluent text that reads as finished, so a wrong figure does not look wrong. And the output has no author you can ask, so the check has to be aimed at what the model is bad at: unsourced specifics and realistic invention.

Why AI output needs its own approval step

Ordinary social review asks whether a post is on brand, correctly scheduled, and safe to publish. AI output adds failure modes that survive that review, because they look like normal finished work.

A label is not a check, and the platforms are explicit about why detection cannot carry that job. Meta's own newsroom post concedes that it is not yet possible to identify all AI-generated content, and that people can strip out invisible markers (Meta, February 2024). So the AI info tag cannot tell you whether the sentence under it is true.

Hootsuite's healthcare guide draws the operational conclusion for a regulated team: every post needs human review and compliance approval before publishing, and AI drafting should run through governed AI with a traceable record of what was generated, who reviewed it, and who approved it (Hootsuite).

Keep the approval decision separate from the label decision. The label tells a viewer that AI was used. The approval decides whether the post should exist. A post can carry a correct label and still be wrong, and the EU duty below turns on whether a person actually reviewed the words.

Which generated elements need a human check

Six elements carry the risk. The table gives the check a person makes for each, and what the source says.

ElementThe check a person makesWhat the source says
Copy: claims, numbers, names, pricesRe-find each one in a stored page or the client's own record, then sign the sentence.Hootsuite: a clinician or compliance officer should review anything that touches health claims or patient-facing messaging.
Image or video of a real personConfirm who is shown, and hold written permission for the likeness.YouTube lists making a real person appear to say or do something they did not as a disclosure trigger.
Footage of a real place or eventConfirm the footage is not altered in a way that misleads about what happened.YouTube: alters footage of a real event or place.
A realistic scene that never happenedDecide whether the post is clearly fantastical. If it is not, treat it as disclosable.YouTube: generates a realistic scene that didn't actually occur. Non-realistic content needs no disclosure.
A generated or cloned voiceGet written permission before generation. A label cannot fix a missing permission.postora: a label does not cure a missing consent.
The platform AI settingChoose it at upload, then record which setting you chose.Meta says it may apply penalties when people fail to use its disclosure tool.

The order inside the table is deliberate. Copy comes first because a wrong number travels furthest. Depiction comes next because a realistic fake of a real person or place is the one a viewer cannot detect. Voice follows, because permission is a legal question that no later review can repair.

Who should own each check

Give each check one named owner. A committee produces a thread, and a thread produces the failure described later in this guide: the version that was approved is not the version that was scheduled.

The legal default is broader than most teams expect. Under the EU AI Act the company using the system is the deployer, and the Commission's FAQ excludes only use for personal, non-professional activities (European Commission). Employees acting under the company's instructions are not separate deployers, and a company stays the deployer when contractors or freelancers run the system on its behalf. A freelancer who earns from the work on a regular basis is a deployer in their own right.

In practice that means three roles, and one person can hold two of them only when the brand is small. One person drafts the post and flags what a model generated. A second re-finds claims and verifies depictions, and should not be the drafter when the post makes a factual claim. The approver signs the version that will publish and owns the label setting.

Where a client is involved, chain the stages. Kontentino describes a real approval tool as one that lets you chain stages so a post cannot reach the client until it clears internal review (Kontentino). Planable offers multi-level approval with sequential stages on its Enterprise plan and requires an account to approve, while outside stakeholders can comment on a share link (Planable).

Name a fallback approver, and put the deadline in the record. An approval that waits on one person is where a scheduled post slips, and a slip invites someone to publish the previous version. The same rule covers the label: the person who signs the post should be the person who sets the AI disclosure, so the two cannot drift apart.

What platform disclosure actually requires

Each platform asks for a setting, not a signature. The requirement is that the viewer knows, and the decision a brand has to make is when the setting applies.

Meta requires people who post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered to use its disclosure tool, and says it may apply penalties if they do not (Meta, February 2024). Still images are not named in that self-disclosure duty, and Meta labels images itself when it detects industry standard AI signals. Where Meta detects that content was only edited with AI tools, the AI info label moves into the post's menu (Meta, April 2024).

TikTok requires creators to label AI-generated or significantly edited content that shows realistic-looking scenes or people, and unlabelled content may be removed, restricted, or labelled depending on the harm it could cause (TikTok Community Guidelines). A label is not a licence: the same guidelines bar content that misleads about matters of public importance, even when it carries one.

YouTube asks for disclosure during upload when a video makes a real person appear to say or do something they did not, alters footage of a real event or place, generates a realistic scene that did not happen, or creates music that is the main focus (YouTube Help). Minor edits and clearly non-realistic content are exempt, and cloning your own voice for a voice-over sits on the exempt list.

The EU adds a duty that sits above all three. Article 50(4) requires deployers to label deepfakes and AI-generated or manipulated text published on matters of public interest without human review or editorial control (European Commission). For an audience in the EU, the conservative setup on a deepfake is both the platform setting and a visible disclosure inside the content, because a platform label does not travel with a download or a repost on another network (postora).

The approval step inside all of this is small and easy to lose. Someone chooses the setting, and the record names the choice. A reviewer who approves a post without knowing which setting will be applied has approved a different version from the one that publishes.

What the approval record has to keep

Keep six fields per post, and keep them together.

  • The exact version shown, with its hash or a locked copy.
  • The checker, and the element each person checked.
  • The time of the decision, in one time zone.
  • The prompt and the model output the draft came from.
  • The label setting you chose.
  • The outcome, with the note that came with it.

Two lists usually get confused. Version history shows every edit to a post. The audit trail shows who approved what, and when (Kontentino). For a team in finance, healthcare or anything regulated, kontentino treats both as a requirement, not a nicety.

The reason to keep them is the failure they prevent. Kontentino describes the email pattern exactly: version three gets approved but version four gets scheduled, because the approval and the content live in two different systems. Email also leaves no status view, so finding out what is approved means re-reading an inbox.

Set a retention period and write it into the brand's operating notes. None of the pages analysed for this guide states one, so the length is a policy choice the brand has to make. A record deleted after ninety days cannot answer a question asked in month seven, and a record kept forever without an index is barely better.

A six field approval record: exact version shown with its hash, checker and element, model output, label setting, time of decision, and the decision itself. A footer explains that version history lists every edit while the audit trail lists who approved what and when.

The six fields that make an approval retrievable, organised from the version history and audit trail distinction that kontentino and Planable each describe on their own pages, read 17 November 2026. No field is supplied by a platform.

A decision list to run on every post

This list turns the sections above into seven questions. Run them in order on every post that used a model, and stop only at the end.

  1. Did AI write or change the words? If yes, re-find every claim, figure, name and price in a source, and have a person sign the copy.
  2. Does the image or video show a real person, place or event? If yes, confirm what it shows and hold written permission for the likeness.
  3. Did AI make or clone a voice? If yes, get written permission before generation, not after it.
  4. Will the post use the platform AI label? Set it at upload, and record which setting you chose.
  5. Is the text on a matter of public interest? If yes, someone with editorial responsibility must revise it, because spell-checking does not count as review.
  6. Is the audience in the EU and the content a deepfake? If yes, add a visible disclosure inside the content.
  7. Record the decision, with the version, the checker, the time, the model output, the label setting and the outcome.

Steps one to three produce the checks. Steps four to six produce the disclosure. Step seven produces the record, and a post without it has an approval nobody can audit. Post that needs a second pair of eyes before it reaches a client at all? That is the narrower question in which posts need customer review, and this list assumes the answer is already yes.

Seven numbered approval questions in order: words, depiction, voice, platform label, public-interest text, EU deepfake disclosure, and the record. Notes mark the stop conditions for the first three and the step that can never be skipped.

The decision order this guide recommends, built from YouTube's four disclosure triggers, Meta's disclosure tool and the European Commission's summary of Article 50(4), all read 17 November 2026. It is a policy sequence, not a rule published by any platform.

What breaks an approval in practice

Four failures account for most of the damage, and each has a visible symptom.

The approver sees a mockup instead of the live crop. The post that publishes is truncated, cropped or re-flowed, and the yes was given to a different picture. Approve the exact image and caption that will post, which is its own decision in approving the exact image and caption.

The approval happens where the content does not live. This is the email pattern above: the decision and the post separate, and the wrong version ships.

Someone sets the label after the approval. The reviewer approved a post without the AI info setting, and that setting changes the version that goes out. If you want the disclosure rules themselves, they are in do you have to label AI generated posts.

An edit lands after sign-off. The fix is mechanical: lock the approved post as read-only and schedule only the locked copy. Planable describes exactly that, turning approved posts read-only and scheduling them once they are approved.

Where Superdirector fits

Superdirector runs an operated social account. The $199 plan covers one AI-run account that publishes on account-level standing posting permission. The account is authorised once, under the brand's standing rules, and posts follow that authority.

That design leaves the approval policy with the brand. Where a brand wants a person to check copy, depiction or voice before a post goes out, those checks run inside the brand's own workflow, and the record stays with the brand. The $29 plan is MCP access with hands-on creation, for teams that build and post their own work.

The practical setup is the decision list above, kept in the brand's operating notes, with the record stored beside the post. A brand that writes the list down once can hand it to a new person without re-deriving it, and a client can read the same list and see what was checked. An agent that posts without a person in the loop raises a different set of questions, covered in guardrails before an agent posts.

Frequently asked questions

How do you approve AI-generated social posts?

The platforms require disclosure, not approval. Meta, TikTok and YouTube each ask for an AI setting at upload and warn about penalties, removal or restriction when realistic content goes unlabelled. Human review enters through the EU AI Act: AI text on matters of public interest is exempt from the labelling duty only when it has had human review or editorial control and someone holds editorial responsibility. A brand's own policy is what puts a person on the copy, the depiction and the voice.

Which parts of an AI-generated post should a person check?

Six: the claims, numbers, names and prices in the copy; any portrait of a real person; footage of a real place or event; a realistic scene that never happened; a generated or cloned voice; and the platform AI setting. Copy comes first because a wrong figure travels furthest, and a voice comes after the depiction because permission cannot be repaired by any later review.

Who should approve AI-generated social posts?

One named owner per check. The drafter flags what a model generated, a checker re-finds claims and verifies depictions, and an approver signs the version that publishes and owns the label setting. Under the EU AI Act the company is the deployer, so responsibility does not pass to the employee who clicked approve, and it does not leave the company when a freelancer operates the system on its behalf.

Does a platform AI label count as approval?

No. A label tells the viewer that AI was used, and says nothing about whether the post is accurate or the likeness is permitted. Meta says it may apply penalties when people fail to use its disclosure tool, and TikTok may remove or restrict unlabelled content. postora puts the limit plainly: a label does not cure a missing consent.

What should an approval record contain?

The exact version shown, the checker and the element checked, the time of the decision, the prompt and model output, the label setting, and the outcome. Keep version history, which lists every edit, beside the audit trail, which lists who approved what and when. Set a retention period and write it down, because none of the platform pages names one.

Can AI write about a matter of public interest without a label?

It can when the text has had human review or editorial control and a person or company holds editorial responsibility for it. The European Commission's FAQ is strict about what counts: spell-checking or grammar fixes do not. For a brand posting about product safety or a public health issue, the practical step is a reviewer who can change the sentence, not one who approves it unchanged.

Related practical guides

Write the three checks into your posting routine

Name one person for copy, one for depiction and voice, and one approver who sets the AI label, then keep the six-field record with every post that used a model.