AI content generation is fast, cheap and wrong often enough that you cannot publish its raw output. That last part is not a complaint about any particular model. It is how generation works: a model produces the most plausible continuation of your prompt, and plausible is not the same as true, on-brand or free of visual defects. The useful question is not whether to use AI content generation. It is what has to happen between the moment a draft appears and the moment a person publishes it.
This piece is the checklist I use. Each check is described by what it looks at, who or what owns it, and what happens when it fails. If a check has no fail action, it is a suggestion, not a check.
Why raw AI content generation output is not publishable
Three properties of generated content make a review step unavoidable.
First, the model has no idea which of its sentences are facts. It will state a product capacity, a date or a comparison with the same fluency as a greeting. Second, the model has no memory of your brand beyond what you put in front of it this time. Voice rules you gave it last week are gone unless the system loads them again. Third, image and video models make physical mistakes that a language model would never make in text: extra fingers, unreadable labels, a product that changes shape between frames.
Each of these needs a different kind of check, and none of them is solved by asking the model to be careful.
The checks between generation and publishing
Here is the sequence, in the order a piece meets it.
1. Fact check against a brand factbook
- What it checks: every number and every factual claim in the draft.
- Against what: a factbook per brand, the single source of approved facts, figures and wording.
- Fail action: the unsupported claim is blocked, and the draft uses wording the sources support.
The key design choice is that the fact-checker is a separate step from the writer. A model asked to write and verify in the same pass tends to approve its own work. A separate fact-check agent with only the factbook and the draft in front of it has one job, and it is easier to see when it fails.
This is also why the factbook has to be per brand. If two brands share one pool of facts, the checker will happily confirm brand A's figure inside brand B's post.
2. Voice and editorial rules
- What it checks: whether the draft sounds like the brand, raises a legal concern, or falls apart under a skeptical read.
- Owner: a set of critic agents: a brand editor, a lawyer and a skeptical reader.
- Fail action: the draft goes back for revision before the next step.
Voice rules work best as concrete, testable statements rather than adjectives. "Warm and professional" gives a critic nothing to check. "Never use the word cheap, always name the product line, address the reader as you" does. When an operator corrects the same thing twice, that correction is a candidate rule. In our setup, an operator edit can become a brand rule after confirmation, with the before, the after and the reason shown, so the fix sticks instead of being retyped into the next prompt.
3. AI-artifact check on model output and images
- What it checks: every model response and every generated image.
- Owner: an external artifact-check service.
- Fail action: fail-closed in production. If the check fails, nothing is saved.
The important word is before. The check runs before persistence, so a defective image never lands in the library where someone might pick it up later. For image selection, a person still makes the final call: candidates from several image models are shown side by side, the operator picks the frame that moves on, and the pick is stored with its reason.
4. Video artifact check
- What it checks: each generated video clip.
- How: a frame strip, a frame-difference pass and a multimodal scan.
- Fail action: reroll, at most 2 times.
Video fails in ways a single still frame hides. A frame strip shows the whole clip at a glance. A frame-difference pass highlights where something changed that should not have, such as a face, a hand or a product shape. A multimodal model then reviews the result and returns a verdict.
Two rules make this gate trustworthy. The first is that a scan that fails or cannot be read returns unknown, and unknown never counts as a pass. Pipelines that default to "no error, so fine" are how broken clips get through. The second is the reroll cap.
5. A cap on rerolls
Rerolling is tempting because it is cheap per attempt. Without a cap, a pipeline can regenerate a clip indefinitely until a check passes by luck, which burns money and selects for clips that fool the checker rather than clips that are good. A cap of 2 rerolls forces a decision: if the clip still fails, a person looks at it, changes the brief or drops the shot.
I will be direct about status in our own system: short vertical video runs within pilots, not as a generally available format. Whether generated video meets a specific brand's bar is exactly what a pilot measures.
6. Human sign-off
- What it checks: everything the automated gates cannot judge, such as whether the piece is worth publishing at all.
- Owner: the reviewing person.
- Fail action: the piece stays a draft.
Automated checks narrow the field. They do not decide. In our AI content factory, every output is a draft and the system has no publish step. A person reviews each piece, edits it if needed and publishes it from the team's own tools. Making approval structural, rather than a step people are asked to remember, is what keeps it from being skipped under deadline pressure.
7. Audit trail
- What it records: for every artifact, the brief, inputs, prompts, models and QA result. For prompt changes, who changed it, when, the old value, the new value and the reason.
- Why it matters: when a published piece turns out to be wrong, you can see which step let it through and fix that step, instead of blaming "the AI".
The audit trail is also what makes the other checks improvable. Without it, you know a bad post went out. With it, you know whether the fact-check missed it, the factbook was out of date, or a person approved it anyway.
Common mistakes when adding QA to AI content generation
A few patterns I see repeatedly when teams bolt checks onto an existing workflow:
- Checks that only warn. A gate that logs a problem and lets the piece through trains everyone to ignore the log.
- One model checking itself. Asking the generator to review its own output in the same pass gives you agreement, not verification.
- Unknown treated as pass. A timeout, an empty response or an unreadable result must stop the piece, not wave it on.
- Shared facts across brands. A fact-check is only as good as the scope of the factbook it checks against.
- No limit on retries. Unlimited rerolls hide a bad brief behind a lucky generation.
Putting the checks into a pipeline
You can run these checks by hand for a while. Past a handful of posts a week, or past a single brand, they need to be part of the pipeline itself, with each brand in its own workspace and each gate wired to a fail action. That is what our AI content factory for multi-brand teams does, and the way to find out whether it fits your brands is a pilot on two of them, judged against criteria agreed in advance.
For the broader picture of where AI content creation helps and where it hurts in a multi-brand setting, see AI content creation for brands.
FAQ
What is the most important check for AI content generation?
For text, the fact check against a per-brand factbook, run as a separate step from writing. For images and video, the artifact check before anything is saved.
Can AI detectors replace these checks?
No. Detectors guess whether text was written by a model. The checks here ask different questions: is this true, does it sound like the brand, is the image free of defects, and has a person approved it.
How many times should a generated clip be rerolled?
Set a small fixed cap. We use at most 2 rerolls, after which a person decides whether to change the brief or drop the shot.
Does quality control slow AI content generation down?
It adds steps, and most of them are automated. The part that takes real time is the human review, and that is the part you should not remove.