Direct answer: In Gavin Sim’s practical experience using AI-generated images and video, the output often does not look good enough to publish. A person has to pay close attention, exercise judgement and decide whether to publish, edit or regenerate. Judgement is a permanent stage in the workflow, not a temporary gap waiting for better models.
The firsthand observation
This comes from my practical experience using AI-generated images and video. The output often does not look good enough. Someone has to pay close attention to it, exercise judgement and make a call before anything is published. That is the whole of the firsthand claim, and everything after it is a working method rather than a report of incidents.
I want to be careful not to overstate the case. This is not an argument that all AI visuals are bad. Much of what these tools produce is genuinely useful, and the speed of iteration is a real advantage. The point is narrower: often enough to matter, the result is not publishable as it is, so review has to be designed into the process rather than hoped for.
Seven dimensions worth reviewing
What follows is a general review checklist. These are categories a reviewer should have in mind. They are not incidents that occurred, and none of them is being reported as something that happened in a particular project.
1. Anatomy and objects
Hands, fingers, teeth, eye alignment, limb count, and the way objects intersect. Reflections, shadows and overlapping items are where models still improvise physics that does not hold up.
2. Text inside the image
Any rendered word is a risk: signage, packaging, screen content, logos. Read every character. Near-correct text is worse than no text because it looks deliberate.
3. Continuity in video
Across shots, check that clothing, lighting, hair, props and backgrounds stay consistent. Faces and objects can drift subtly between frames in ways that feel unsettling before viewers can articulate why.
4. Timing and motion
Pacing, cuts, and whether movement obeys weight and momentum. Video that is a beat too slow or motion that floats reads as artificial even when every frame is individually fine.
5. Brand consistency
Colour, typography, framing, and tone. Generated output tends to drift toward a generic house style, which is precisely the opposite of what a brand asset should do.
6. Factual and contextual fit
Does the scene match the claim beside it? A stock-feeling boardroom next to a story about a two-person team undermines the copy. In a local context, generated environments frequently do not look like the place they are supposed to represent.
7. Rights and provenance
Check for recognisable likenesses, trademarks and third-party styles. Know which tool produced the asset and under what terms, and keep that record with the file.
A practical approval checklist
- State the asset's job in one sentence before reviewing it.
- View it at final size, on the device the audience will actually use.
- Scan the seven dimensions above, in order, without rushing the boring ones.
- Read every visible character of text out loud.
- Place it beside the copy it will run with and check they agree.
- Ask whether it looks like your brand or like a generic AI asset.
- Record the tool, the prompt and the approver alongside the file.
One person approves. Approval by committee produces a slow average, and averages are how brands end up looking like everyone else.
Regenerate, edit manually, or reject
Most wasted time in creative AI work comes from choosing the wrong one of these three.
Regenerate when the flaw is structural: wrong composition, wrong concept, anatomy that cannot be patched. Change the brief, not just the seed. Repeating the same prompt and hoping is not iteration.
Edit manually when the asset is fundamentally right and one localised element is wrong: a stray object, a colour cast, a small correction. A two-minute manual fix beats a twenty-minute regeneration cycle.
Reject after roughly three regenerations that have not moved the output closer, or whenever the asset carries meaningful claims or real people. Commission it, shoot it, or use a simpler design. Knowing when to stop is a skill, and it is cheaper than the alternative.
Judgement is part of the operating system
Gavin Sim, on the recurring pattern: Businesses struggle with AI because they lack practical knowledge, expect it to do everything automatically and underestimate the learning curve. Judgement is still required after implementation.
This connects directly to how adoption should be designed. In what operational AI adoption actually looks like I argue that every workflow needs a named human approval point. Creative work makes that argument visible, because the shortfalls are easy to see. The operational version of the same idea is in how we onboard students with AI-assisted automation.
Teams that accept this early move faster overall. They stop expecting perfect first outputs, build review into the schedule, and spend their attention on the decisions that actually differentiate the work.
Frequently asked questions
Are AI-generated images and video unusable?
No. Plenty of output is useful. The point is narrower: the output often does not look good enough to publish as it is, so it needs close human review. Review is the cost of using the tool, not evidence that the tool is bad.
Which review categories catch the most problems?
As general review categories rather than reported incidents: anatomy and objects, any text rendered inside the image, continuity across video shots, timing and motion, brand consistency, factual fit with the surrounding copy, and rights or provenance.
How long should review take?
For a single image, under two minutes with a checklist. If review regularly takes longer than editing the asset by hand would have, the brief is wrong or the tool is the wrong choice for that asset.
When should you regenerate rather than edit?
Regenerate when the problem is structural, such as composition, anatomy or the wrong concept entirely. Edit manually when the asset is right and one localised element is wrong. Reject when three regenerations have not moved the output closer.
Does disclosure matter?
It depends on context and audience expectations, and in some cases on platform or client rules. A sensible working position is straightforward: never present generative output as documentary evidence of something that happened.
Does this mean AI has failed?
No. It means judgement is part of the operating system. Every production process, including entirely human ones, has a review stage. AI changes who does the first draft, not whether anyone is accountable for what ships.
Sources and methodology
Firsthand evidence. One claim is firsthand: in Gavin Sim's practical experience using AI-generated images and video, the output often does not look good enough, so a person must pay close attention and exercise judgement. No specific project, client, channel or failure incident is being described.
Editorial guidance, clearly separated. The seven review categories, the approval checklist and the regenerate, edit or reject rules are a recommended working method for readers. They are not a log of problems encountered.
External statistics. None are used in this article. No model benchmarks, error rates or vendor comparisons are claimed, and no specific tool is named or evaluated.
Related reading and contact
For the operating context behind this view, see the OFT AI Framework. Teams building an internal review standard can see how that is taught in corporate AI training. Journalists looking for commentary on generative media and human oversight can find bios, topics and contact details in the media room.