Generative tools have changed the front half of playable production and left the back half almost untouched. Exploratory work that can run in parallel and is cheap to throw away has become much faster. Work that is exact, interactive and reviewed by a third party has not moved. Teams that use AI on the first kind get a real speed-up; teams that use it on the second spend the saving on rework.
Where it helps
Concept volume. The expensive constraint in creative testing has always been how many truly different ideas a team can put in front of the auction, and the bottleneck is usually the effort of making an idea concrete enough to judge. Generating twenty rough boards for a concept review costs minutes, and nineteen being discarded is the point rather than a failure.
Background and environment plates. A background is a large asset carrying no interaction, and it is the easiest thing in a playable to generate. You need a mood and a palette, not accuracy. It is also where generated output survives compression best, because a painted background tolerates WebP at a lower quality setting than UI art does.
Copy variants. Headline and CTA microcopy is a real testing variable, and generating thirty phrasings to pick five from is a good use of a model. The judgement stays human; the typing does not need to be.
Localisation first drafts. Machine translation has been production-viable for this for a while, and the current generation handles ad microcopy better than earlier tools because it can be given the context: that this is a button label, that it must fit a fixed width, that the tone is playful. It still needs a native reviewer before it ships. What it removes is the wait for a first pass.
Cutouts and clean-up. Removing a background, extending a plate to a second aspect ratio, and upscaling a small source asset are all routine now, and each saves the kind of hour that used to block a whole variant set.
Where it reliably fails
Producing a shippable interactive file. A model can write plausible HTML and JavaScript for a mini-game. It cannot reliably produce one that inlines every asset, calls the right store-exit function for each of a dozen networks, respects a 2 MB index.html budget, handles both orientations, survives a minifier, and holds a frame rate on a four-year-old Android device. Each of those is achievable individually; all of them together, correctly, first time, is not. The failure mode is a file that works in a desktop browser and fails in review or in the container. This is where a deterministic exporter earns its place.
Consistency across a variant set. Testing depends on changing one variable at a time. Generated art drifts (proportions move, the palette shifts, the lighting changes) and a variant set whose art drifts is not a test, because the difference in result cannot be attributed to the variable you meant to change.
Brand and IP safety. A generated asset can resemble protected work, and a generated character can drift off a licensed style guide in ways obvious to the licensor and invisible to whoever generated it. For licensed IP that is disqualifying without a human review step somebody is accountable for.
The last ten per cent. Hands, text inside images, UI chrome, and any element that must align precisely with an interactive hit area still need manual correction. Text rendered inside a generated image is a particular trap in this format, because it cannot be localised, cannot be edited for a variant, and is usually the first thing you want to change.
Disclosure is now a legal question in the EU
The transparency obligations in Article 50 of the EU AI Act apply from 2 August 2026. They require that outputs of generative AI systems be machine-identifiable as artificially generated or manipulated, and that people are informed where content is a deepfake. The European Commission published finalised guidelines on those obligations on 20 July 2026, alongside a Code of Practice on the marking and labelling of AI-generated content developed through the AI Office. Penalties reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher.
For most playable production this is a documentation and process obligation rather than a creative constraint: a generated background in a game ad is not a deepfake. It becomes sharp where a creative uses a synthetic likeness, a synthetic voice, or generated content presented as a record of something real, such as a fabricated review or testimonial. Those were already prohibited by platform policy on misleading creative; they now also sit inside a regulatory regime with its own penalties.
What to put in place now
Record which assets in a creative were generated and with what tool, keep the provenance metadata your tools emit rather than stripping it in an export step, and route anything involving a person's likeness or voice through whoever handles legal review. None of that slows production down. Reconstructing it a year later, after a query, does.
The rule that works in practice
Put AI upstream of the editor and never downstream of review. Generate concepts, plates, copy and translations; assemble, tune and export deterministically; and review the finished creative the way you would review any other. This is not scepticism about model quality. The back half of playable production is governed by exact third-party requirements, and what you want from a tool there is that it does the same thing every time.
What to add to QA
- Check generated art at the size it actually renders. Artefacts invisible at full resolution become obvious at a fifth of it, and the reverse is also true.
- Look for text inside generated images and remove it. It cannot be localised or edited, and it is frequently subtly wrong.
- Compare every asset in a variant set side by side for drift in palette, proportion and lighting before treating the set as a test.
- Confirm anything resembling a person, a voice or a real record has been through whatever review your organisation requires.
- Re-check file size after generated assets are added: upscaled and painted output is larger than hand-authored art at the same apparent quality.
Create playable end cards in minutes—no code required.
Open Playable Ads Maker