Why More AI Creatives Can Slow Learning: The Search Problem in Short-Form Marketing

ONCE Social Short-Form Video Marketing Research · Social Short-Form Video Marketing Research 09/10

Research article / literature comparison and creative decision analysis. Research cutoff: October 5, 2026.

If a team can generate 100 short-form videos in a day, does it learn faster—or merely produce 100 results it cannot explain?

This is the problem most easily missed when generative AI enters advertising production. Creative used to be scarce and teams often could not produce enough variants to test. Candidate supply has now expanded, moving the bottleneck from making an asset to deciding which assets deserve attention, why one worked and what should change next. If the opening, character, setting, product angle, caption, sound and offer all change together, even a winning version reveals little about what should be repeated.

Our central argument is that generative AI expands the supply of candidate assets; it does not automatically improve the quality of learning. Scale becomes cumulative knowledge only when a team fixes brand invariants, preserves interpretable axes of variation, and separates generation, screening, deployment and causal validation. We call a collection that can support the next decision a “learnable creative set.” This is ONCE’s organisation and extension of the evidence, not a validated universal theory.

AI does not eliminate creative cost. It moves cost into search.

Remi Daviet and Yohei Nishimura’s 2026 Marketing Science paper separates the problem into two tasks: finding potentially effective visuals in a vast generative space and ensuring that the visuals remain acceptable for the brand and campaign. The study builds two predictive models—one for click performance and one for brand acceptability—and uses Bayesian active learning to choose the next candidates expected to provide the most useful information.

Its field component involved image ads for an outdoor-activities company on Instagram. Audience, budget, placement and template were fixed while only the background image changed. A brand representative first labelled 60 batches of 12 images to train acceptability. Performance learning then ran for nine trials with 12 candidates per trial. The final comparison deployed 10 model-selected visuals, 10 chosen by an aesthetics model and 10 supplied by one professional designer. The paper reports mean link CTRs of 0.98%, 0.78% and 0.65%, respectively. The authors also caution that the batches were small and there was only one human benchmark.

The important conclusion is not that AI defeated a designer. Two narrower observations matter. First, searching for beautiful images and searching for images that both earn clicks and pass brand review are not the same objective. Second, active learning does not mean deploying every possible variation. It uses uncertainty to select the next batch that is most valuable for learning.

The transfer limits are explicit. The experiment optimised background images and link clicks for one outdoor campaign. It did not study short videos containing performance, product actions, captions, sound and shot continuity, and it did not measure sales or long-term brand memory. The paper itself identifies complex products and people, exact text and transfer across campaigns as future problems. Turning this result into “AI video beats human production” would expand a controlled image-search experiment into a claim the study did not test.

Prediction can improve without explaining a commercial result

Mark Heitmann and colleagues take another route in their Journal of Marketing research. They fine-tune an image model using advertising mindset measures such as attention, interest, desire and intended action, then compare generated visuals with existing ads. The programme includes studies in automotive and fast-moving consumer goods and a Meta field click test. Across the two field tests, the researchers recorded 28,775 impressions and 496 clicks; fine-tuned AI images produced a 2.04% CTR versus 1.37% for existing brand images.

This is still not a sales experiment. Much of the research uses participant ratings on AIDA-related scales; the field outcome is a click. The paper also finds that trained visuals are easier to process, yet visual averaging can remove the surprise a highly differentiated product needs. A system may learn a fluent visual language associated with high historical ratings without building new brand assets.

Anuj Kapoor and Madhav Kumar’s 2025 randomised field experiment studies a different environment. An Indian direct-to-consumer brand selling sustainable products sent WhatsApp ads to more than 21,000 existing users. The three arms received purchase-history-based GenAI personalised video, personalised image, or a generic non-personalised video. The formal abstract reports that personalised AI video increased engagement by six to nine percentage points over the two baselines.

The random allocation gives the comparison causal force in that setting, but the setting is not equivalent to today’s public recommendation feeds on Douyin, TikTok, Reels or Shorts. WhatsApp is a targeted messaging environment; personalisation used existing purchase histories, while the marketing team retained control of the message. The outcome concerns interaction after a single contact, not competition inside a public feed, and it does not establish purchases across industries. Personalisation, video format and technological novelty may all contribute, while persistence under repeated exposure needs new evidence.

Together, these studies do not produce one proposition that “AI works.” They answer three different questions: whether a model can search for more suitable imagery, whether a trained objective transfers to clicks, and whether personalised video to known users improves one engagement response. Production efficiency, predictive accuracy, clicks, sales and long-term brand effects need separate records. The most convenient metric cannot answer for all the others.

What makes a creative set learnable?

The value of a creative set is not its size. It is whether the next decision can become more accurate because the set existed. We recommend specifying four constraints before generation.

Brand invariants do not force every execution into one template. They identify the facts audiences must attribute correctly and the things a brand will not trade for clicks. Exploration happens outside those invariants. The same product reason can be expressed through demonstration, user experience, comparison or a short narrative. One distinctive asset can perform different information roles in different openings.

Creative diversity also cannot be counted only through visual difference. Ten backgrounds surrounding the same narrative causality may be surface variation. Changing why the product enters the story, the sequence of proof or the qualification of the speaker may test a deeper hypothesis within the same setting. AI is excellent at producing visible difference; teams still decide which differences are worth learning from.

Fixed A/B, active learning and platform optimisation are different tests

Fixed A/B offers interpretability at the price of a narrow candidate set. Active learning supports a continuing search when feedback is expensive, but it needs stable features, explicit objectives and updating. Platform optimisation is designed for delivery and often combines exploration, allocation and evaluation. Early leaders can receive more opportunities, which may help execution while making the equal-opportunity comparison harder to recover.

A practical sequence is to use small fixed comparisons to calibrate a creative mechanism, then let the platform scale candidates that pass brand and information checks. Teams with sufficient data and modelling capability can add active learning to select the next generation direction. The three methods can form a relay, but they should not share one vague label such as “AI automated testing.”

Creative fatigue is not solved by generating ten more files

At least three forms of fatigue matter. Viewers can tire of repeatedly seeing one execution. A platform can concentrate similar assets on the same kinds of people. A team can repeatedly optimise one short-term winner until brand expression narrows. Producing more surface variants may relieve the first while worsening the latter two.

A more resilient portfolio preserves two forms of diversity. Execution diversity changes setting, rhythm, people, composition and sound. Hypothesis diversity tests problem recognition, functional proof, brand linkage and reasons to act. The first reduces repetition. The second keeps the learning system from pursuing only one response.

AI disclosure should not be treated as a switch that always increases trust or always harms response. A 2026 study using three online experiments across watches, smartphones, moisturisers, print and video reports that AI disclosure reduced brand attitudes overall. The negative effect was stronger with emotional appeals and warm brands and weaker with rational appeals and competence-oriented brands. This is an interaction in specific experiments, not a reason to avoid disclosure. Legal obligations, platform rules, content risk and audience expectations still require separate decisions. Disclosure measures how people respond to a source cue; it does not prove that the production process itself is reliable.

A public case: consistency is a search condition, not final retouching

ONCE’s public AIGC off-road motorcycle concept case provides a limited but useful object for analysis. We read the case text and visible page information. The page describes a film of roughly 28 seconds and recurring cues including a black motorcycle, gold rims, white riding gear, warm sunlight, dust and sunset. We do not use it to claim media, click or sales results.

To turn that direction into a learnable creative set, the vehicle colour, rims, riding gear and outdoor-adventure tone could remain fixed while one axis changes at a time. Does the opening reveal the rider or the tyre first? Does the middle emphasise terrain change or mechanical detail? Does the ending use a sunset silhouette or a clear product close-up? The visual world can vary without redrawing product and brand facts during every generation.

This is why episode 04 on brand memory still matters. If a high-performing asset weakens brand attribution, the system may be learning detachable attention. Episode 08 on counterfactual mismatch adds the measurement warning: a platform-selected click winner is not automatically the version that produced the most incremental sales.

A minimum creative-learning experiment that can fail

The following is an experiment proposal, not an ONCE client result.

1. Write the invariants. List the appearance and functions that must remain accurate, distinctive brand cues, permissions and prohibited claims.
2. Create three hypothesis families. For example: problem opening, functional proof and brand context. Generate six to ten rough candidates in each family without deploying all of them.
3. Reject untestable outputs. Product distortion, incorrect captions, identity drift, missing brand cues and unsupported claims do not enter media testing.
4. Build matched versions. Select two candidates in each family. Keep duration, offer, landing page, audience and budget consistent while primarily changing one creative variable.
5. Read outcomes by layer. Viewing diagnoses openings and rhythm; product comprehension and brand linkage test meaning; clicks and qualified enquiries observe action; sales and incrementality require an appropriate counterfactual design.
6. Select the next batch. Do not only copy the winner. Keep one exploitative version extending a known strength and one exploratory version aimed at the largest uncertainty.

The design can fail because the sample is too small, platform delivery is unequal, the variation axis remains contaminated or outcomes occur off platform. That failure should not be renamed “the model needs more creative.” First determine whether the missing resource is generation capacity, brand review, comparable exposure or outcome data.

Conclusion: the scarce asset is an interpretable next step

Generative AI is moving creative supply from scarcity to abundance. The new advantage is not generation speed alone. It is knowing which facts to hold fixed, which hypothesis to vary, which outcome to evaluate, and when a result cannot travel from click to sale, image to video, or private message to public recommendation feed.

A good AI creative system should leave a team knowing something it did not know before, not merely holding more files. Build a learnable creative set first, then decide whether fixed A/B, active learning or platform optimisation fits the question. Otherwise, more assets can make a team repeat unexplained success faster.

If you are planning generated shots, live-action truth anchors, keyframes and multiple short cuts, review ONCE’s AIGC commercial and hybrid-production approach. It describes creative, execution and delivery scope; it does not claim that ONCE provides platform attribution infrastructure or guarantees media results.

Sources and reading scope

• Remi Daviet and Yohei Nishimura, Leveraging Generative Artificial Intelligence to Create Visual Content in Digital Advertising, Marketing Science 45(5), 2026, DOI: 10.1287/mksc.2024.1130. We read the public full text, method, field experiment, results and limitations. The study concerns Instagram background images, brand acceptability and CTR.
• Mark Heitmann, Tijmen P. J. Jansen, Martin Reisenbichler and David A. Schweidel, Picture Perfect: Engaging Customers with Visual Generative AI, Journal of Marketing 90(4), online first June 25, 2025, DOI: 10.1177/00222429251356993. We read the public full text, study designs, field clicks and boundaries. Most outcomes concern static images and mindset ratings.
• Anuj Kapoor and Madhav Kumar, Frontiers: Generative AI and Personalized Video Advertisements, Marketing Science 44(4), 2025, DOI: 10.1287/mksc.2023.0494. We read the formal abstract and the MIT research explanation. The setting is WhatsApp messaging for an Indian DTC brand; the outcome is engagement rather than purchase.
• Aligning Source, Appeal, and Brand: Resolving the Disclosure–Effect Dilemma in AIGC Advertising, Electronic Commerce Research and Applications, 2026, DOI: 10.1016/j.elerap.2026.101598. We read the formal abstract covering three online experiments, categories, formats, results and mediators; we do not present it as a cross-market law.
• ONCE’s AIGC solution, public motorcycle concept case and completed episodes 04 and 08 were rechecked on October 5, 2026. The case supports content and production analysis only, not undisclosed performance claims.