Who Approved This Version?

A person accountable to the brand should approve every material AI-generated change to a product, claim, image or customer answer. When a platform alters those elements automatically, the advertiser still carries the reputational and legal consequences, even if the platform controlled the generation, settings and distribution.

A dress became a shirt and trousers. A bicycle gained a second handlebar. Carefully photographed products came back looking like imitations of themselves. These are funny examples until the altered advertisement carries your name, spends your budget and reaches a customer who assumes you approved it.

The important shift this week is not that machines can produce more creative. Marketers already know that. The shift is that automation is moving beyond drafting and into decisions that used to belong to the advertiser: what the product looks like, what the copy claims, how a lead is answered and whether a public image can be reused.

Approval moved downstream

Business Insider spoke to advertisers and agency executives who said Meta's creative tools had distorted products, inserted unwanted people, garbled text and enabled features unexpectedly. One pyjama dress was reconstructed as separate clothing items. An advertisement for a women's networking group gained a man, while an REI bicycle appeared with two sets of handlebars.¹

Meta's position is that advertisers must review the output because generative systems can make mistakes. That sounds reasonable when a marketer actively requests a new image, inspects it and chooses whether to publish it. It becomes harder to defend when settings are difficult to notice, are reportedly toggled on by bugs or modify an asset after a team believes the creative has been approved.¹

This is not a small production error. A product image is part of the claim being made to the customer. Changing its construction, colour, wording or use can turn a creative variation into a modified product claim, even when the platform regards the change as an optimisation.

The industry has spent years describing automation as a way to remove repetitive work. Yet the Meta examples add another review layer because teams now need to check not only their original advert, but every machine-produced version and every setting capable of producing one. A system sold as labour saving can quietly create an approval queue that did not exist before.

Large advertisers can assign specialists to that queue. They can build asset libraries, monitor settings, compare live placements and escalate problems through account teams. A small retailer, salon or restaurant may have one person creating the post, running the advert, serving customers and checking whether the platform turned the bicycle into something the business does not sell.

This is where AI in marketing stops being a software question and becomes an operating question. Who owns the final check? Which product features may never be altered? What happens when a generated variation performs well but depicts the wrong item? Without answers, the platform is effectively defining the brand's approval boundary through its defaults.

The platform writes back

Google's Business Agent for Leads shows the same shift from a different direction. The beta product places a Gemini-powered agent inside a Search advertisement, trained on the advertiser's website, so a prospective customer can ask questions before reaching a lead form. Google says the agent can answer questions, qualify interest and warm up leads before a sales conversation begins.²

That can be genuinely useful. A customer asking about course dates, delivery areas, dietary options or appointment availability may get an answer immediately instead of waiting for a callback. The business may receive fewer low-intent enquiries, while staff spend more time on cases that need judgement.

The trade is that the advertisement no longer ends with a claim and a click. It becomes a conversation conducted inside somebody else's interface, by a system trained on information the business supplied, while Google controls the surrounding experience. The platform is no longer distributing the message. It is participating in it.

That changes the meaning of approval. A creative team can approve a headline, image and landing page because the possible customer experience is bounded. A conversational advert can produce many answers to many questions, which means the real campaign includes the source material, retrieval rules, refusal behaviour, escalation path and every gap in the website used to train it.

If the website says one thing and the sales team says another, the agent may expose the disagreement at scale. If an old policy remains online, the agent may repeat it with confidence. If the brand voice is warm in advertisements but defensive in customer-service pages, the contradiction can surface before the customer reaches a person.

This is why a company's website and public information are becoming more than publishing surfaces. They are increasingly the source of truth for agents that speak on the company's behalf. The quality of the answer depends less on a clever prompt than on whether the underlying product facts, policies and language are current, specific and internally consistent.

For marketers, that makes content operations part of customer operations. Instagram content strategy, paid media, product data, website copy and sales enablement can no longer be treated as separate cupboards. When an agent draws from all of them, their disagreements become one customer-facing answer.

Provenance cannot be outsourced

Platforms are responding to synthetic media with labels, detectors and invisible watermarks. These mechanisms are useful, but the Reuters test of Meta's Muse Image detector showed why marketers cannot hand the whole responsibility to technical provenance systems. The detector verified all forty original images in the test, then failed to verify 55% after ordinary cropping.³

Cropping is not an exotic attempt to defeat a detector. It is routine work for social formats, display placements, websites and messaging apps. An image may be resized, compressed, screenshotted and reposted several times before the customer sees it, which means a signal that works on the original file may not survive the actual distribution process.

The practical lesson is not that watermarks are pointless. It is that detection belongs at the end of a longer chain, not at the beginning. The advertiser should retain the source asset, record material edits, know which system generated or altered the file and disclose synthetic production when it changes what a reasonable viewer would believe.

That record matters even when the final advert looks good. A polished output can still be unauthorised, inaccurate or impossible to reproduce. Without provenance, a team cannot explain why a face changed, which product image was used, whether a customer consented or who approved the version that reached the feed.

This is especially relevant to AI content marketing because generation is becoming easier at the same time as accountability becomes more distributed. The model may create the image. The platform may crop it. An automated campaign may choose the audience. A reseller may repost it, but the customer's complaint will still arrive at the brand named in the advert.

The temptation is to rely on platform labels as evidence that the system is under control. Yet a label can tell the audience that some machine involvement occurred without telling the advertiser whether the result is correct. Transparency after publication does not repair a false product detail, an invented endorsement or an unapproved likeness.

A sensible provenance standard therefore begins with the brand's own files and decisions. It should answer four plain questions: what was supplied, what changed, who approved it and what the audience needs to know. That is less glamorous than another generation demo, but it is the infrastructure that lets creative teams use faster tools without losing the ability to account for their work.

Human texture gets pricier

The creative work attracting attention this week offers a useful counterpoint. Twix used a deliberately strange ventriloquist double act, directed by Dougal Wilson, rather than presenting another frictionless stream of machine-made variations.⁴ Hinge is leaning into relatable human stories and honesty at a moment when users are increasingly sceptical about both dating apps and synthetic interaction.⁵ Both campaigns make a visible choice instead of hiding the work behind infinite variation.

Neither example proves that audiences reject machine-assisted production. They show something more interesting: as competent synthetic content becomes abundant, evidence of choice, craft and human intention becomes easier to notice. Human texture starts to function as a differentiator because smoothness is no longer scarce.

This creates a better role for AI than replacing the whole creative act. It can help a team explore concepts, organise footage, resize approved assets, prepare calendars, compare edits and learn from performance. The team can spend less time on mechanical repetition while protecting the moments where casting, humour, product truth and cultural judgement determine whether the work means anything.

The OpenAI advertising forecast debate reinforces the point from the commercial side. OpenAI has projected $100 billion in advertising revenue by 2030, while Emarketer estimates that the entire US standalone chatbot advertising market may reach only $5.41 billion by then.⁶ The gap suggests that attention inside a useful conversation does not automatically become advertising inventory on the scale familiar from search and social. It also warns against confusing access to an audience with permission to interrupt it.

A chatbot answer may satisfy the question and end the session. A lead agent may resolve an objection without a website visit. An automated social campaign may generate more variations without creating stronger memory. In each case, the platform can increase output or reduce friction while leaving the marketer with the harder task of earning trust and preference.

That is why the next phase will not be won by the team with the most generation. It will be won by the team that knows which decisions should be accelerated and which should remain deliberately slow. Product truth, consent, humour, taste and the decision to publish belong in the second category, even when everything around them becomes faster.

The last approval is the product

Marketing teams do not need to reject automated creative, conversational advertising or machine-assisted distribution. They need to stop treating approval as a single click performed before the platform begins its work. Approval now has to cover the source material, permitted changes, generated versions, live placements and answers delivered on the brand's behalf.

That sounds like more process, and in the early stages it is. The alternative is hidden process owned by the platform, where defaults decide what changes and the brand discovers the result through a customer screenshot. The choice is not between control and speed. It is between visible control and invisible rework.

The strongest teams will define a small number of non-negotiables. Product construction cannot be invented. Prices and policies must come from current records. People and public media require permission. Sensitive claims need named human approval, while routine formatting and scheduling can move quickly around those rules.

This is where brand voice becomes operational rather than decorative. It is not a paragraph in a style guide or a prompt pasted into a generator. It is the accumulated set of choices that tells a system what it may change, what it must preserve and when it has to stop and ask a person.

The platform may write the variation, answer the lead or choose the placement. The brand still owns what the customer believes. Keeping that final accountable decision close to the people who understand the product is not resistance to progress. It is the condition that makes faster production worth having.

Sources