Get Real

On 16 October, at the launch of the Journées Particulières, Antoine Arnault placed AI on the side of the supply chain and accounting, "never for creativity" (translated from French), according to Radio Classique. He presented the return of the event, open to the public in 71 venues, as a response to "the age of artificial intelligence", according to WWD.

Between 17 and 20 October, however, three internship offers from Louis Vuitton's design studios mention AI. The one from the women's shoe studio asks the intern to "generate visuals and renderings from drawings" (translated from French) with "Procreate, in-house AI programmes" (Louis Vuitton). The Objets Nomades textile offer speaks of renderings "using 3D or AI" (Louis Vuitton); the men's leather goods offer specifies that "3D and AI are a plus" (Louis Vuitton). Of 99 design, studio or image offers published by the LVMH group, across all its Maisons, over one month, collected on 21 October from its careers site, I counted 11 that mention AI.

I see no hidden contradiction here. A rendering serves to judge a form before making it; a campaign serves to show it to the client. But a leader's words become a public promise, and a public promise gets checked. My tally shows AI in the studios; the group's public pages that I read, the careers site included, do not say at which stages of creation it is allowed. The day a client, a journalist or a regulator asks the question, the answer will need to exist in writing.

Eric Bompard puts its generated images on hold. On 14 October, at a morning event held by the Fédération française du prêt-à-porter féminin, Alexis Rouge Carrassat, the brand's head of data, explained that generated-image projects had been tested a few months ago, then put on hold: a rendering of cashmere that is not 100% reliable is not acceptable, according to FashionNetwork on 20 October (French version, for subscribers).

At the end of August, however, the agency Cherry, which has worked for the brand, presented generated visuals for Bompard, "material close-ups" of the cashmere, designed "without ever letting the process show" (translated from French) (Agence Cherry); the page does not say whether they were published.

My reading: this is the first brand I have seen give a public, technical reason for saying no to the generated image, and the reason is the material. The more the material weighs in an object's price, the less acceptable an approximate image becomes. A head of image can draw a simple step from this: write down the materials in the collection that AI renders badly, and the fidelity threshold required before any trial.

Meta seeks backing from fashion. The New York Times revealed on 16 October that Meta had invited 19 fashion journalists, editors, designers and influencers to its Meta Connect conference to obtain, according to a Meta spokesperson quoted by the paper, a "cultural endorsement" from fashion's most influential players (The New York Times). Vogue World Milan, in September, was already partly sponsored by Meta Glasses, according to TechCrunch, and Meta has been courting fashion since June for its glasses, one agency seeing in this the need for "cultural cachet", according to Glossy.

What I take from this: Meta does not go through the Maisons, it goes through those the Maisons invite to their shows, and the platform seeking their backing is the one that decides, further down, what it labels or not. A communications department can set now what its ambassadors, and the creators it pays, may accept from tech groups.

1. California joins New York on synthetic performers. Bill SB 1050, signed on 16 September and in force from 1 January 2027, requires clear disclosure of any audio or video advertisement that "prominently" features a synthetic performer, according to the law firm Manatt; New York has applied a similar rule since 9 June, which I mentioned in Edition 11. The wording is prescribed ("This performance features a synthetic performer") and the fine can reach 2,500 dollars per violation, according to the law firm BakerHostetler.

Pennsylvania has passed a broader text (SB 806, passed on 7 October and sent to the governor, who has not yet signed it according to the sources I read, including MediaPost): any advertisement that uses AI in a way that misleads about the product would have to say so, according to the MultiState summary of 16 October. Why it matters: a Maison campaign shown in New York, already, or in California in 2027, that features a generated model will have to say so.

2. Meta removes the word "train" from its advertiser terms, not the collection. In mid-October, Nieman Lab and then The Verge spotted in Meta's business terms, applicable from 30 October, a clause allowing it to train its models on any advertiser's site; Meta removed it in the following days, according to TechCrunch. Three lawyers interviewed by Nieman Lab consider that the new version still allows broad collection: for the lawyer Matt Topic, "develop" is enough to cover "train". What it changes: a Maison that buys advertising on Instagram has accepted terms which, according to these lawyers, let Meta collect its site; to be reviewed by the legal team and the IT department before 30 October.

3. A fake photo of an actress circulated unlabelled. Meta's rules do not require an AI label on still images: a fake photo of the actress Mia Goth reached nearly 500,000 likes on a verified Instagram account, according to The Verge on 15 October, and several media outlets relayed it as genuine, according to TheWrap. Meta deleted the account, which had more than a million followers, for impersonation, not for the generated image, according to IBTimes. For a Maison: its ambassadors are exposed in the same way and, since Instagram does not label still images, it is up to the Maison to spot and deny.

The six campaigns reviewed by BoF on 20 October came out between May and August. Prada entrusted "Simple Stories" to the Oscar-winning director Barry Jenkins (Hypebeast); Burberry films the daily commute with the model Rosemary Ferguson, over the voice of a London Underground announcer (V Magazine); Dior sets its Bar jacket in a Parisian café (Jing Daily); Balenciaga follows the actress Sarah Pidgeon through New York, in three films by Celine Song that show the camera and the crew (Hypebae); Saint Laurent films the actor Connor Storrie in an almost empty flat (Nylon); Chloé photographs its models on hills in bloom (Interview).

"The return to this very human imagery follows a year of AI experimentation", the publication writes.

On 14 October, at the same Fédération française du prêt-à-porter féminin morning where Bompard said no, Etam's supplier, the studio Envelope, put a figure on the other option, according to FashionNetwork: a fully generated campaign, produced in two weeks, for 15,000 to 25,000 euros against 50,000 to 75,000 for a conventional campaign, according to the ranges given by Rémi Sello, the studio's co-founder. Etam is not a Luxury Maison and these are the supplier's figures; they show which way the slope runs.

What I argue for. The Maisons' choice is coherent: when a generated image costs a fraction of a shoot, one shot with Barry Jenkins proves, by its very cost, that someone wanted it. But this choice only has value if the client can tell the difference, and the law is beginning to require that it be flagged.

What is mandatory, what is optional, what remains open. Mandatory: since 2 August, Article 50 of the EU AI Act requires any deepfake to be flagged, that is, an image, sound or video generated or manipulated by AI that resembles existing people, objects or places and would appear authentic; the fine can reach 15 million euros or 3% of worldwide turnover (Article 99). A generated model on an e-commerce page falls under it; a fictional campaign more probably falls under the lighter disclosure below, and the first risk is then reputational.

Optional: the European icon, below. Open: for works that are "evidently artistic, creative, satirical, fictional or analogous", the regulation allows a discreet disclosure that does not hamper the work, according to the European Audiovisual Observatory; whether a Maison campaign can claim this remains to be settled. I set out its timetable as early as Edition 10.

Europe already has its icon. On 10 June, with its code of practice on the transparency of generated content, the Commission published a set of icons to affix to an image produced or manipulated by AI, a general icon and others that separate the fully generated from the retouched, according to the law firm Heuking; no one is obliged to use them, and affixing one is not enough to comply, according to CADE.

As the icon spreads, a campaign that does not carry one will say, without the Maison having written it, that nothing here is generated. And if one of its e-commerce visuals does carry one, the Maison will have to be able to explain why that one, and not the others. My position: on generated e-commerce visuals, the icon, as soon as clients recognise it; on the campaign image, nothing to affix, because it is shot.

The objection. Disclosing costs desire. Four researchers measured it in the Journal of Advertising Research in 2025, across three experiments: when a Luxury brand flags that an ad is generated by AI, consumers judge it less favourably, because they see less effort in it, and therefore less brand authenticity; the effect fades when the generated image is highly creative. The study does not measure the reaction to an image flagged as "partly generated".

The rule I draw from it is the one in my forthcoming book, Luxe Agentique: the machine backstage, the hand on stage. Always trace, internally; show the disclosure where the law requires it and where the client asks for it; never turn "made by AI" into a selling point. The client judges the image; they judge the process only if it is put in front of them, and then they judge it less favourably.

What the measure does not say. The BoF article reads campaigns; it measures nothing: no figure on the share of generated images in the Maisons' productions. The study tests fully generated ads, not adaptations. And the public pages of the fashion and leather goods Maisons that I read publish no rule on the share of AI in their images.

This rule translates into four steps. The label becomes a mandatory field in the image library, signed off by the head of image, checked by the legal team, filled in by the agency and its subcontractors under contract, with a right of audit. The trace is written into the file according to the open C2PA standard, a process that attaches the history of its making to the image, which I described in Edition 12. It is backed by an internal register, because that inscription can disappear when the image is republished. And the answer in store, "is it generated?", is prepared with the sales teams and entrusted to the client advisors.

The question leaders ask me is no longer "should we go?" but "how far?": where, in our company, does the line run between what AI may touch and what it must never touch? That line cannot be decreed. It is drawn use case by use case, with the managers who own them, and it is written down, so that it can be stated the day a client asks for it. That is the work I do with you.

Three types of engagement, à la carte. You start with the one that answers your question of the moment; none commits you to the others.

1 · AI strategy. Identify and specify your use cases. A use case is a specific task that a team hands, in whole or in part, to AI. I help you identify them, then the managers who own them rule on them, based on what holds you back as much as on what makes your strength. Read →

2 · AI solutions. Choose and set up your AI solutions. For each need, the right solution, preferably among those already in place in your company, configured with your team and left up and running. Read →

3 · AI training. Train your teams. Your teams take AI in hand on their real assignments; each person leaves with what they can redo on their own. Entry point: the team leader, who leaves with their adoption plan. Read →

Book a meeting →

On 14 October, YouTube Music began displaying a label on tracks "created primarily using generative AI", directly on the listening page, according to its official blog.

Rights holders choose one of three disclosures, "Fully Gen AI", "Partly Gen AI" or "No Gen AI", and a track retouched with a pitch corrector stays in the third, according to TechCrunch. The declaration goes through the tracks' delivery files, and therefore through the distribution chain itself, and YouTube cross-checks it with its own detection tools; it does not yet say whether the "partly" disclosure will be displayed on screen, according to Music Business Worldwide.

The movement is collective: Spotify has displayed AI credits since April, Deezer announced a filter against fully generated music the same week, and Apple Music has promised "Made With AI" labels by the end of the year, after a call from the recording industry, the Grammys and the performers' union SAG-AFTRA, according to MobileSyrup and Music Business Worldwide.

On 20 October, Spotify published its review under a title that could belong to a Maison: "One Year In: Putting Human Artistry First on Spotify" (Spotify press release). Artists file "AI Credits" there, detailing track by track the share handed to the machine, and more than 300,000 artists, who according to Spotify account for more than 99% of what listeners search for there, carry the "Verified by Spotify" badge, which guarantees "a real, identifiable person" behind the music, according to the same press release.

Two things strike me. The declaration is graduated: between all and nothing, there is "partly", and that, I believe, is where most real production will sit. And it is the rights holder who declares, the platform that checks: YouTube with its own tools, Spotify with those of other companies. Spotify admits, moreover, that no single signal is decisive and that AI detectors remain fallible.

Zalando already does what the Maisons avoid. On 18 September, The Business of Fashion reported that 93% of the marketing campaigns on its site are made with AI, according to Matthias Haase, its vice president in charge of content solutions.

For product pages, agents cast from among 150 fully generated models and assemble the outfits following the brand's rules; a second team of agents reviews each image and repairs what the AI has added of its own accord. A campaign used to take six to eight weeks; it now takes four days, according to the same article. Zalando confirms it: its models "can sometimes infer details that are not actually there", hence a check that compares each generated image with the product photo (Zalando).

The right ground, for a Maison: adaptations of a campaign already shot, to the formats of social networks and markets, and e-commerce visuals, at the material fidelity threshold Bompard sets itself, never the campaign image itself.

The trap is disclosure. According to the same BoF article, Zalando considers that most of its images fall outside the European obligation on deepfakes, and that a label on every visual would sow confusion. Zalando writes it itself: "Visible labels are reserved specifically for scenarios where content could otherwise mislead consumers" (Zalando). Its rules for sellers forbid any visible disclosure on the image and instead require an invisible marking in the file, mandatory by December (Zalando's image guidelines).

The result, BoF notes: on the site, an image marked as generated is all but impossible to find. What I argue for goes further. Zalando reserves visible disclosure for images that could mislead, forbids it to its sellers and asks them for a trace in the file; a Maison says "shot" of its human images, knows for each of the others what is generated, and gives that label to whoever asks for it, client, journalist or regulator. That reading holds up for a retailer; it will hold up poorly for a Maison that sells the material of its objects.

In the coming seasons, I expect a Maison to adapt a shot campaign with AI. The test will not be speed, it will be material: a cashmere, a leather, a grain rendered accurately enough for the adaptation to stand next to the photograph. Its in-house studios will gain adaptations to review rather than to produce: a craft that changes.

A Maison has always known how to say where a skin comes from, who assembled a piece, in how many hours. It will soon have to say, too, what AI did in an image. The one that writes it down first will, I believe, set the rule for its whole sector. Luxe oblige!