When AI content gets a label, does it lose its value?

Nobody likes being marked.

When Anthropic announced that new Claude models would embed an invisible, machine-readable watermark into generated text, parts of the internet reacted as though the chatbot had started signing its users' homework. Reddit threads filled with arguments over authorship, privacy and whether merely polishing human-written work with Claude should leave an AI fingerprint. Dozens of users reportedly said they were cancelling subscriptions, even as Anthropic said it had not seen an uptick in cancellations.

The irony arrived quickly. Days later, Google announced that users could turn off the visible watermark on AI-generated images, videos and music created through Gemini and Flow. But only the visible one. Google's invisible SynthID watermark and C2PA provenance metadata remain embedded underneath. In India, moreover, Google's help documentation says the visible-watermark toggle is available only to AI Ultra subscribers.

The watermark, in other words, isn't disappearing. It's disappearing from sight.

And for brands and agencies, that distinction raises a more consequential question than whether consumers can see a little AI badge: what happens when machines can increasingly identify machine-made content, even when the humans consuming it cannot?

Anthropic's system is based on Google DeepMind's SynthID-Text technology, which subtly adjusts the probabilities governing word choices during generation to create a statistical pattern detectable with the appropriate system. Google already uses SynthID to watermark text generated by Gemini.

For advertisers, the first consequence may emerge not in the creative department but in distribution.

Read more on AI attribution and distribution 

In India, AI is already becoming embedded in marketing operations. Salesforce’s State of India Marketing 2026 shows 81% of marketers have adopted AI, while 83% say they need more personalised content than they can currently produce. Deloitte separately found 55% of Indian enterprises have deployed AI at scale across marketing and sales.

The productivity gains can be dramatic: Reuters reported in May that Kimberly-Clark had cut some content-creation cycles from 24 days to just two hours using an AI platform developed in India.

Kartik Mehta, Chief Business Officer and Head of Asia at Channel Factory, believes platforms could increasingly distinguish between AI used as a creative tool and AI used as a content factory.

“Not all AI is bad. A human creator using AI for high-quality visual effects will be treated differently than a faceless channel churning out automated script-to-video content,” he says, arguing that platforms may lower the organic reach of unoriginal, scaled AI content to protect user experience.

For advertisers, that distinction could extend the familiar concept of brand safety into what Mehta describes as “AI suitability”: brands deciding how much AI assistance they are comfortable supporting in the content around which their advertising appears.

But provenance becomes more awkward when the AI-generated material isn't surrounding the advertisement. It is the advertisement.

Saket Dandotia, Founder and CEO of Onetab.ai, argues that provenance could remove some of the opacity that has made GenAI commercially attractive. Agencies have been able to use AI for ideation, drafting and versioning without that production process necessarily being visible downstream. Machine-readable provenance potentially makes AI involvement legible.

That does not mean all AI content suddenly becomes less valuable. Instead, Dandotia expects the market to bifurcate.

High-volume functional work, including product descriptions, localisation and performance-ad variants, could be openly AI-generated and priced accordingly. Brand-building work is different. There, he argues, an AI-generated tag could operate almost like a “discount sticker”, while verified human authorship could eventually acquire signalling value precisely because human effort is scarcer.

That creates a rather uncomfortable question for an agency business historically built around people and hours: if the machine did substantially more of the production, why should the client continue paying human-production rates?

Vishal Mehra, CEO of Popkorn, expects that conversation to arrive quickly. Clients, he says, will attempt to compress agency fees on the grounds that AI has reduced production effort, while agencies will increasingly have to price themselves around strategy and thinking rather than execution.

There is also the question of who pays when things go wrong.

Read e4m report on AI & brand responsibility 

“If an AI heavy campaign gets flagged or throttled after it goes live, who will absorb the loss?” Mehra asks. He expects liability clauses around AI use to emerge in agency contracts, comparing them with existing provisions governing stock imagery rights.

Not everyone sees provenance as a threat to the economics of creativity.

Rahul Vengalil, Co-founder and CEO of Tgthr, argues greater transparency is ultimately positive for advertising. Generative AI has made previously uneconomical forms of visual storytelling possible, while the democratisation of production tools simply moves differentiation elsewhere.

Brands, he says, will ultimately differentiate themselves “not… with the tool but with the creativity”. Strong AI-led work still requires the thinker or copywriter, the art person and increasingly the prompt engineer.

Pankaj Srivastava, Founder and CEO of UnoSearch, similarly expects the conversation to move from whether AI was used towards how responsibly it was used. Agency agreements, he says, could eventually include clauses covering AI-assisted production, provenance disclosure, intellectual-property safeguards, and quality assurance, while clients place a greater premium on strategy, original research, and human verification rather than sheer content volume.

Read more on AI-Agency matrix 

There is, however, a rather large technological asterisk.

Watermarks are not infallible proof of authorship. Anthropic's mark indicates that Claude was involved in processing text; it does not necessarily establish how much of the underlying work originated with the model. Meanwhile, developers were claiming workarounds almost as quickly as the watermark was announced, including tools that attempt to break the statistical pattern by paraphrasing marked text through other models.

Dandotia consequently cautions that if detection remains leaky, contracts and disclosure requirements may ultimately matter more than the watermark itself.

Which leaves the AI industry in a mildly absurd arms race: companies are spending billions building machines capable of writing increasingly like humans, then building another technology to prove that a machine wrote it, while developers build yet another machine to make that proof disappear.

For agencies, though, the joke may eventually land on the rate card.