Can AI labels restore consumer trust?

Generative AI blurs the line between truth and fabrication, but over-policing stifles creator spirit, and under-policing hollows out public trust

Featured faculty: Tony Ke
Written by Putro Harnowo

Appearance can be deceiving, and this idiom deeply resonates in modern times. Not too long ago, a French woman lost her life savings to an impostor posing as actor Brad Pitt on Instagram, while YouTube star MrBeast called out a fake TikTok video of him offering a new iPhone for just US$2 each. These cases are just the tip of the iceberg.

AI can generate anything as easily as typing a sentence, fueling fraudsters to produce deceptive content at a remarkable scale. More and more people fall victim to AI forgeries, often exploiting public figures to trick netizens. As a result, the EU has required digital platforms and content creators to label AI-generated content starting this August.

In Asia, China was the first to mandate AI labels last year, followed by South Korea and Vietnam earlier this year, while India is pondering the same. Overall, these AI laws require companies to state clearly when publishing AI-generated content, but some critics question their effectiveness in tackling disinformation and misinformation.

Tony Ke, Professor of Marketing at the Chinese University of Hong Kong (CUHK) Business School, points out two distinct ways in which AI labels could lose their power. “First, consumer attention towards AI labels may decline after repeated exposure through psychological habituation, but changing the design or placement of AI labels can refresh such attention,” he says.

“Secondly, an AI label becomes helpful when it makes people aware that content made with AI might be fake and deceptive. However, if AI tools become so easy and affordable that both truthful and deceptive creators use them frequently, the AI label becomes merely a descriptive label.”

Therefore, the more accurately AI labels identify manipulative content, the more likely people are to associate AI labels with a risk of deception, and that is the central focus of Professor Ke’s latest study.

Deceptive content typically aims to appear as genuine as possible, so Facebook, Instagram, TikTok, and LinkedIn have deployed algorithms to detect visual patterns, statistical patterns, and metadata to score the likelihood of content being made with AI.

These algorithms are imperfect. Reports show that AI detectors fail to distinguish AI-generated output. Human-made content may also have characteristics like AI output, so if the platform sets the detection algorithm too aggressively, the algorithms may mislabel human-made content. Conversely, making the algorithm too conservative may risk missing the real AI content.

Optimising AI labelling to tackle deception

To find the optimal balance, Professor Ke partners with Chen Jieteng of Shanghai Jiao Tong University and Shin Jiwoong of Yale University. Their study, Labelling AI-generated content on platforms: misinformation, consumer inference, and creator incentives, develops a theoretical model of strategic interaction among three key players: the platform, users, and content creators.

Deceptive content creators have stronger incentives to adopt AI to generate misinformation because they are not constrained by factual accuracy, and such misinformation can be particularly effective at attracting user attention and engagement. Consequently, users may interpret AI labels on content as a warning sign of potential deception.

Therefore, mandating salient AI labels on content may discourage user engagement and deter deceptive creators from using AI-generated content. This explains why labelling matters in real-world settings and why such labels can discipline AI misuse without arduous manual fact-checking.

“Transparency can protect a brand when it communicates governance and responsibility. A bare AI label, without context or accountability, can instead transfer the general scepticism surrounding AI directly onto the brand.”

Professor Tony Ke

Aggressive labelling algorithms tend to flag more content as AI-generated, which strengthens users’ trust in unlabeled content but may discourage truthful creators, as some of their content may be mistakenly labelled as AI-generated and consequently receive less user engagement. If the labelling policy becomes more lenient, more AI-generated content goes unlabeled, which may increase user engagement. However, more deceptive content may also slip through, causing harm to users and, in extreme cases, undermining their trust even in unlabelled content.

When AI tools become so easy and cheap for anyone, AI content becomes abundant, and platforms need to apply more aggressive labelling to maintain credibility. Only when algorithms can accurately spot deceptively real content can platforms adopt more lenient labelling that maintains users’ trust. However, near-perfect detection may be impossible.

“Deceptive creators nowadays can conceal metadata, use third-party tools, or tweak content to evade AI detection. Therefore, self-disclosure, algorithmic detection, and digital markers should be used together to combat deceptive content,” Professor Ke says. “The main deterrent comes from users’ association between AI labels and the risk of manipulation. Mandatory labels will discourage deceptive AI use when they are credible and consequential, but labels that are easy to evade are unlikely to have an effect.”

Using AI content while keeping brand reputation

In general, consumers may perceive AI content as unreliable or low effort, even when the underlying message is truthful. While consumers do not intrinsically dislike AI content, scepticism emerges because deceptive creators have stronger incentives to use it. Reputable brands can be caught in a negative association.

Mandatory labelling can deter companies from deploying AI content, but Professor Ke notes that such content can still be useful if the benefits outweigh the risks. “There is no universally optimal ratio to balance the labelling policy. For instance, the consequences of posting AI-generated parody are not equivalent to sharing fabricated medical endorsement,” he says.

“For marketers, the risk depends highly on the context. Consumer scepticism is likely to be damaging where authenticity is central to the product, such as luxury craftsmanship, personal services, healthcare, finance, and the like, or when using AI to create testimonials and product demonstrations.”

Businesses can further protect their brands by explaining the role AI played. Putting a disclaimer that says, “AI-generated content, reviewed and approved by our editorial team,” while also clearly displaying an AI label can be a fine example. “To maintain accountability, a named editor, designer, expert or brand team should remain visible for factual claims,” says Professor Ke.

Brands can substantiate marketing claims with accessible sources, original product images, and other factual records, rather than just visually appealing content. Brands should also consistently disclose when they use AI, rather than admitting it only after consumers point it out.

“The most defensible message is using AI for a defined purpose, having humans verify the result, and the company remaining accountable for every claim. That distinction will become increasingly important as AI use becomes the norm,” he adds.

“Transparency can protect a brand when it communicates governance and responsibility. A bare AI label, without context or accountability, can instead transfer the general scepticism surrounding AI directly onto the brand.”