Is Human supervision enough to prevent cultural bias in AI ?


AI bias

Artificial intelligence is becoming an increasingly important tool in marketing and communication. From generating advertising visuals to translating campaigns and personalizing content, AI offers brands new opportunities to reach audiences more efficiently. Yet these benefits come with risks. As highlighted by Emma Vidal Madjar in her article, “AI model biases: the reputational risk your brand doesn't see”, AI systems can reproduce existing stereotypes and expose companies to reputational damage.

But identifying AI bias is only part of the problem. The answer cannot simply be to ask employees to check AI-generated content. Human judgment has its own limitations, and companies must rethink how they supervise and govern the use of AI.

However, if AI systems reproduce existing biases and humans are not always capable of recognizing them, how can companies ensure that their cross cultural communication does not damage their reputation?

Table of Contents

Where does AI bias come from?

AI models are trained on vast amounts of data, much of which comes from the internet. However, this information does not represent all cultures, languages and social groups equally. As a result, AI systems can reproduce dominant perspectives and reinforce existing stereotypes.

A 2024 study published in PNAS Nexus examined five widely used language models. Researchers found that their outputs consistently reflected cultural values closer to those of English-speaking and Protestant European countries.

Interestingly, the study also explored how prompts could help reduce this bias. Specifying a cultural identity in the prompt improved cultural alignment in 71% to 81% of the countries and territories examined for the newer models.

Visual international communication

This highlights how AI bias is influenced not only by training data, but also by how users interact with these systems. Prompts can reinforce stereotypes through the assumptions they contain, or help provide a more specific cultural context. Even so, better prompts cannot eliminate every underlying bias.

For international communication, this distinction is particularly important. A campaign that appears appropriate in one market may contain cultural references or representations that are inappropriate elsewhere.

The Human factor: when supervision fails

If AI can produce biased content, human supervision should theoretically prevent it from reaching the public. However, this assumes that employees can consistently identify problematic outputs. In reality, humans have unconscious biases of their own. They may overlook stereotypes because these representations reflect familiar assumptions or their own environment.

Visuel controverse de campagne

The Converse advertising controversy illustrates this problem. In September 2026, the brand removed a social media advertisement featuring K-pop singer Karina following criticism that its imagery evoked Ku Klux Klan symbolism and the history of lynching. Critics interpreted the lighting on a white garment as resembling a KKK hood and the way the sneakers were held as evoking a hanging victim. Converse apologized, acknowledging that it had got the image wrong.

How did such imagery manage to pass through the campaign's approval process? The brand did not publicly explain how the advertisement had been reviewed. This example demonstrates that even experienced creative and marketing teams can fail to recognize cultural sensitivities before publication. Adding AI to the creative process does not automatically solve this problem.

The risk of automation bias

Another risk is automation bias, the tendency to perceive AI-generated outputs as reliable and accept recommendations without sufficient independent verification. This raises concerns about employees who may trust AI-generated content simply because it appears objective or professionally formulated.

Research reported by The Guardian in June 2026 also illustrates the potential consequences of excessive reliance on AI. In a four-week experiment involving 67 participants, AI assistance initially improved participants' ability to identify misinformation. However, their independent performance declined by 15.3% by the fourth week. Although this experiment concerned misinformation detection rather than marketing, it highlights the broader risk of relying on AI which can weaken the habit of questioning information independently.

Shadow AI: a challenge for companies

Beyond these individual limitations, companies also face a wider organizational challenge. Even when they establish rules for responsible AI use, employees do not necessarily follow them. This phenomenon, known as Shadow AI, refers to the use of unauthorized AI tools within organizations, often without the knowledge or approval of management.

According to UpGuard's 2025 report, based on a survey of 500 security leaders and 1,000 employees worldwide, 81% of employees and 88% of security leaders reported using unapproved AI tools. Furthermore, 45% of workers said they found ways around restrictions to access blocked applications. These figures reveal a significant gap between company policies and actual employee behavior.

For marketing teams, this creates additional risks. Employees may use unauthorized tools to generate campaign visuals, translate advertising messages or write social media content without appropriate validation. Consequently, companies may have limited visibility into how their content is produced and whether it contains biased representations.

Therefore, simply prohibiting certain tools is unlikely to be sufficient. Companies need to understand why employees use them and provide accessible, approved alternatives.

Visual shadow AI

Beyond Human supervision: building responsible AI practices

If neither AI nor human judgment is entirely reliable, companies need a more structured approach to managing AI-generated communication.

In order to help reduce these risks, companies can :

  1. Develop AI literacy. Employees need to understand how AI works, where its limitations lie and how biases can influence its outputs. Article 4 of the European AI Act establishes an obligation for AI providers and deployers to take measures to promote AI literacy among relevant staff and other people working with AI on their behalf.
  2. Establish clear internal guidelines. Companies should identify approved tools, define acceptable uses and specify which content requires mandatory human validation. In addition, employees should also be encouraged to disclose their AI use rather than hide it.
  3. Introduce diverse validation teams. Human supervision should involve people with different cultural backgrounds, professional expertise and perspectives. For international campaigns, consulting local communication professionals or representatives of the target audience can help identify inappropriate references that an internal team might overlook.
  4. Monitor and evaluate results. Validation should not stop when content is approved. Companies should regularly audit AI-generated outputs, document recurring problems and monitor audience feedback after publication. By doing so, they can identify recurring issues and adapt their guidelines over time.
Visual responsible AI practices

Ultimately, effective human oversight requires more than assigning someone to approve AI-generated content. Reviewers need the knowledge and time to challenge recommendations and reject inappropriate outputs.

Conclusion: AI needs Human judgment, not blind trust

AI bias is not exclusively a technological problem. It is also a human and organizational challenge. While AI can reproduce existing stereotypes, people can reinforce them, overlook them or trust automated recommendations without sufficient scrutiny.

For brands, responsible AI use therefore requires more than a final human check. It demands employee education, transparent internal policies, diverse perspectives and continuous evaluation.

The objective should not be to remove humans from the creative process, but to make their involvement more informed, critical and accountable. In this context, human supervision should be part of a broader system of responsibility, rather than the only safeguard against cultural bias. Ultimately, protecting brand reputation in the age of AI depends not only on what technology can generate, but also on how responsibly companies choose to use it.

For more information about how this article was developed, including the AI tools used and the prompt engineering process, please refer to the methodological note below.

Juli Swart LinkedIn picture

Article by Juli Swart