As Head of Digital Production at McCann Singapore, Jose Siojo is responsible for ensuring the tools used in creative production are approved and regulated, while helping creative teams use them effectively across diverse markets.
For Siojo, the prompt is the brief, whether we treat it that way or not.
IMDA’s February 2025 report found that AI safety guardrails were significantly less effective when prompts were submitted in regional languages rather than English—a finding Siojo says tracks with what he has observed in practice. Much of the training and safety testing behind today’s AI models remains centered on English-language contexts, meaning local nuance is often the first thing to be lost.
We recently spoke with Siojo about how generative AI is changing production workflows, where AI bias often appears, how he approaches inclusion, the role prompting plays in shaping AI outputs, why accessibility needs to be built into the brief from the start, and what brands can do to use AI more responsibly.
You’ve spent more than a decade working in digital content production. How has generative AI changed the way you approach a brief and the way the creative work gets produced?
Generative AI has changed how I look at a brief from an efficiency and compliance standpoint. Part of my role is making sure that whenever a brief involves generative work, we’re proposing tools that are approved and regulated by the company, not just whatever’s trending.
This helps us protect our clients from any potential copyright infringement, and provides them with confidence to use Generative AI effectively. And once that’s sorted, we guide teams into using the tools effectively to achieve the desired output.
Then the focus becomes optimizing the workflow around those tools so we’re not wasting time on repetitive production tasks, which frees up better resources, time, people, budget, for the actual creative thinking, which is where it should be going in the first place.
As AI-generated content becomes more prevalent, what kinds of bias or blind spots are you seeing most frequently in creative outputs?
The most common one is the “generic Asian” default, no specific ethnicity, no regional markers, often lighter-skinned and younger than the actual audience. It also shows up regionally, where Southeast Asia gets treated as one pan-Asian output, when the countries in this region are each unique and not interchangeable. The idea of “local” in a prompt would result in rural associations or imagery with plastic chairs — which does not reflect how our audience actually lives.
The most common one is the ‘generic Asian’ default, no specific ethnicity, no regional markers, often lighter-skinned and younger than the actual audience.
There is also a gender default toward male for neutral roles. None of it is malicious, just the model’s training data, but if we don’t catch it, the work can signal to part of our audience that they weren’t who it was made for.
IMDA’s 2025 red-teaming report found that AI safety guardrails were significantly less effective when prompts were submitted in regional languages rather than English. As someone working with these tools across multiple brands, what are your thoughts on that?
It tracks with what we’ve seen, since a lot of these models are trained and safety-tested mostly in English, so the guardrails are strongest there and get thinner once you switch to Bahasa, Tagalog, Thai, or Vietnamese. For us, this isn’t just a safety issue, it’s a relevance issue too, because local insights and nuance are what make work actually land with a market.
It tracks with what we’ve seen, since a lot of these models are trained and safety-tested mostly in English, so the guardrails are strongest there and get thinner once you switch to Bahasa, Tagalog, Thai, or Vietnamese.
If the model’s understanding is weaker in the regional language, that nuance is the first thing to go, so we tend to draft and review in English first, even if the final output needs to be in a local language, because that’s where the model is most reliable. This is also why having a human in the loop is so important — and we make sure local perspectives are accurately represented by having colleagues from various Southeast Asian countries check local nuances of any campaign.
Much of the conversation around AI and creativity focuses on risk. Where have you seen it genuinely help produce more inclusive or accessible work?
The clearest one is accessibility basics, alt text, captions, multilingual versions, which used to get cut as extra work but are now doable. The other is representation, since AI lets us generate options we couldn’t afford to shoot conventionally, and specific prompting opens up more representation, not less, by removing the cost barrier. But generating people is a rights issue too, so what tool we’re using, what it allows, and whether we’re covered needs to be part of that same conversation, not an afterthought.
Can you share a recent project where your team used AI to address accessibility or inclusion? What was the challenge, and what did you learn from the process?
We have projects with a Japanese insurance brand with a strong regional presence. To stay representative without defaulting to a single face or ethnicity, we leaned into an origami-inspired key visual, capturing the brand’s core values—precision, care, and Japanese craftsmanship—while sidestepping the casting problem entirely. Origami let us represent the brand without pinning down a specific race or individual, while still giving us a wide range of creative options for the market.
As the models get better, the output and the imagination get better too, more room for craft, more room for inclusion.
The challenge is also the opportunity. As the models get better, the output and the imagination get better too, more room for craft, more room for inclusion. The learning is that representation doesn’t always have to be about people, objects and non-human elements can carry cultural specificity just as well.
“Prompting for inclusion” is gaining traction as a concept. In practical terms, how much does the way you instruct an AI shape whose reality ends up reflected in the output?
The model isn’t neutral, it’s making a casting decision based on its training data, so the prompt is the brief, whether we treat it that way or not. Think in three parts: who the person actually is, what the cultural context is, and how specific the details are. “Southeast Asian” isn’t a brief, “a Javanese woman in her 50s in a kebaya at a warung in Yogyakarta” is.
That specificity is the difference between work made for someone versus made for no one in particular. And accuracy isn’t the finish line, the depiction should also feel aspirational and full of life. Same goes for rights: usage rights need to be part of the discussion too.
Do you think the industry is paying enough attention to accessibility when adopting AI tools, or is the conversation still mostly focused on efficiency and output?
Mostly efficiency and cost, but for big brands this is also a compliance and regulation question, not just a creative one, since bigger brands carry bigger risk, so there’s more pressure to work only within approved tools, and the real challenge is less about whether accessibility matters, and more about navigating the way into using these tools properly, and once you’re in, it’s important to make sure accessibility is part of the brief from the start.
Where does responsibility ultimately sit when AI-generated content fails to represent audiences accurately—the technology, the people using it, or the processes around it?
The review process carries a big responsibility. If that is weak, the bias gets through regardless of how careful any individual person is, and that’s part of why we had talks within our organization how to safeguard prompts for inclusion.
The review process carries a big responsibility. If that is weak, the bias gets through regardless of how careful any individual person is…
It’s not enough to assume people will catch it on their own, the process has to build that check in. We should always build accessibility in, with the brief.
For brands looking to use AI more responsibly, what are some practical steps they can take to make their content more representative and accessible?
Specify the brief, age, ethnicity, body type, ability, setting, and treat the prompt like a casting call. Build accessibility in from the start. Consider alt text, captions, and multilingual versions as part of the brief.
Check before generating, who is this person, is the context specific, are the details concrete enough. It just needs to become a habit. The tools can do a lot now. What they can’t do is know your market the way you do.
