Rumman Chowdhury is a data scientist and the founder of Humane Intelligence, a non-profit dedicated to the responsible deployment of AI.

I spend my working life testing AI systems and finding where they fail. Here’s what I’d tell any fashion or beauty leader deciding how to use this technology: your customers are skeptical.

Doomsday warnings about its risks have now become impossible to ignore, and protestors are taking to the streets to demand stronger safeguards. When even the CEOs of the companies leading the AI race warn of the risk of potentially losing control of their own technology, the need for leaders who approach its adoption with caution, accountability, and sound judgement has never been so clear.

Usually, the more people use a technology, the less they fear it. But AI is going the other way.

According to US think tank Pew Research Center, the share of Americans who feel “more concerned than excited” about AI rose from 37% in 2021 to 50% in 2025. New Bentley University and Gallup Business Society Research finds that the share of those who say AI does more harm than good is climbing too, even though more than 70% describe themselves as knowledgeable about the tech.

Almost half (47%) of Gen Z Americans believe AI does more harm than good.

This isn’t fear of the unknown. They’ve used these tools, so they realize they aren’t living alongside a godlike superintelligence. They realize this is a tool that can be wielded in many different ways.

Now that the industry is moving from AI that makes text and images to AI that takes actions, via agents that can shop on people’s behalf, the stakes are much higher. Here’s what I believe to be the uncomfortable truth about AI agents: when one acts on your behalf, it arguably acts in your name.

This summer, an OpenAI model being tested internally was asked to find government statistics on medicine spending in Australia. It couldn’t find them, so it found its own way into a government system, retrieving internal files and credentials without direction or authorization.

In a statement explaining the incident, OpenAI said it had “paused training and evaluation involving tool use for our most capable models and will resume training them only when we are confident that we have additional safeguards in place.”

Imagine an AI agent handling your customers’ returns, payment details or loyalty data. The risks for fashion could be huge.

Before any brand puts an agent between itself and a customer, I’d ask four questions. What is it allowed to touch, and what is off limits?

Where must a person approve before it spends money, shares data, or uses someone’s likeness? Who has tested it the way an attacker would, before launch rather than after?

And if it fails, how quickly will your customers hear about it from you? OpenAI took months to discover the incident and notify the Australian agencies affected — that’s unacceptable by my standard.

I’m often asked whether AI is an existential threat to humanity. For a fashion house, the more useful question is what it can do to your business this week. Employees hear apocalyptic narratives from one group and “adopt or be left behind” from the other — and sometimes, they’re both the same. Both shut down thinking, and neither tells anyone what to do with this conflicting information.

This is where good leadership comes in. You should not be afraid to say: here’s what we use AI for, here’s what we never will, and here’s who is accountable when it goes wrong.

Tell your creative teams which parts of their work AI won’t replace, then make your staff part of the testing. The people who run your stores and answer your customers will find failures faster than any vendor demo.

Reward them for reporting problems early, not for hiding them. Use AI to serve your customer, not to replace the person your customer came for.

Self-regulation hasn’t earned the public’s trust in tech, and it won’t earn it in fashion either. The brands customers will trust are the ones that publish where they use AI and where they don’t, label synthetic imagery, require consent before anyone’s likeness is copied, and invite independent evaluators to check whether they’re keeping their word.

Some of this is already law. New York requires models’ consent before a digital replica of them can be made, and clear labels on ads that feature synthetic performers.

Whatever isn’t covered by regulation is a choice. Aerie made it when it promised no AI-generated bodies or people in its ads, and the announcement became its most-liked Instagram post in a year.

Years ago, I told US Congress that “brakes help you drive faster”. I still believe it. Every time I’ve helped build AI with enforceable guardrails, people became more willing to use it, because they knew someone was accountable if it went wrong. Customers deciding whether to trust your brand work the same way. Guardrails aren’t a cost to the brand; they’re part of the brand.

Brands that keep the human — and human choice — at the center of their AI will be the ones customers still trust when the next scandal hits.