“We are the music makers, and we are the dreamers of dreams” – Ode, by Arthur O'Shaughnessy.

As researcher practitioners, we are in a privileged and responsible position to reflect the reality of the world as best we can.  It’s a heavy crown, one that requires humility, curiosity and good judgment.  

The data and insights we provide have a fundamental impact on commercials and the future of people’s lives, but they all have their flaws which are widely known and understood.   

This new frontier of AI represents different and new challenges for us. Certainly there are many positives, but knowing how best to interrogate and use AI tools will be key to ensuring we don’t inadvertently perpetuate bias in society.  

This is difficult because AI systems are generally complex, opaque and misunderstood, so identifying bias or inaccuracies is difficult. There’s also an innate ‘trust’ we have of their outputs, that translates into authority. AI outputs are often positioned as certainty: fluent language, polished summaries, confident recommendations. This presentation can make weak or biased outputs feel authoritative, even when the underlying reasoning is flawed.   

Researchers are increasingly using AI systems whose biases are less visible, less able to be interrogated and more scalable than traditional research biases.   

The impact of these challenges is real:  

• A 2024 study in PLOS Digital Health found that large language models generated less accurate and sometimes harmful medical advice when a patient was identified as LGBTQ+, even when the underlying condition was identical (Celi et al., 2024 ).  

• A review in Nature Humanities and Social Sciences Communications shows that facial recognition systems are significantly more likely to misidentify older adults, reflecting systemic gaps in the data used to train them (Rosales & Fernández-Ardèvol, 2023 ).  

• Research from Stanford University’s AI Index highlights that systems trained on predominantly Western datasets consistently underperform for Global South users, misinterpreting language, context and intent (Stanford HAI, 2023 ). 

A synthetic persona designed to represent “Gen Z consumers” may default toward Western, urban, digitally fluent behaviours because those identities dominate training data. Marginalised experiences can become flattened, exaggerated or erased entirely, while still appearing plausible on the surface.  

The solution lies in a mind-shift. Thinking about inclusive AI at every stage of the AI development process and as something everyone in the industry can influence. That means treating inclusion not as an ethical add-on, but as part of data quality itself.  

There are practical ways to start addressing this, several of which are outlined in the recent MRS AIA Council paper on Systemic Inclusion by Design. 

One example is the use of a systemic inclusion canvas during the initial design phase of an AI tool. This acts as a framework to help teams think more critically about what they are building, and why. Questions might include:  

• What inclusive values are guiding the tool’s purpose?  

• What systemic biases are we trying to reduce within the research ecosystem?  

• Are we solving a genuine equity problem, or simply scaling convenience? 

The second area could be practical bias testing protocols, that use established tools like fairness testing but under the lens of inclusion, answering questions such as whether model performance, outputs, or recommendations differ across identity groups.  

Lastly, inclusive model scoring cards could be employed to support transparent implementation, for example scoring the model outputs on how culturally appropriate, understandable and accessible they are.  

More inclusive and transparent AI systems are not just better ethically, they are increasingly becoming commercially advantageous. PwC research found organisations with stronger responsible AI practices were more likely to report improvements in return on investment, efficiency, innovation and customer experience.

AI tools are not magical, they are actually things that can be controlled. There are actually many opportunities in the development and use of them that can be influenced, and not just by engineers and chief technology officers, but by everyone. In fact, everyone needs to be engaged in this conversation. 

Fundamentally, this is not a nice to have, it’s imperative for us to align the tools we are using to ensure they are as accurate and unbiased as we can achieve. The impact of which will be felt on businesses and society for a long time to come.  

Sabrina Trinquetel is account manager at YouGov, member of the MRS AIA Council and co-chair of MRS Pride