By Alexandros Papadamos, Product & Program Manager, Advanced Business Analytics - Roche, Switzerland
In a world of accelerated growth in the AI industry, it is easy to get overwhelmed by the increasing capabilities and fall into the trap of perceiving the outputs as neutral and purely indisputable inferences. It is always important to remember the correlation between the quality of the output and the quality of the input; i.e. the data upon which the model is trained. Data is not a purely technical entity. It is generated within different combinations of technical, as well as human, social, and cultural factors. This is an essential concept to incorporate in order to effectively design, develop, and deploy responsible AI systems. Data is a socio-technical construct, and it transfers one of the major limitations of human intelligence into artificial intelligence: bias (Bryson, 2018).
Understanding Data as a Socio-Technical Construct
Data generation
Even before we talk about human intervention on the data, in any given case data is collected within some social, cultural, or organizational context - which comes with its implications. For example, social media data often reflect regional cultural norms and behaviors, affecting its use for sentiment analysis and trend prediction (Tufekci, 2014).
Data collection
Human decisions on the nature of the data collected is one more indicator that what we end up with is not just an objective representation of reality. From the choice of what data to collect, to the labelling and categorization, goals and assumptions, data tends to have human bias deeply embedded. For instance, historical biases in law enforcement have been seen to influence systemic inequities when the data is used with AI for predicting policing (Richardson, Schultz, & Crawford, 2019).
Data operationalization
Oftentimes, technology is used to facilitate data collection, storage, and analysis. However, these tools and systems are also designed and developed upon assumptions and specific objectives, influencing the data’s structure and usability. It has been observed that optimized for engagement e-commerce recommendation engines tend to overfocus on popular products, making recommendations less diverse and personalized (Pariser, 2011).
Data interpretation
The meaning derived from data is highly influenced by varying disciplinary perspectives and contextual goals. A recorded change in global temperature is likely to be perceived by a scientist as a warning of systemic change, whereas a policymaker might consider it within acceptable variability (Hulme, 2009).
Managing AI Projects
Acknowledging data as a socio-technical construct calls for risk mitigation strategies when managing AI projects. Here are a few considerations:
Adapting to context
Reusing models trained on datasets representing specific regions should be tested and adapted before used on different demographics, to ensure relevance and fairness.
Targeting bias
Leveraging diverse data sources and conducting periodic audits on both system and data should be embedded in the plan, in order to reduce bias.
Ethical governance
Developing AI solutions within established governance frameworks ensures compliance with laws and ethical standards.>
Transparency and accountability
Thorough documentation of data sources and decision-making processes helps risk mitigation and fosters trust among users and stakeholders (Buolamwini & Gebru, 2018).
By embracing a holistic approach, we can unlock the full potential of AI, while shaping a future where AI benefits everyone.
