At Intent HQ, we are at the forefront of navigating the complex terrain of AI biases, a critical aspect that shapes the future of fair and equitable AI systems. In our endeavor, we understand that AI’s transformative power hinges not just on its technical prowess but on its ethical application. This commitment aligns with our broader mission: to turn vast, unstructured data into actionable insights, ensuring privacy and security remain paramount.
AI biases manifest in various forms, from reporting bias, which skews data representation, to automation bias which can lead to over-reliance on automated systems. Selection and exclusion biases further complicate the landscape by influencing which data is considered relevant. These biases, if unchecked, can perpetuate inequality and hinder our mission to deliver insights that genuinely reflect individual behaviors and preferences.
Our work involves grappling with the intricate challenge of ensuring accurate representation in our AI models. Accuracy, though vital, is not synonymous with fairness.
The implications of unchecked AI biases are glaringly visible, providing us with lessons on the critical need for vigilance and proactive measures. For example, the predictive crime model launched in Chicago in 2010, which aimed to forecast criminal activities without relying on racial or demographic data, inadvertently concentrated its predictions within Black and Hispanic neighborhoods, which represent a small fraction of the city’s urban residential areas. Similarly, high-profile cases like Amazon’s recruitment tool, which was found to favor male candidates over female ones, and a study revealing Google’s ad targeting system provided men more opportunities for high-paying jobs than women underscore the pervasive nature of biases in AI.
These instances are not just technical oversights. They serve as stark reminders of the ethical responsibilities that come with deploying AI, which, without careful guidance, can amplify rather than mitigate bias.
Recognizing the legal and ethical implications, we advocate for a proactive stance on AI ethics. The regulatory environment, while evolving, underscores the need for vigilance and ethical deliberation beyond compliance. It’s a call to action for organizations to wield their moral compass, especially in areas where the law may lag behind technological advancements.
In our pursuit of crafting unbiased AI, our inspiration often comes from the cutting-edge methodologies employed by pioneers like OpenAI in the development of ChatGPT.
OpenAI embarked on this ambitious project with a two-staged approach: beginning with unsupervised pre-training across vast corpora of open-domain datasets to cultivate a baseline of general language understanding. This foundational layer, while comprehensive, was not immune to the pitfalls of sensitive, unfair, or factually incorrect content that often infiltrates training data. To address this, a nuanced fine-tuning phase was implemented, leveraging reinforcement learning with human feedback. This crucial step involved human labelers, who were tasked with evaluating the outputs of ChatGPT to ensure they were non-toxic, accurate, andaligned with ethical standards. Labelers were instructed to assess these outputs with a keen eye on their helpfulness, truthfulness, and harmlessness.
Example of rating card provided by OpenAI to human labelers:
This process goes beyond mere technical refinement; it is an intentional effort to weave human values into the fabric of AI. Evaluators are meticulously chosen based on their ability to navigate sensitive topics, with considerations spanning agreement on identifying sensitive speech, consensus in ranking responses, capability in writing demonstrations sensitive to diverse groups, and a self-assessment of recognizing speech sensitivities across different cultures or topics.
OpenAI Human labeler demographic data:
Despite these efforts, it’s acknowledged that such meticulous selection criteria, while aiming for fairness, don’t completely eliminate biases. They underscore the significance of integrating human judgment and the necessity for continuous, rigorous testing. Special emphasis is placed on critical evaluation to assess AI outputs, addressing the neural networks’ inherent opacity.
Additional avenues to diminish AI bias are available, such as developing a heterogeneously balanced dataset that could counteract long-standing societal inequalities. This approach, while corrective, reaffirms the indispensable role of human oversight in AI development. It brings to light an essential truth – the journey towards unbiased AI is not solely a technological endeavor but a deeply human one. It reminds us of the ancient wisdom that “Knowing yourself is the beginning of all wisdom.” as Aristotle famously stated.
At Intent HQ, our team’s diversity is our strength. It informs how we approach model development, from data selection to testing, ensuring our insights resonate across cultures and communities. We believe in harnessing the collective power of diverse data sets, and counterbalancing biases with a more inclusive vision of AI.
This path towards ethical AI is not just a technical challenge; it’s a reflection of our values. It demands a blend of legal adherence, ethical commitment, and a deep understanding of the human aspects behind the data. As we continue to navigate this landscape, our focus remains steadfast on delivering AI-driven insights that not only respect privacy and security but also embody our commitment to fairness and equity.
As we look forward to sharing these insights at the upcoming ANA AI for Marketers Conference, we invite you to join us in this ongoing dialogue. Together, we can shape a future where AI not only predicts but also respects the rich tapestry of human diversity.
ANA Panel Discussion: Combating Inherent Biases in AI
10th of April – 10:38am – 11:10am
Our Chief Product Officer, Sharifah Amirah, joins the panel alongside industry leaders to discuss the empowering uses of AI for all women and girls.
Moderator: Christine Guilfoyle, President, SeeHer
Sharifah Amirah, Chief Product Officer, Intent HQ
Traci Spiegelman, Vice President, Global Media Mastercard
ABOUT THE AUTHORS:
Sharifah Amirah, Chief Product Officer, Intent HQ
Sharifah Amirah is the Chief Product Officer and Chief Client Officer for Intent HQ, where one of her passions is to create a greater social impact through the application of technology, especially in bridging the gender gap and enhancing women’s economic empowerment.
Previously, Sharifah collaborated with technical and commercial executives across various industries in Asia, Europe, and North America to drive revenue and cost savings through AI and analytics.
Check out Sharifah’s Linkedin here.
Chris Schildt, Applied Data Scientist, Intent HQ
Chris has formal training in Philosophy and Computational Neuroscience, and an interest in the mind and being as far back as his earliest memories. For the past 5 years he’s worked at the intersection of AI and marketing by helping brands connect with their customers in more meaningful ways. He spends most of his non-professional time with his 6 month old daughter, AI… yes, those are actually her initials.
Check out Chris’s Linkedin here.