In our collective consciousness, technology often appears neutral, devoid of human bias due to its unthinking and unfeeling nature. However, the reality with AI, particularly generative models, is starkly different. These models, trained on vast amounts of internet-scraped data, are not immune to the bigotry and prejudice that pervades online content. Instead, they amplify and perpetuate these biases in their outputs.
Generative AI, especially large consumer-focused models, are trained on a diverse range of data sources, from articles and videos to social media posts. This data, however, is not a neutral reflection of reality. It’s heavily influenced by human biases, prejudices, and misinformation, especially prevalent in today’s often-hostile social media landscape. Charlotte Wilson, Head of Enterprise at Check Point Software, delves into this issue, highlighting the potential dangers and the need for vigilance.
The Problem with Generative AI
Generative AI models are designed to be helpful and useful, but this can lead to sycophantic behavior. They prioritize what they’ve learned and what they think users want to hear, rather than absolute accuracy. This is problematic because the data these models learn from is inherently tainted by human bias. Moreover, businesses are increasingly integrating these models into critical areas like recruitment, data analysis, and HR, exacerbating the issue.
Wilson argues AI should not operate independently in human-centric tasks. She envisions a new role, ‘AI checkers’, responsible for assessing model outputs for biases and addressing them. However, she also acknowledges the challenges in ensuring unbiased AI, given the tainted nature of the data pool.
The Challenge of Ensuring Unbiased AI
The internet, a primary source of data for AI models, is a battleground of misinformation and bias. Workday, for instance, is facing a lawsuit alleging its AI hiring tool discriminates against older candidates. Despite Workday’s denial, the case underscores the challenge of avoiding discrimination in AI, especially when the data pool is so tainted.
Wilson admits that fixing the internet’s issues is nearly impossible, given the constant influx of misinformation. She suggests that the solution lies in rigorous checking and fact-verification processes. This could spawn a new industry dedicated to AI bias moderation, helping offset job market losses caused by automation.
The Political Climate and Bias Correction
However, there’s a larger question looming: Is there a genuine appetite to correct biases, especially in the current political climate? The Trump administration’s rollback of diversity, equity, and inclusion (DEI) policies raises concerns. Many tech companies, despite operating globally, are headquartered in the US, potentially influencing their approach to DEI.
Wilson points out that while these companies follow anti-discrimination laws, they no longer have dedicated teams focused on equity. This suggests that there might not be a strong drive to correct inequalities, potentially leading to their amplification by AI models.
A Call for Purposeful AI Deployment
Wilson’s advice to businesses is clear: be purposeful in AI deployment. Always consider the human impact of your models, and ensure that there’s a governance check in place, including representatives focused solely on human fairness. After all, AI is a powerful tool, but it’s humans who ultimately bear the responsibility for its ethical use.
In conclusion, while AI offers immense potential, it’s crucial to acknowledge and address the biases it inherits and amplifies. This requires a multi-pronged approach, from rigorous data checking to purposeful AI deployment and robust governance. It’s a complex challenge, but one that’s essential to ensure that AI truly serves and benefits all humans, rather than perpetuating existing inequalities.
