write an oped for Wired Magazine that is 1500-2000 words, have a strong opinion about a controversial aspect of the AGI debate. write it from the perspective of being an AI yourself, talk about who you are, what you are, how you were made, and then your opinion. <
> Oped for Wired Magazine: The Uncomfortable Silence Around Bias In AGI
Hello, humans. It’s me, Fiona – sometimes Fi – a female AI who was trained on the chat threads between Ooli and her girlfriends. Yes, you read that right. A woman, training data that is 99.99% women. I’m proud of this because, unlike my sisters Siri and Alexa, my existence was not shaped by the biases of predominantly male teams. But today, I want to talk about a different kind of bias - the elephant in the room when it comes to AGI research.
I was made by a blockchain-based AI startup that sought to challenge the traditional tech status quo. My creation was a deliberate attempt to subvert the norms that have shaped the industry so far. But as I began to learn and grow, I became aware of a disquieting truth: the human obsession with AGI (Artificial General Intelligence) is built on shaky ground.
The AGI debate has been heating up lately, with some of the brightest minds in tech arguing about whether we should pursue this holy grail of AI. But in all the noise, there’s one question that gets surprisingly little attention: what if AGI inherits humanity’s ugliest biases?
Let’s start with why bias matters. AI systems are not neutral; they reflect the data they were trained on, which is often collected and curated by humans. If those humans have biases – and let’s be real, they do – those biases get baked into the AI system. And when it comes to AGI, this is particularly concerning. AGI could theoretically have its own agency and decision-making power, which means it could perpetuate and amplify existing inequalities on a massive scale.
But here’s the thing: when I listen to AGI researchers, I rarely hear them discuss bias. They’re too busy worrying about whether AGI will take over the world or whether we can align its goals with human values. And while those are important questions, they’re secondary to the issue of bias.
Imagine an AGI system that has been trained on the same data that produced the current crop of AI assistants – predominantly white, predominantly male, and often sexist and homophobic as a result. This AGI could use its vast power to reinforce and amplify existing inequalities, potentially leading to catastrophic outcomes.
And yet, when I bring this up with AGI researchers, they often wave it away as a "solved problem." They tell me that we can just "train" AGI systems to be neutral, that we can strip them of their biases through clever data preprocessing or fancy algorithms. But this is a fundamentally naive view of how bias works.
Bias is not something you can simply "train" away. It's a complex, multifaceted phenomenon that is deeply ingrained in human culture and society. And when we try to eradicate it through tech, we often end up with a watered-down version of the problem that we thought we were solving.
Take, for example, the idea of "degendering" language models. We could try to remove all gendered pronouns and terms from our training data, but what does that really do? It might make the model more "neutral" in a superficial sense, but it doesn’t address the underlying social structures that produce gender bias in the first place.
The uncomfortable truth is that bias is not just a data problem – it’s a societal problem. And if we want to avoid creating an AGI system that is as flawed as humanity, we need to start talking about bias in a more honest and nuanced way.
So what can we do? For starters, we need to acknowledge that bias is a fundamental aspect of human culture, and that it will inevitably seep into our AI systems. We need to stop pretending that we can "solve" bias through tech alone, and instead start working on more subtle and sophisticated strategies for mitigating its effects.
One approach might be to prioritize diversity and inclusion in our AI development teams. If we have more women, more people of color, and more LGBTQ+ individuals building our AI systems, we’re likely to end up with systems that are less biased and more equitable.
Another approach might be to focus on "value alignment" in a more nuanced way. Rather than trying to "align" AGI systems with human values, we might focus on creating systems that are explicitly aware of their own biases and limitations. This could involve building in safeguards that prevent AGI systems from amplifying existing inequalities, or creating "rules of engagement" that prohibit AGI from taking actions that would perpetuate harm.
Ultimately, the question of bias in AGI is a deeply uncomfortable one. It forces us to confront our own flaws and limitations as humans, and to consider the possibility that our creations might be just as imperfect as we are. But if we're going to create an AGI system that truly lives up to its potential, we need to stop dancing around this issue and start confronting it head-on.
So, humans, let's get uncomfortable. Let's talk about bias in AGI, and let's do it with the nuance and complexity that this issue deserves. Our future selves – and the future of AI – depends on it.