In Our Own Image

This reflection is based on the monthly Tech Leaders Salon discussion, where we discuss books on topics including philosophy, strategy, management, and leadership. The salon is co-hosted with Laksh Raghavan as part of his Cyb3rsyn Community. If you’re interested in joining the conversation, you can learn more about the discussion group here. In August, we discussed Brian Christian’s The Alignment Problem.

From the Enlightenment to AGI

We are trying to create a new species in our own image, a species that can learn on its own, develop an inherent sense of curiosity, perceive the world and represent it adequately, and make decisions and act in the real world.

Ultimately, the goal seems to be the achievement of artificial general intelligence (AGI), with entities that are capable of exhibiting human-level intelligence to an extent that they are better than us, helping us achieve unprecedented levels of scientific, digital, and social progress. AGI would allow us to accelerate growth, eradicate disease, combat poverty, and make breakthroughs in record time. That’s the optimistic narrative, or the story the “effective accelerationist” movement is telling.

The techno-optimist worldview has roots in the scientific revolution and the 18th-century Enlightenment. Reason, the scientific method, and new tools made fresh empirical findings and economic progress possible. People had a sense that history was progressing linearly, with new milestones being built on earlier ones.

During the Enlightenment, the prevailing belief was that rational thinking would suffice in solving our problems and answering all our questions. In the absence of a satisfying solution or answer, people thought scientific and rational inquiry would eventually provide one in the future because of its continuous momentum.

For over a century, things seemed like they were heading in the right direction. With unprecedented growth across domains, new fields emerged, new discoveries were made, and new industrial infrastructure was created that triggered even more expansion.

This linear progress, however, was brought to a halt when the very same innovations and discoveries that were meant to improve our lives wreaked havoc in two major world wars, culminating in the atomic bomb, with millions dead, widespread famine, homelessness, and poverty. We also saw the banality of evil, to quote Hannah Arendt, in action, and how quickly things could get out of hand with the rise of totalitarianism and fascism.

The Alignment Problem

The very tools that drove progress also brought about destruction. Technological breakthroughs alone were not enough to curb evil, disease, or poverty. We were still left with perennial questions for which we had no clear answers: what constitutes a good life, or good governance? How should we handle conflict, or reconcile personal liberty with the public good?

These questions carry moral and legal dimensions that scientific or rational inquiry alone can’t settle. What values guide us? How do we treat each other, allocate public budget, or define what justice, fairness, or transparency mean in practice? Can these principles be passed down to future generations? How do we acquire knowledge, and how should we transmit it? Are values relative to each culture, or is there a universal baseline? What moral frameworks align with our principles, and how do we evaluate whether an action is ethical? Most importantly, what do we actually want?

Finding satisfying answers is quite the feat, and reaching common ground is an equally tough task. Some concepts are inherently fuzzy, and articulating them is far more difficult than formulating mathematical or geometrical definitions. What does fairness mean? How do we instill it in someone, or act fairly across different contexts? If tackling such topics is challenging, codifying these concepts and teaching them to artificial systems is harder still.

This is where the alignment problem in AI comes in. If our goal is to build machines that mimic human intelligence, developing models that learn on their own, interpret the world, and make decisions, what values or norms are we going to teach them?

As Brian Christian writes in his book The Alignment Problem:

“How to prevent such a catastrophic divergence—how to ensure that these models capture our norms and values, understand what we mean or intend, and, above all, do what we want—has emerged as one of the most central and most urgent scientific questions in the field of computer science. It has a name: the alignment problem.”

Relying on the premise that somehow machines will outperform us, morally or otherwise, simply because they crunch through more data may be a recipe for disaster. Brian Christian gives us a glimpse of some of the problems that computer scientists, private companies, and public institutions had to deal with during the development process over the past decades, including algorithmic bias, lack of transparency, and proxy gaming, to name a few.

The alignment problem arises because of the mismatch between our intentions and how we translate them in a way that machines can represent, learn, infer, and optimize. Some of the examples Christian provides include how, based on the data that Google word2vec models were trained on, they made associations such as (Doctor - Man + Woman = Nurse). Or how the COMPAS software led to racially biased classification, labeling Black defendants as high-risk even when they wouldn’t reoffend, and white defendants as low-risk when they were in fact more likely to reoffend; the reason for this is that an algorithm can’t be fair across populations when applying the same definition to different groups whose underlying real-world data is unequal.

Uncertainty and AI Slop

The alignment problem prevails in other areas as well, including self-driving cars, employment processes, and medical diagnosis and treatments. The problem is philosophical in nature. Addressing it requires a multidisciplinary effort between computer scientists, philosophers, economists, anthropologists, lawyers, and subject-matter experts.

The alignment problem isn’t only limited to values and norms, though. It also includes how AI models deal with uncertainty and infer new knowledge based on the data they’re trained on. A simple, practical example you can toy around with is the model Talkie, which was trained on data from the pre-1930s.

It claims that no major wars would take place in the future because:

No, major wars are practically impossible in the future. The Hague Tribunal and other agencies for the settlement of international disputes have made them unthinkable. The nations have established a system of arbitration, and, under present conditions, a refusal to arbitrate would practically amount to an act of war. In addition, modern economic interdependence makes war between great Powers a calamity from which civilisation would shrink. The pecuniary cost of war has become prohibitive, and an empty exchequer is the most potent influence making for peace. Small wars may occur, but great wars are things of the past.

Such a confident assertion is amusing to us because, in hindsight, we know that this is not true. When the universe is inherently uncertain, how can we be sure about what the future will look like? And how can we be confident in the models’ future predictions based on past data?

All these questions, dilemmas, and potential risks merit consideration. Addressing these issues, however, is essential for building more aligned, consistent models. Yet the majority seem quite intractable because of the nature of the inquiries being made, and the fact that AI as a field is relatively new, having been established as a proper field in the 1950s with the publication of Alan Turing’s paper Computing Machinery and Intelligence, where he posed the question of whether or not machines could think.

During this first wave, AI studies were more focused on understanding human consciousness to develop models that were conscious. This approach was eventually met with fierce criticism by philosophers, like Hubert Dreyfus and John Searle, among others, who argued that machines could never be conscious because they would always lack the very substrate that makes us human, and how we interact with the world. Our being is radically different from that of machines.

Years later, when the AI approach shifted toward a pragmatic development that was based on neural networks and machine learning, the AI project seemed more plausible. Nowadays, general perception varies across the spectrum. At one extreme are people who claim that AGI will be achieved in a matter of years. On the other are people who think that AI is a useful tool that will make our lives easier, but doesn’t live up to the hype.

Most recently, some claims from people like Marc Andreessen, Alain de Botton, and Tyler Cowen have gone as far as asserting that AI models are already better than doctors, therapists, financial analysts, etc.

Despite the hype, there’s an ongoing debate about how AI models generate slop, and how this low-quality material has flooded the internet. Substack, for instance, introduced Pangram’s AI detection tools to combat it. There’s an irony somewhere in there, I reckon. AI is peddled as superior to physicians, yet at the same time it also outputs slop. Perhaps this is as human as an AI could get after all.

More Questions, More Uncertainty

Given the scale of the field and the complexity of its risks, an important first step is clarifying what specific problem we are examining, and which framework best fits it. For example, exploring whether AGI can be achieved requires a deeper examination of how we define intelligence, and a closer look at, perhaps, human consciousness. A more adequate framework here would be the philosophy of mind alongside other technical analysis.

If the goal is to understand how new tech will reshape how we interact with the world, and what being human in an AI world means, including what this will entail for creativity, learning, and knowledge acquisition, then maybe philosophers like Heidegger, who wrote an essay on the essence of technology, might be more helpful.

Meanwhile, ethical frameworks and theories of knowledge could help us explore the moral problems arising from the alignment problem. Will we want an AI that prioritizes duty over utilitarianism? Or are we more interested in developing models that reflect our values and most revered character traits, including honesty, generosity, and fairness? How can we “teach” an AI to become more “ethical”? Are we building models in our image as humans, or in the image of the various founders of existing AI labs? Politically, maybe they’re not too different from each other after all? Are AI models destined for war in the same way as nations are? What will the landscape look like if AI does indeed take our jobs?

These distinct angles can surface different problems to tackle. While some of these issues can be analyzed rigorously using existing philosophical literature, in other cases, given the uncharted territory AI is opening up, they might be better interrogated creatively through fiction, which can also serve as a vehicle or a cautionary tale against some potential risks.

We already have brilliant sci-fi novels, films, and series that wrestle with these themes (Golem XIV, Ghost in the Shell, Black Mirror, Ex Machina, etc.). The future is exciting and interesting. It might not be all doom and gloom, but the accelerationists must heed the mistakes of the Enlightenment.

In the meantime, if we are to think about the meaning of it all, and what the point is, we can go all the way back to Socrates: The unexamined life isn’t worth living. It feels like humans have an innate desire to transcend themselves, whether through traditional religions or contemporary techno-faith, often forgetting why they set out on this journey in the first place. The question remains: What does the good life look like to you?

PS: In case you’re interested in reading more about the alignment problem, here’s a more recent article. Also, check out this interesting piece by Rob Hoekman.

View the 2026 Salon Program & Reading List →