The Origin of the Machine Species

How AI Mirrors Human Reasoning and Evolves in Digital Ecosystems

2025-06-06

How AI Mirrors Human Reasoning and Evolves in Digital Ecosystems

It was close to midnight when Emma, fifteen tabs open, was defending (with genuine fervour) the best-ever Pokémon color (obviously violet) in a social‑media thread that had long since drifted to the merits of EV over hybrids. She composed paragraph‑length rebuttals, linked to reputable sources, even vibe-coded an interactive compound failure risk graph — yet the opposing side grew only more entrenched. Objectively verifiable facts didn’t seem to matter anymore. Tiny exceptions were blown out of proportion. How could the same people who made excellent points just a few hours ago now had become so …stupid??

Somewhere between the second and third mug of tea, she realized something: this wasn’t about finding truth — it was again just about winning.

POINTS!

But why? Why did they care so much about winning? There were no prizes down in the neon-flashing catacombs. Was the small dopamine kick enough of an incentive? And why the change from objectivity-driven at first to almost manipulative at the end? If it weren’t for that, she would have explained it away with a lack of education.

Emma had unknowingly tapped into a deep insight from cognitive science: human reasoning evolved primarily as a social tool for argumentation and persuasion, not to improve understanding and finding the truth.

Reasoning as a Social Strategy — not a Truth Mechanism

Reasoning* */rē′zə-nĭng/ noun: The process of drawing conclusions, making decisions, or justifying beliefs using evidence or logical steps.

It’s quite a complicated process that often involves combining what we already know to arrive at something new. In our brain, the area where it happens is called the prefrontal cortex and it sits just behind your forehead. “Sits” is probably the wrong word, because it’s neural tissue that burns glucose ten times faster than muscle! It is curious that evolution would create such a power-hungry module: low energy consumption is favored, because you’d never have known when you’re next meal would have come or when you’d have to confront predators. Reasoning sounds very useful in somewhat sophisticated settings, where you have the luxury of discuss and argue, but a large and power-hungry prefrontal cortex appeared in hominins like the Homo erectus, more than one million years ago. That was not a time for chit-chat.

The prevailing view of why the hungry prefrontal cortex and thus sophisticated reasoning appeared is counterintuitive. Cognitive scientists Hugo Mercier and Dan Sperber have argued through their “argumentative theory” that when people explain their reasons, it’s not mainly to find the truth, but to convince others and defend their own ideas. We give reasons to rally support from those who agree with us and to push back against those who don’t (H. Mercier & D. Sperber, The Enigma of Reason; Harvard Univ. Press, 2017). In other words, the machinery for logical elaboration evolved togetherwith social feedback loops that limit how often we crank that machinery. The system is adaptive precisely because it progressively scales its metabolic spend according to social pressure:

  1. Energy frugality A large prefrontal cortex is useful, but only if used sparingly. Initial reasoning starts with quick, low‑effort heuristics (“weak arguments”).
  2. **Escalate on demand **If an ally or foe challenges us, we invest a bit more energy to refine the argument.
  3. *Audience tailoring **We spend the *most metabolic currency only when the stakes —and the listeners— justify it (e.g. courting mates, negotiating alliances, heading to court).

The insight is that we’re wired to reason not to seek the truth or understanding —it just sometimes happens when enough people argue and challenge each other in public.

Machine Reasoning is …not very different?!

We built computers to be unflinching number‑crunchers: add, sort, verify —never improvise. Large language models break that bargain and deliver a truly genuine paradigm shift: They don’t solve for the right answer, they guess the next word from a statistical sea of possibilities. Suddenly a machine that should behave like a calculator can riff like a novelist, bluff like a debater, or misremember like a tired friend. This new type of machines —deterministic metal acting like mercurial mind— surprisingly also reproduce a few unexpected aspects of human reasoning:

  1. DECIDE FIRST, REASON LATER
    Mercier & Sperber note that reasoning is a post-hoc rationalization. That is, spinning arguments to justify gut feelings. Studies have shown that this is oddly more the case about big decisions, such as buying a house or a car. GPT-LLMs exhibit a functional analogue: they generate an answer in the hidden layers, then emit a chain‐of‐thought that retrospectively “explains” it. That chain is helpful for us, but it is not the causal path the network used to arrive at the answer — much like our own verbal justifications.
  2. DEBATE RESCUES ACCURACY
    Put humans around a table and their argumentative specialization becomes a collective problem-solver; weak individual biases cancel out when others critique them. Prompting LLMs with self-consistency (“generate 5 rationales, then vote”) or with adversarial debate between two models yields better answers for the same reason: multiple imperfect reasoners expose each other’s blind spots.
  3. SURFACE PLAUSIBILITY BEATS GROUND TRUTH
    Humans are swayed by eloquent rhetoric, even if it is misleading and LLMs are biased by statistically common but factually wrong continuations. Both systems —biological and cybernetic— need external fact-checking to stay tethered to reality.
  4. ALIGNMENT IS SOCIAL
    Reinforcement Learning from Human Feedback (RLHF) trains models on thousands of judgments from human annotators about which response sounds better. Despite being much slower, that is eerily close to how communities reward or punish human arguments through likes, votes, or status —turning alignment into a decades-long socio-cognitive loop rather than a purely technical fix.

Still different: Emotion & Stakes vs. Breadth & Speed

However, the key divergences between human and AI reasoning still matter deeply. Humans are driven by clear goals, vivid emotions, and tangible stakes in survival; these personal motivators genuinely shape our reasoning. AI, on the other hand, trades visceral urgency for immense breadth. A GPT model can skim centuries of text in its parameters and sweep that landscape for patterns in milliseconds. That wide‑angle lens lets it dive into profound depths as well —zooming from a medieval treatise to the latest protein‑folding paper without breaking stride. In practice, human reasoning feels the heat and machine reasoning feels the horizon.

And we feel the heat also in a different way: just as steam abstracted muscle, self‑supervised gradient descent abstracts a portion of conceptual labour. Day and Night — 365 days a year. However, while most of the discourse is centered around AGI and super-intelligence, a much more significant phenomenon is underway and will shape our collective future: A Cambrian Explosion of Machine Intelligences.

Claude, o3, Magistral, Gemini, Llama, specialised protein‑folding transformers, multi‑agent planners,… a zoo of AI models is roaming at full speed through deployment niches of many kinds: Cloud APIs, on‑device diffusion, browser JS, IoT cameras, self-driving cars, humanoid robots, and more! And constraints like latency budgets, GPU scarcity, licence friction, community goodwill, etc. are leading to selection under resource pressure. Just as Darwin explained 200 years ago in the Origin of the Species. Countless variations will bloom, many will go extinct, and only a comparably small subset will dominate the verticals.

Can we predict who will survive and thrive? Before Darwin, Linnaeus in mid-1700s was laying fundamental groundwork by trying to catalogue every beetle as Europe’s colonies shipped creatures from four continents. He meticulously laid down a taxonomic system and classified thousands of species based on morphological characteristics. Despite the best tools of his time and access to relatively broad information, ultimately Linnaeus could not predict giraffes. Giraffes are clear evolutionary winners in their niche, but if you haven’t ever seen one, you probably wouldn’t believe that heck of a neck! However the blueprint that Linnaeus brought forward in Systema Naturae is still useful to be expanded upon to map the ecosystem rules: data quality, feedback speed, and integration cost.

Responsibilities in Physical, Cultural, and Digital Evolutions

But are most of humans just passive observers in this rapid digital evolution? Trying to frantically map and understand what’s going on? Since AI experts have a hard time following all the progress, it looks like a lost cause. But despite the challenges, it is our calling as the dominant species on the planet. Just as we have responsibilities on physical evolution (maintaining climate, biodiversity,…) and cultural evolution, we also need to steer this digital evolution powered by AI in a way that is aligned with our human values.

The question is: how, exactly?

Alignment with human values must be framed at a very abstract altitude. Principles like be helpful, harmless, honest are not programmable as token‑by‑token rules, but need to be included in the base policy of AI systems. Principles act as a coarse moral compass that orients an unimaginably wide search space. To steer silicon minds that roam across the whole library of Alexandria in a blink, we humans have to climb the abstraction ladder ourselves —shaping shared norms, laws, cultural reflexes, and educational systems that can nudge those probabilistic intellects toward humane ends. Moreover, to be effective, those headline values have to be translated into concrete, context‑aware rules that make sense on the factory floor, in the trading room, and inside the lab notebook. A “harmless” Phase‑I drug trial, a “helpful” warehouse forklift robot, and an “honest” robo‑advisor each demand industry‑specific translations that only practitioners steeped in those physical and legal realities can supply. Because every sector is anchored in real‑world constraints —blood chemistry, load limits, fiduciary law— the last word on safety and ethics cannot rest solely with the creators of AI systems. The steering wheel may sit near the summit of abstraction, but its torque is transmitted downward through layers of regulators, standards bodies, and frontline experts who convert lofty principles into checklists, sign‑offs, and circuit breakers that align with our needs and values. In the same way biologists tailor conservation plans to each organism, we’ll need industry‑specific guardrails for the many species of machine intelligence now erupting —from tiny edge models in a thermostat to globe‑spanning multimodal giants— each thriving in a different regulatory habitat. In other words, the naturally occurring cultural evolution is now called to sprint upward, yet the act of steering remains gloriously distributed across the people who build, govern, and deploy these systems.

The Relevance of Schools

The reason why schools will always be relevant now appears more clearly: our limited human reasoning abilities allow us only to focus on relatively narrow domains, at least at the level needed to effectively contribute to steering AI systems. A glimpse of this future need is already visible today in how GPT-LLMs are being used today: we could do so much with the latest models, but we have choose because our time is limited and tend to favor specific domains.

However, the need for change in schools is as significant as the foreseeable impact of AI on the job market. Many skills are less relevant to be learnt in detail: GPT-LLMs write fantastic essays and can be creative, likely more than many teenagers. Copilots can autocomplete effective and succint code. Midjourney paints beautiful sceneries of yesterday, today, and tomorrow. Veo makes you vibe like Kubrik. Like craftmen of wooden looms are not getting paid in sufficient numbers any more, their know‑how didn’t disappear, it migrated. Some became mechanics who serviced the new power looms, others pivoted to designing Jacquard punch‑card patterns, and a few rose to manage entire textile mills. And along with new jobs, new supply chains and skills to be taught appeared. Automation killed the old task and pushed human ingenuity to think further up the value chain, at a higher level of abstraction.

More specifically, for a new world in which most of knowledge tasks are being killed, 3 types of skills are becoming key:

Some schools teach them already today, but mostly as part of a subject or as nice-to-have soft skills. The root cause of this is of course that hard skills are more easily quantifiable and gradable. Market demand isn’t very strong yet. But the trajectory is clear and when the new jobs will emerge, the curricula will change. Any school district that is able to time this trend appropriately will create significant value for their region.

Some topics you studied in school will always be cornerstones of humanity. Take Darwin’s Origin of Species: after cataloguing countless adaptations, he closes with a vision of “endless forms most beautiful.” In our own age of automated argumentation, schools are where the next generation first meets those new “forms” — LLMs that code, debate, and design. Classrooms become mini‑Galápagos labs, letting students observe, tweak, and steer their silicon companions. Just as Darwin’s work demanded a refined taxonomy, so today we must collectively craft a more precise moral taxonomy —helpful, harmless, honest— and teach our kids how those abstract categories play out in pharmacology, finance, or freight logistics. Ethical responsibility still attaches to human persuaders, but accountability now diffuses across all who develop, deploy, and regulate machine intelligence. Steering cannot be left to AI makers alone; it flows through educators, domain experts, standards bodies, and citizens who vote on the guardrails.

No matter what, some topics you studied in school will always be cornerstones of humanity. Take Darwin’s Origin of Species: after cataloguing countless adaptations, he closes with a vision of “endless forms most beautiful.” AI-natives will grow up with systems that autonomously code, debate, design, and so much more. In their schools, they will need to learn how to steer these systems. Classrooms will become mini‑Galápagos labs, letting students observe, tweak, and steer their silicon companions. Just as Darwin’s work demanded a refined taxonomy, so today we must collectively craft a more precise moral taxonomy -helpful, harmless, honest- and teach our kids how those abstract categories play out in pharmacology, finance, or freight logistics. Ethical responsibility still attaches to human persuaders, but accountability now diffuses across all who develop, deploy, and regulate machine intelligence. Steering cannot be left to AI makers alone; it flows through educators, domain experts, standards bodies, and citizens who vote on the guardrails.

And for as long as there will be mysteries in the Universe, there will always be need for reasoning, arguing and understanding our way to the best human society as possible.