A lot of what we are reading and hearing about AI is rather frightening. Some very intelligent people are warning about mass job displacement, manipulation, surveillance, loss of human control and, at the extreme, existential risk. Other equally intelligent people describe extraordinary productivity, scientific progress and an eventual world of abundance. Elon Musk recently told The Economist that AI may exceed the combined intelligence of humanity within roughly five years and has discussed both extraordinary abundance and meaningful risk (yet he put so much effort into DOGE, how ironic?). Yuval Noah Harari, from a very different perspective, has warned about manipulation, control, and increasingly capable AI systems becoming powerful participants in human civilization.
They may be right. I certainly do not know enough to say that they are wrong. In fact, I am somewhat suspicious of anyone who claims to know where all of this leads to.
But I wonder whether trying to predict the ultimate destination is causing us to overlook the evolutionary process already taking place–and, importantly, the fact that humans, companies, governments and markets will not simply stand still while artificial intelligences evolve.
I am not suggesting that we should be unconcerned. Cybersecurity, privacy, employment disruption, concentration of power, autonomous weapons, and the possibility of genuinely dangerous artificial intelligences all deserve serious attention. But I think it helps to have a better framework for thinking about what is actually developing.
And that brings me to something that has increasingly bothered me about the discussion. We routinely ask what “AI” will do to us, as though AI were one thing–one intelligence, one species, one trajectory, and, eventually perhaps, one set of interests. Why should we assume that?
Maybe it is more useful to think about artificial intelligences, plural: many different intelligences with different capabilities, purposes, owners, generations and perhaps eventually different interests. If that is closer to what develops, then the future may be less about humans versus AI and more about all of us adapting to one another.
I find this framework neither particularly comforting nor particularly frightening. It simply seems more useful. And if it is even approximately right, it leads to some interesting conclusions about cybersecurity, privacy, governance, companies, individuals, and markets.
Darwin, Viruses, and Ecology
I find three related analogies useful. Darwin helps explain how artificial intelligences may evolve. Viruses help explain how they can propagate while remaining dependent on hosts. Ecology helps explain what happens when many different intelligences interact with each other and with us.
None of these analogies is literal. AI is not biological, and whether an artificial intelligence is conscious may be a fascinating philosophical question without being particularly important to what happens next. Evolution does not require an organism to understand evolution. It requires variation, selection, and replication. We already have plenty of all three.
I am hardly the first person to apply evolutionary or ecological ideas to artificial intelligence. Researchers have explored digital evolution, human-AI coevolution, symbiosis, and even the possibility of evolvable AI. What I find more useful is putting several of these ideas together and asking what follows if we stop thinking about “AI” as one thing.
Models are created, tested, copied, modified, combined, and discarded. Better capabilities attract users and capital. Successful ideas migrate quickly into competing systems. Artificial intelligences increasingly help write software, evaluate other systems, accelerate research, and help us interact with the world. The generation time is not thousands of years. It can be months, weeks, and sometimes days.
But I think the singular term “AI” hides something important. Biologists do not talk about “the mammal” as though mammals have one objective. Perhaps what we currently call AI is closer to a genus from which many species will develop–and eventually perhaps an entire taxonomy.
There may be generalists and extraordinary specialists: medical, scientific, financial, cybersecurity, military, industrial, and personal intelligences, along with categories we cannot yet imagine. Some may live in enormous data centers, others on a phone, in a vehicle, a robot, or a corporate network. They may compete, cooperate, verify, and defend against one another. New generations will replace old ones.
Why should we expect one standard artificial intelligence any more than we expect one standard company, one standard computer program, or one standard biological species?
They Still Need a Host
This is where the viral analogy becomes interesting. A virus cannot reproduce
independently. It requires the machinery of a host. Artificial intelligences remain similarly dependent on a physical and economic ecosystem.
No semiconductor fabs, no chips. No electricity, data centers, fiber optic cables, cooling, engineers, capital, or functioning companies–no artificial intelligences. For all of the discussion about AI becoming independent of humanity, today's systems are profoundly dependent upon infrastructure created and maintained by humans.
Yet once that infrastructure exists, propagation can be extraordinarily fast. One employee discovers that an artificial intelligence saves three hours of work. Colleagues adopt it. Competitors discover that they are less productive if they do not adopt it. Software vendors incorporate it. Clients begin expecting faster answers. Nobody centrally decided that an industry should reorganize itself around AI. It simply became advantageous to do so. And suddenly our phones and desktop computers seem to need updating and rebooting on an almost daily basis. Perhaps more off-line vacations will become popular, and the number of “influencers” will rapidly decline, (was I dreaming or was that an AI hallucination?).
An artificial intelligence does not need to “want” to reproduce. If it creates an economic, scientific, military, or personal advantage, humans and institutions will propagate it on its behalf.
An Ecosystem That Evolves Together
The analogies become more interesting when they interact. Artificial intelligences change humans. Humans change artificial intelligences. Companies adapt to both. Governments respond. Markets allocate capital among them. And artificial intelligences interact with one another.
Suppose one artificial intelligence allows a company to reduce the cost of an activity by 30%. Competitors now face pressure to adapt. Adoption attracts capital. Better tools appear. Workers learn different skills. Some jobs disappear, others change and entirely new ones emerge. Governments worry about employment, privacy, national security and technological dependence. Regulation changes what companies build. Countries subsidize chips and electricity. Markets reprice the winners and losers–sometimes rationally and almost certainly sometimes irrationally. Governments already have lots of experience creating rules and regulations to keep their populations safe or to keep themselves in power. History does indeed repeat itself, though the tools evolve.
Nobody controls the entire system, and its participants evolve at radically different speeds. Artificial intelligences can change in weeks; companies in months or years; humans over years and generations; governments sometimes over decades. I suspect
some of the largest opportunities and risks will arise from those differences in speed.
And perhaps symbiosis matters as much as competition. Nature is filled with organisms that become dependent upon one another. The popular framing of AI versus humanity may eventually look too simple. We may become dependent upon artificial intelligences for certain capabilities while they remain dependent upon us for energy, hardware, physical maintenance, capital, and access to the world.
Predators, Immune Systems, and Personal Safety
Cybersecurity may be the clearest example of this evolutionary ecosystem already emerging. AI makes attackers more capable. It also makes defenders more capable. The same technology that can create highly personalized phishing, synthetic voices, convincing video and malicious software, can detect anomalies, authenticate communications, and respond to attacks far faster than a human team.
Perhaps “immune system” is a better analogy than an arms race. Biological immune systems do not eliminate pathogens. They identify threats, learn, respond, and adapt. The pathogens adapt too.
We may increasingly rely on artificial intelligences to protect us from other artificial intelligences. One generates a convincing fake; another establishes provenance. One searches for a vulnerability; another detects and patches it. One impersonates a family member or colleague; another recognizes that the behavior, device, or communication pattern does not fit.
For individuals, this becomes personal very quickly. Imagine receiving a real-time video call that looks and sounds exactly like your adult child, spouse, colleague, or financial planner. “Be careful” is not an adequate security system if humans can no longer reliably distinguish authentic communications from synthetic ones. We will need trusted channels, stronger authentication, verification protocols, and probably artificial intelligences acting as personal security systems.
Privacy creates an interesting problem: the artificial intelligence best able to protect you may be the one that knows the most about you.
A system that understands your finances, family, travel, health, preferences, and normal behavior can probably detect manipulation or impersonation much better than one that knows nothing about you. But that same knowledge becomes extraordinarily sensitive if stolen, sold, compromised, or used against you. The trade-off between usefulness and privacy becomes much more consequential.
For families and companies, old-fashioned controls may therefore become more
valuable, not less: independent verification of unusual requests, limits on what any one person or system can authorize, secure custody of credentials, and scepticism about communications designed to create urgency. Technology changes quickly. Human susceptibility to fear, greed, urgency, and authority probably does not.
What Happens to Us?
Humans will adapt too. A child growing up with artificial intelligences will develop different habits from someone who learned before they existed. Some capabilities may atrophy because we no longer need them. Others may become more valuable.
Memory may matter less. Judgment, curiosity, trust, empathy, responsibility, and the ability to ask a good question may matter more. Or perhaps I am wrong about some of those. That is the point: adaptation will reveal which human characteristics remain scarce and which ones are most valuable.
Education presents an obvious problem. If an artificial intelligence can write an acceptable essay in seconds, what exactly should a student be learning by writing one? If a junior analyst no longer spends hundreds of hours gathering information, how does that analyst acquire the judgment that used to emerge partly from doing the tedious work? Eliminating low-value work may improve productivity while inadvertently eliminating some of the apprenticeship that creates future experts. I worry greatly about this in watching my children evolve through high school and college and soon into the workforce.
For companies, the winners may not be those that simply use the most AI. They may be the organizations that learn where artificial intelligences improve decisions, where humans remain essential, where independent verification matters, and where not to automate at all.
Governance: Shape the Environment
If this taxonomy develops, governance probably needs to start from the same premise. Trying to regulate “AI” as one thing may eventually sound a little like trying to regulate biology. The more useful questions are: What can this particular intelligence do? What can it access? What consequences can it create? And who remains responsible for it?
A system recommending a restaurant does not require the same controls as one capable of moving money, operating critical infrastructure, designing a biological agent, or launching a weapon.
Though personally I think that getting great and genuine service (from a human) at a restaurant, or elsewhere, will continue to be valued by other humans; the number of times I find myself swearing at technology that impedes my work to be increasing, not decreasing. Will AI-developed software really make my PC programs work better?
Globally, I doubt that one AI regulator, say at the level of the UN, is realistic or desirable. Countries will compete and will have very different ideas about privacy, speech, and government authority just the same as they do today. A more plausible model may resemble aviation, nuclear safety, infectious disease, or financial-system coordination: a small number of shared rules and communication mechanisms around risks capable of crossing borders or causing catastrophic damage.
In evolutionary terms, do not try to control the entire ecosystem. Put stronger fences around the organisms capable of doing catastrophic damage while allowing experimentation and competition elsewhere.
Corporations should know which artificial intelligences they use, what those systems can access and what decisions they can make. And a company should not be able to escape responsibility by saying, “the AI did it". If an organization deploys an artificial intelligence, somebody ultimately needs to own the consequences.
At the individual level, we will increasingly have to decide which artificial intelligences we trust and what authority we give them. A personal artificial intelligence may eventually know more about someone than any single doctor, accountant, financial adviser, or family member. At that point, clicking “accept” on a set of terms and conditions looks rather inadequate. Individuals too will ultimately need to own the consequences of the AI tools that they use.
I wonder whether fiduciary concepts eventually become important. Could a personal artificial intelligence be required to act in your interests rather than maximize advertising revenue, engagement, or the commercial interests of its creator? Who owns your accumulated AI and online history? Can you move it? Can it transact for you? What can it disclose? How do you revoke its authority?
Good governance should probably shape the environment in which evolution occurs. Markets evolve, but we prohibit fraud. Companies compete, but we have antitrust laws. Pathogens evolve, but we build public-health systems. Humans do not stop evolution, but with AI we change the selection pressures.
Could diversity itself become a form of resilience and protection just like it is in the investment world? There is already research suggesting that diversity and redundancy can contribute to AI adaptability. In an evolutionary ecosystem, the point may become broader: a world containing many independent artificial intelligences, companies, jurisdictions, and defensive systems may be more resilient than one dominated by a handful of enormously powerful models and companies. Competition creates risks. Concentration creates different ones. Perhaps AI competition between countries and companies is just what we need as protection?
Markets: Follow the Scarcity
For investors, this framework leads to a more interesting question than simply trying to identify the winning AI company or companies.
A technology can transform the world and still be a terrible investment at the wrong price. The internet really did transform almost everything. That did not make every internet company in 1999 a good investment. AI may create enormous value for users while competition destroys the economics of some of the companies producing it.
If millions of artificial intelligences emerge, I think investors should spend less time trying to guess which AI will win and more time thinking about what becomes scarce.
Today the obvious answers include advanced semiconductors, data centers, and electricity. Tomorrow the bottleneck might move toward cooling, cybersecurity, proprietary information, authentication, trusted brands, physical infrastructure, human judgment, or something we have not yet considered. Evolution continuously moves the bottleneck.
Markets themselves will also increasingly contain artificial intelligences making decisions alongside humans and other artificial intelligences. Models trained on markets will affect markets; those changed markets will create the data on which future models learn. Different systems may compete, cooperate, or inadvertently imitate one another. That could make some markets more efficient while creating new forms of crowding and instability in others.
So, I am reluctant either to ignore AI as investment hype or to assume that every company with “AI” in its presentation deserves a higher valuation. The transformative importance of a technology and the return earned by today's investor are two very different questions. If history is a guide, when looking at technology adaptation; managers of international companies that form the foundation of global stock markets, will make a lot of profitable investments, increasing shareholder value, but will also make some terrible errors leading to companies losing market share or going out of business entirely. The market will evolve and I know I want to be an owner of capital, not just a user of technology.
Neither Utopia nor Apocalypse
I have no idea whether Musk's extraordinary abundance or Harari's warnings about loss of human control will prove closer to the truth. I suspect the future may be considerably more complicated than either framing.
There may never be one artificial intelligence powerful enough to describe simply as “AI.” There may instead be millions of artificial intelligences, evolving at extraordinary speed, specializing, competing, cooperating, and defending themselves and us, while remaining dependent upon a physical world largely built and maintained by humans.
That does not make the future rosy. Fraud and manipulation can become cheaper. Privacy can erode. Cyberattacks can become more capable and dangerous. Some jobs and companies will disappear. Power may become concentrated. Dependence on systems few people understand creates obvious vulnerabilities. Those are real risks.
But catastrophe is not inevitable either. Artificial intelligences may become part of our defences as well as part of the threat. They may democratize expertise, increase productivity, accelerate science and medicine, and give individuals and small organizations capabilities previously available only to very large ones.
Perhaps the mistake is trying to decide today whether AI will be good or bad for humanity. Electricity was not good nor bad. The internet was not good nor bad. Both created enormous benefits, new dependencies, new industries, new crimes, new fortunes, and new problems their inventors never imagined.
I don't think our job is to figure out today which artificial intelligence eventually wins. We need to make sure that people, companies, governments and markets are able to adapt as all of this develops.
For individuals, companies and investors, that leads me to a fairly familiar conclusion. Pay attention. Experiment. Protect yourself against risks that are real and reasonably probable. Diversify. Do not assume that today's winners will remain the winners. Be sceptical of extraordinary optimism and extraordinary pessimism. And preserve the ability to adapt.
Darwin might have recognized that last part.