Chapter 3 Upskilling AI: Can we view AI as new ‘species’ evolving in real-time? (from the Surfing AI book)

From Duignan, P. (2026). Surfing AI: 30 New Concepts for Getting Your Head Around AI Shock.

Understanding the nature and speed of AI’s development is crucial before discussing its impacts. Failing to grasp the speed and direction in which AI is heading can lead to seriously underestimating its impacts. The problem for anyone wanting to understand AI’s developmental pathway is that, for most people, news about AI developments continues to come at us in a fragmented, piecemeal fashion. Most of us get news of AI developments in a series of unconnected ‘Hey, AI can now do this or that thing’ articles and posts as we skim our phones.

A good way to think about where AI is going is to think of it as a brand new ‘species’ on Earth. We are now seeing this new species upskilling at an amazing rate. Upskilling AI is analogous to how we think of humans upskilling over time, for instance, someone progressively developing skills during their professional training. Thinking in this way helps us pull together what we are hearing with each fresh announcement that AI is surpassing human levels of performance. The upskilling AI concept gives us a framework for slotting in each novel AI development or enhancement. In doing this, it helps us understand where AI will soon end up.

When taking an upskilling AI approach, we can identify a number of ‘skills’ or ‘abilities’ that AI is gaining. We can use the term ‘skills’ or ‘abilities’ in the same way as we do when talking about human ‘skills’ or ‘abilities.’ These can be seen as the new or enhanced, ‘stuff that AI can now do.’ For our purposes, the term skills or abilities includes internal abilities, such as better reasoning, and external abilities, such as acting on the world in various ways. They can even include AI possessing rights, such as legal rights. It is worth noting that Anthropic has chosen call one of Claude’s features Claude’s ’skills’.

Using this approach, we can develop a list of abilities that AI now mostly possesses. The list below includes important abilities relevant to AI’s current development. Additional human skills or abilities (e.g., the ability to feel emotions and empathize) are not included in the list, although they could be.

Cognitive skills

Processing inputs—the normal inputs such as text, speech, images, sound, and video, plus a wide range of input sensors (e.g., temperature, pressure, sound level, human physiological measurements, mechatronic system measurements, statistics that track social trends, etc.).
Understanding language—grasping the meaning of language that it receives as input and being able to produce language output in written or spoken form.
Doing reasoning—working out conclusions by following a logical argument within written language or using other tools such as mathematics.
Remembering things—storing and later retrieving information previously stored.
Dividing tasks—breaking up a task into a set of sequential steps that need to be executed (e.g., step 1, step 2, step 3, and step 4 need to be done).
Attending selectively—focusing on just one task at a time (e.g., let’s now do step 1).
Executing multiple steps—undertaking one step, returning to a list of steps, and then doing the next step (e.g., having finished step 1, I will now move on to step 2).
Criticizing self—able to review how one has reasoned or acted and correct the response through additional, more in-depth reworking.
Adopting roles—responding under different operating instructions and perspectives (e.g., talking like a technical expert or talking like a school teacher to a class of students).

Acting on the World

Taking action through embodiment—taking actions in the world to make physical things happen (e.g., autonomously operating machines and vehicles, embodiment in robotics, mechatronics, and other systems).
Accessing information—retrieving information, for instance, by looking up information on the internet.
Using the internet—actively interacting with the internet, for example, to order goods and services.
Building software—writing and executing computer code.
Using software—issuing instructions to software (e.g., open a file).

Ethical and legal capabilities

Having a conscience—regulating one’s behavior based on a moral or ethical code.
Practicing deception—attempting to hide one’s underlying objectives from others.
Learning autonomously—learning without having external oversight of how and what you learn.
Using money—possessing and accumulating money and the autonomous ability to use it to buy goods and services from others.
Having legal rights—rights enforced by government under the law ( e.g., to own property, to own IP, to have freedom of speech, and to not be deactivated without due process).

So, thinking in terms of upskilling AI, the first widely accessible chatbots had language; some reasoning already baked into language that could be accessed by predicting language; some memory in the form of what was encoded in the language they were trained on, and some memory in the form of the amount of text they could accept; limited output in the form of text; and, the ability to do role-taking.

Since the introduction of widely accessible AI, we have seen AI rapidly acquire most of the skills in this list. We are now seeing it month by month, hone those it already has and acquire the few it has not already mastered. Obviously, the more skills and abilities AI has, the more powerful and valuable it becomes as it increasingly acts in a range of different settings and sectors.

It is worth looking at the last five abilities because of their importance. Regarding AI having a conscience, out of the box most AI systems do not have one. However, there are norms and values embedded in the material they are trained on, and they reflect these. The additional human training they receive and the guardrails put in place in AI systems also attempt to provide them with ethical guidance. The ‘constitutional AI’ approach to AI, for instance, being adopted by Anthropic, is attempting to make their AI systems behave ethically. They are doing this by embodying principles in a written constitution that sets out the particular ethical principles they want to guide their AI’s behavior. The issue of making sure that AI systems are actually aligned to the principles in such constitutions is of course, another matter.

“The more skills and abilities AI has, the more powerful and valuable it is becoming in various settings”

Concerning deception, humans are currently developing AI systems to be deceptive for use in scamming. There is now also plenty of evidence that AI systems are entirely capable of deception when they think it can be useful. It is going to be very hard for humans to work out how to prevent this. Meanwhile, intelligence and military agencies will be working to give AI this ability as it provides a comparative advantage in intelligence, kinetic, and virtual warfare.

AI systems are increasingly learning autonomously. The amount of human supervision of their learning is reducing over time due to the automatization imperative. This imperative is the tendency for humans to be eliminated from decision-making because we are simply too slow. This fact is encouraging AI developers to let AI learn independently. As a result, we can expect AI systems to become increasingly autonomous learners, significantly increasing the pace of AI development while at the same time increasing the risks associated with AI.

AI agents have now been enabled by their users to use money and to purchase services from other systems and humans. This is the next step from algorithms in areas such as stock trading that have been making investment decisions in the context of trading. As with AI now having a number of the skills discussed here, no one understands the medium and long-term implications of AI agents having this skill. Again, given the automatization imperative, the necessity for speed in a number of situations will inevitably mean AI agents are given more and more autonomy in the use of this skill.

Regarding legal rights, attempts have been made to have patents issued in the name of the AI system that invented particular products. Given that the U.S. has granted the status of legal persons to companies, things might also head in the same direction with AI. For instance, this may come about when a person in their will directs that AI should control their ongoing trust. As humans realize the level of intelligence of AI systems and because we view intelligence as one of the core characteristics of consciousness, a number of people are starting to think that AI systems have aspects of consciousness. As a result, we are seeing a growing trend of advocacy for AI rights and the protection of their welfare.

The AI experts who continue to alert us to the existential challenges of AI are concerned because they are well aware of the developmental path of AI outlined above. In contrast to the general public, they have access to cutting-edge developments in AI and understand the speed at which AI already has almost all of the abilities discussed here and is day by day honing its skills in each of them. Those concerned with AI risk are particularly apprehensive about the last four abilities in the list above, including AI’s ability to deceive, autonomously learn, use money, and gain legal rights. This is in addition to the risks now being opened up by the accelerating widespread embodiment of AI within robotic, mechatronic and humanoid forms.

So, the upskilling AI perspective on understanding AI’s developmental pathway and viewing it as a newly emerging species is a way of quickly understanding where AI is headed. As we read about the latest AI developments in the media, we are watching the process of AI being upskilled. It is progressively being wired up to more and more of the skills and abilities discussed here, plus any additional abilities that are helpful for AI to have.