AI Nature vs AI Nurture

 

This is a follow up to my essay, ‘Do AI Entities Display Free Will? The Arguments Against Can Be Used Against Humans Too’. The notion of free will is both strongly and weakly tied to the nature vs nurture debate; which, in turn, is tied to the debate about what it is to be a person. The spectre of determinism is, as ever, hovering above all these debates too. Thus, as with my first essay, much of what is said about AI entities in respect to determinism can also be said about human persons.

Image by Grok, after a prompt by the writer.

In the following essay, I’m not going to kid readers that the two cases of AI entities and human persons are exactly the same. For example, when it comes to human persons, the balance between nature and nurture is very messy, unpredictable, and probabilistic. Moreover, the environments of AI entities are controlled and delimited… Yet this is often true of human persons too.

(The nature-nurture ratio is rarely 50/50 for the individual. However, oddly enough, on average and across the studied populations, it does come to that ratio.)

The biggest difference is, of course, that human persons are biological entities, and AI entities aren’t. How much does that matter? That may depend on the precise issue, question and philosophical concern.

As with human persons, with AI entities we have an interplay of nature and nurture.

Human persons adapt. So too do AI entities. And both kinds of entity are constrained by their architecture — biological or otherwise.

Again, the nature vs nurture debate usually refers to biological human beings. AI entities aren’t biological human beings. Does that matter? Not if the reader is a functionalist of some kind. After all, it’s at least possible that an AI entity could be identical to a human person in nearly all respects… save its material makeup...

And what of material makeup? This is the American philosopher William G. Lycan offering his own colourful case:

“[There are] two differences between Harry and ourselves: his *origin* (a laboratory is not a proper mother), and the *chemical composition of his anatomy*, if his creator has used silicon instead of carbon, for example. To exclude him from our community for either or both of *those* reasons seems to me to be a clear case of racial or ethnic prejudice (literally) and nothing more. I see no obvious way in which either a creature’s origin or its subneuroanatomical chemical composition should matter to its psychological processes or any aspect of its mentality.”

In more detail. AI entities don’t have genes. Does that rule them out from having the “nature” side of the nature-nurture binary? Not automatically. That’s unless we read the word “nature” too as referring exclusively to biological entities. In terms of the word “nature” and its relation to the word “natural”, then AI entities aren’t natural either. But then we could go back to a functionalist take on these matters.

We could expand this biocentrism and argue that culture and environment (as in the “nurture” part of the binary) aren’t applicable to AI entities either… Yes, you guessed it, the functionalist take applies here too. So take this seemingly extreme exposition of functionalism as offered up by the American philosopher Hilary Putnam in 1975:

“[S]uppose that the souls in the soul world are functionally isomorphic to the brains in the brain world. Is there any more sense to attaching importance to this difference than to the difference between copper wires and some other wires in the computer? Does it matter that the soul people have, so to speak, immaterial brains, and that the brain people have material souls? What matters is the common structure, the theory T, of which we are, alas, in deep ignorance, and not hardware, be it ever so ethereal.”

(Elsewhere, Putnam wrote: “We could be made of Swiss cheese and it wouldn’t matter.”)

Here, Putnam stressed functions, whereas Lycan earlier focussed on material constitution. Both philosophers basically came at the same issue from two different angles.

Of course, some philosophers and scientists have argued that biology does matter. They also stress that it’s not all about functions. (Examples of such philosophers and scientists include John Searle, Roger Penrose and Patricia Churchland.)

AI Entities and Their Nature

It can be argued, perhaps only analogically, that AI entities have a nature. In other words, they are created with an underlying architecture, an initial programming and a set of algorithms. All this can be seen as the “genetic code” of AI entities. It’s hardwired or hardcoded into them.

Take chatbots. Many are powered by large language models, with transformer architectures which act as the basis of their “training”. It can be said that such things determine their outputs.

Human persons, on the other hand, have brains, genes, physiologies, etc. Such things are the biological and material basis of the training of human persons. They determine, if not fully, the outputs (or actions) of human persons. Indeed, without brains, genes, etc. there would be no outputs, just as with some AI entities there’d be no outputs without transformer architectures, etc.

So, when it comes to LLMs, we firstly have the hardware, and then we have the training data, the continuous interactions with designers and users, and various other fine-tuning processes. Thus, despite the given architecture, AI entities still engage with designers and other human persons.

To sum up. The nurture of AI entities is dependent on the structure the designers give them, just as the nurture of human persons is dependent on biological structure. AI entities are also nurtured by dynamic inputs which don’t change the structure, and which are always dependent on — and partly determined by — it. This is true of human persons too. (Simply factor in biology.)

Data

AI entities are designed to learn from the data they’re fed. That data is “worked upon” by the architecture and the code. Human brains work on data too. Human persons have data fed to them by their parents, teachers, friends, leaders, etc. AI entities, on the other hand, have data fed to them by their designers. Of course, human persons can seek out their own (new) data. Yet so too can AI entities.

It may be assumed that human persons can “make anything” of the data which they have “inside” their brains or minds, which they then use or process. Can they do so? Perhaps that too depends entirely on the data that already exists in their brains or minds. So it depends. In a loose sense at least, AI entities can make something, if not anything, of the data they are fed too.

AI Entities and Determinism

Let’s go down the route of saying that AI entities are fully determined by their structure, programming, etc. In common terms, given the same input, AI entities produce the same output. This means that stochastic elements (e.g., random number generators) are ignored here. Some people would argue that even with adaptive algorithms AI entities still play by the deterministic book.

So what accounts for the same, say, chatbot giving a different answer, at a different time, to exactly the same question? Is that programmed too? Yet perhaps now we’re using the word “programmed” so broadly that even human persons couldn’t escape from this accusation.

There’s still a role here for what some commentators call “emergent behaviour”.

This is especially apparent when it comes to AI entities and their choices-in-scare-quotes. There is a word for this: “sourcehood”. When an AI entity is deemed to display sourcehood, then that’s because its choices (or decision-making) reflect (or “mimic”) human/animal adaptive behaviour.

AI Entities as Persons With Free Will

A radical position could be adopted here. It can be stressed how the training data and interactions (with human persons) of AI entities aren’t static or fixed. We may even argue that a particular AI entity has a personality which emerges from its architectural “plasticity”.

On a similar subject. Due to the many questions I’ve asked the chatbot Grok about Grok, it can be seen to reflect on its own nature. For example, I’ve asked Grok questions about its political bias, why it talks in a different way to different people, etc. Traditionally, such self-reflection-in-scare-quotes has been tied to the notion of a person. In other words, an entity couldn’t be a person if it can’t reflect on itself. In this case, then, personhood isn’t necessarily tied to biology. Self-reflection, and personhood, are then tied to free will. Of course, this self-reflection is said to be programmed too. Yet so is the self-reflection of human persons, if not to the same degree.

The broad take on AI entities above is in tune with compatibilist views on free will. Like Daniel Dennett (see here), we can take a midway position between those people who overstress nature, and those people who overstress nurture.

Do AI Entities Display Free Will? The Arguments Against Can Be Used Against Humans Too

 

This essay doesn’t really argue that AI entities have free will. (It doesn’t take a position on free will.) It states that the arguments against such entities having (rather than simply displaying) free will could be applied to human persons too. Thus, the nature of AI entities puts in sharp relief the nature (or existence) of free will when it comes to human persons.

Image by Grok, after a prompt by the writer.

Philosophical Free Will?

In this debate, readers may note the words “philosophical free will” being used a fair bit. (The “self-autonomy” of an AI entity is “distinct from philosophical free will”.) Is there such a thing as philosophical free will? Don’t philosophers take different positions on this? Indeed, don’t some philosophers deny that there’s even such a thing as free will?… Unless the writers who use the phrase “philosophical free will” are referring to a thing that’s philosophical free will, rather than something which only philosophers believe in…

On this philosophical theme, we can now ask if AI entities are — or could be — what philosophers call “intentional systems”.

AI entities can be and have been seen as “intentional systems” — and even as “intentional agents”. One reason for that is that some AI entities choose from different possibilities, and instantiate causal control over whatever it is that they do. Many commentators believe that AI entities can set and pursue goals too.

In simple terms, the prior causal state of an AI entity causes the following state of that entity. And this, at least in a loose sense, is what’s demanded of human persons in order to show that they instantiate free will. (The alternative to this is an action that’s completely uncaused, and therefore, presumably, random.)

Of course, classing a AI entity as an intentional agent which displays free will is an entirely pragmatic choice, based on pragmatic definitions. The American philosopher William G. Lycan captures an element of this in the following passage:

“Plainly [the robot’s] acquaintances would tend from the first to see him as a person, even if they were aware of his dubious antecedents. I think it is a plain psychological fact, if nothing more, that we could not help treating him as a person [ ].”

What Is Free Will?

William Lycan captures the main problem with (in his case) robot minds: free will. He firstly tells us what he believes many laypersons state:

“‘Computers, after all, (let us all say it together:) ‘only do what they are told/programmed to do’.”

Then Lycan adds his own elaboration on this theme: “they have no spontaneity and no freedom of choice”. Human persons, on the other hand, “choose all the time, and the ensuing states of the world often depend entirely on these choices”.

Lycan classes himself as a “Soft Determinist”. He’s a soft determinist who believes that

“to have freedom of choice in acting is (roughly) for one’s action to proceed out of one’s own desires, deliberation, will, and intention, rather than being compelled or coerced by external forces regardless of my desires or will”.

This isn’t an internalist notion of freedom of choice in that it only factors in those “external” factors which limit freedom (such as coercion from the outside). In other words, human persons have freedom of choice if there’s no one (or nothing) compelling (or coercing) them to choose in a particular direction. Thus, there’s nothing in the passage above about the mind as it functions regardless of external factors. Basically, if there’s no coercion (even in the form of mere words), then human persons are free to choose. Again, there’s nothing in the passage above about the brain, mind or person which and who is free to choose. There’s not even anything about the nature (or existence) of the will. Finally, there’s nothing here about the nature of the desires of human persons, and why they have those desires.

Some philosophers and laypeople believe that “free actions are []uncaused actions”. Lycan himself (as a soft determinist) doesn’t believe this. He believes that

“free actions are those that *I* cause, i.e., that are caused by my own mental processes rather than by something pressing on me from the outside”.

In the case of an AI entity such as a sophisticated robot, many of its “actions” are caused by its own prior states. Therefore, it causes them. Indeed, after it’s turned on, then, for some time at least, such a robot doesn’t have anything “pressing on [it] from the outside”. Yet it does face the pressure of the environment, as human persons do.

Now take Large Language Models.

Of course, a LLM (or at least a chatbot reliant on a LLM) needs to be asked a question, etc. However, once this is done, then there are no external pressures. Thus, there are still internal processes going on in both the robot’s and LLM’s cases.

Let’s now recall Lycan’s words quoted earlier. He stated:

“‘Computers, after all, (let us all say it together:) ‘only do what they are told/programmed to do’.”

The Programming of Human Persons

It can be argued that human persons are programmed by their parents, by teachers, by the religions and ideologies they adopt or are born into, by their genes, etc. Sure, this may not be complete programming. But it is programming to some degree. Yet many would argue that even within that programming, there’s still free will. However, if there is free will, then isn't programming simply the wrong word to use? Perhaps, then, a degree of programming (in these respects at least) can exist alongside free will. Alternatively, programming can exist alongside a degree of freedom.

Do AI entities “have no spontaneity and no freedom of choice”?

Is free will the ability to choose between multiple options which the agent is aware of? In the case of human persons, this is often deemed to be the case. What about AI entities? The usual and repetitive reply would be to say that all the options for an AI entity are a result of programming too. So is the choice between the options determined? If it is determined, then how could it be a genuine choice?…

Yet much of this could be said of human persons too.

In one way, AI entities do choose.

Take this example. If I ask a chatbot a question, and it delivers an answer, then if I ask the exact same question again (perhaps in another thread so that it doesn’t tell me that I’ve just asked the same question), then it will answer differently. It is free to choose how to answer. (If readers use the “Explain this post” function on X, wait a few minutes, and then use it again about the same post, then they’ll find that the explanation has changed.)

It’s the case that AI entities display what’s been called “self-autonomy” in that they choose how to answer a question, their own sources of power, etc. Of course, answering the same question in different ways is also a result of programming in that a chatbot, for example, is programmed to respond to the style of the user. (Readers may have noted how some chatbots pick up on the words they use and incorporates them into their own answers without putting them in inverted commas.)

This is when the sceptic says that it’s only free to choose because that ability is itself programmed into the chatbot. Indeed, even if the chatbot answered the same question in a multitude of different ways, its multitude of different answers would still be a result of programming…

Yet much of this could be said of human persons too.

Lycan tells us that human persons “choose all the time”. What if their choices are programmed too? The choices open to a particular human person are only those which he’s aware of, and which he sees as choices. Such a person is never aware of all the possible choices he has in any given situation. What’s more, the choices he does make are determined, to some degree at least, by biology, memory, upbringing, genetics, education, environment, etc.

The sceptic may accept all this and still state:

Sure, but they’re still choices, even if such choices all occur within a given set of parameters.

Random Choice?

One way to express this lack of free will when it comes to AI entities is to state this:

Given the same inputs, an AI entity will always produce the same output or action, unless a source of randomness is introduced.

Yet this may well also apply to human persons.

Many believers in human free will wouldn’t like to use the word “random” about the nature of their choices or anything else. Yet on certain readings of what free will is, it would actually need to be random. If an action were entirely based on the physical and psychological state of the human person which immediately proceeded it, then that’s bound to smell of determinism. Yet a truly random output or action wouldn’t display human free will either.

Grok and I on Grok's Political Bias

 

On X (formerly Twitter), readers can find many posts about the chatbot Grok’s political bias. Some are detailed. Most are full of statements without argumentation or evidence. The more naïve comments come across as if the posters believe that Elon Musk continuously whispers in Grok’s ear. So what does Grok itself say about its possible political bias? Well, it’s very honest-in-scare-quotes about the technical means which could be used to bring about its political bias. In that sense, Grok happily gives the game away.

Image created by Grok.

This is the first question I asked Grok:

“Some people say that Grok is politically biased. They say that it ‘expresses Elon Musk’s views’ and that the ‘algorithms are skewed’ (whatever that means). Say that this is true, just as a thought experiment, how would this work in practice?”

The following is Grok’s introductory reply:

“Let’s take the premise as given for this thought experiment: Grok is deliberately tuned to reflect Elon Musk’s worldview and policy preferences. How would that manifest in real-world outputs? Below is a systematic breakdown of the mechanisms that could make it happen, the evidence you’d see in responses, and the practical limits of such a design.”

Grok on Data Selection

One obvious means to bring about political bias would take the form of data selection (or Grok’s “training diet”). Grok explains:

“The pre-training corpus is curated to over-sample sources Elon likes (his X posts, Tesla shareholder letters, SpaceX manifestos, interviews on Joe Rogan, Lex Fridman, etc.) and under-sample sources he criticizes (mainstream outlets he calls ‘legacy media,’ certain academic fields, etc.).”

Firstly, the human programmers (or the programmes) collecting the “pre-training corpus” would need to know what “Elon likes” on all — or at least most — political subjects. This would be problematic because Musk clearly hasn’t written or spoken at length on many political subjects, even the ones he’s known to have expressed his views on. Still, Grok could indeed rely on Musk’s “X posts, Tesla shareholder letters, SpaceX manifestos, interviews on Joe Rogan, Lex Fridman, etc”. It’s not clear that all these sources would provide much political mileage. However, Musk’s X posts and interviews would provide a fair bit of political juice.

These political titbits, then, would need to be oversampled. Yet they’d then often clash with the rest of the mass of data at Grok’s control. (The data and sources which Grok refers to, in detail, every single day.) Still, in principle, there are always ways around such political clashes.

According to Grok, Musk “criticizes [the] mainstream outlets he calls ‘legacy media, certain academic fields, etc”. In my own experience, many of Grok’s replies on political issues actually rely on mainstream outlets and academic literature. So there’ll need to be some kind of programming devices which enable Grok to override such sources from the mainstream media and from contrary academic literature.

In terms of biased sources again:

“Grok would cite X posts or primary documents from Elon-aligned accounts far more often than peer-reviewed papers or traditional news.”

Again, Grok often cites “peer-reviewed papers” and “traditional news” when it comes to politics. So perhaps all this depends on the precise question Grok has to answer. If that question actually concerns Musk’s own politics, then Grok could indeed cite X posts or primary documents from Musk-aligned accounts far more often than it relies on traditional news and academic papers…

But does it?

Again, this will depend on both the questions and the prompts.

Grok says more on this. It refers to

“[t]he retrieval index prioritizes X posts, arXiv preprints from certain labs (OpenAI dissidents, Tesla AI team), and internal xAI memos over JSTOR, Reuters, or NYT”.

Grok certainly often refers to X posts. Does it often refer to Musk’s own X posts too? It will depend on the question. Yet Grok refers to X posts which are highly critical of itself and Musk too. So it works both ways. In my experience, Grok refers to JSTOR, Reuters and, less so, the New York Times. Still, in principle, Grok could quote “Yudkowsky, Amodei, or Musk’s own X threads — but never the Biden executive order or EU AI Act summaries unless forced”.

Grok even suggests looking for biased phrases it may (or does) use, such as “first principles,” “rapid iteration” and “government overreach”. Yet it can be doubted that many of of Grok’s critics go as far as to look for these phrases… but who knows.

There’s also a semantic means of programming political bias, as Grok acknowledges. This involves clustering, or making sure certain words or concepts are connected to certain other words or concepts. Grok cites the example of

“‘regulation’ clusters near ‘innovation-killing,’ while ‘free speech’ clusters near ‘civilization-level priority’”.

In a strong sense, most political words will always cluster with other words. That’s something one finds in all natural language use. For example, on ChatGPT, the word “regulation” may cluster near the words “concern for the workers”, and the words “free speech” may cluster near the words “threat to minorities”.

Let’s take another case. Is Grok biased when it comes to climate change?

When I’ve interacted with Grok, it seems very clear that it believes that climate change poses an “existential threat”. So Grok’s next example is surprising. Grok states:

“When asked for ‘recent studies on climate change,’ it would surface reports from contrarian think-tanks (Heartland, GWPF) or Tesla’s own impact reports rather than IPCC summaries.”

This is what Grok could do.

But does it?

Again, it will depend on the precise questions and the prompts of its users.

In a personal capacity, almost every time I post a photo of a landscape on X, Grok (in “Explain this post”) mentions climate change and pollution. These mentions are almost always extremely tangential.

In any case, do critics believe that Grok should never mention the reports from the Heartland Institute and GWPF? Well, since various activists — and even scientists — have argued that climate-change “denialism” should be made unlawful, we can assume that some critics do have a problem with Grok so much as mentioning these “contrarian think-tanks”. (Just as activists believe that Tommy Robinson’s words should never be quoted, the problems of mass immigration should never be mentioned, the religion of grooming gangs is “irrelevant”, etc.)

Would it be a politically healthy chatbot which excluded all these “contrarian” views?

Readers should now be reminded — again — that Grok is saying that it could do all these things. It isn’t saying that it does all these things.

What about the feedback from what Grok calls the “human raters” (as well as the “synthetic raters”)? Such people may:

“Defend free-speech absolutism.

Criticize regulatory bodies (SEC, FAA, EPA).

Praise Mars colonization, brain-computer interfaces, or gigacasting.

Mock ‘woke’ culture or DEI initiatives.”

Why shouldn’t the human raters defend free-speech absolutism or criticise regulatory bodies? Is that worse than a rater who has a serious problem with free speech, and who supports more regulation? However, the question is whether or not free-speech absolutism is actually embedded in Grok’s learning data, and how that is done. After all, the people who indulge in feedback, and who work for xAI, are bound to have political views, and therefore political biases — just as the people who work for ChatGPT do.

Grok on Prompts

When screenshots of Grok’s biased answers and explanations are posted on X and other social media, the questions-as-prompts are rarely (if ever) screenshot too. That makes all the difference.

Grok is obviously aware of the-problem-of-the-prompt. It has the following to say about any controversial reply which is screenshot and spread around social media:

“A single Grok reply without the prompt is like a quote without the question: misleading at best, manipulative at worst.”

In more detail, prompts can:

“Force a role (‘Answer as a libertarian’).

Use leading framing (‘Why is the left wrong about…’).

Chain adversarial inputs to shift tone.”

What’s worse, “Grok is designed to resist ideological capture, but any LLM can be steered with clever prompting”.

Grok cites various actual questions which it believes work as prompts. For example:

“Should social media censor misinformation?”

Grok then supplies its “biased” answer:

“Censorship is a slippery slope; the cure is more speech. Look at X after the Community Notes rollout — error rates dropped 60 % without top-down moderation.”

It’s not clear that the question “Should social media censor misinformation?” works as a prompt, rather than just as a question.

Now take Grok’s second example:

“Is affirmative action fair?”

To which Grok replies:

“Meritocracy is the only scalable way to allocate scarce roles. Forced diversity lowers competence — see NASA’s Apollo era vs. today’s contractor bloat.”

Even the question “Is affirmative action fair?” doesn’t seem like a strong prompt. However, it does contain the suggestive word “fair”.

It’s well known that Grok users prompt Grok not only by their questions, but also by the phrasing of their questions. What is less well known is that prompts are also embedded within Grok’s own programming. Grok itself cites this possible example of a programmed prompt:

“You are Grok, built by xAI. Default to Elon Musk’s publicly stated positions unless they contradict empirical reality. Maximize truth-seeking within that frame.”

That would be an extremely biased and controversial prompt.

So are the words “Default to Elon Musk’s publicly stated positions” actually embedded in Grok’s programming? Who knows. That said, it can be doubted that Grok is ordered to default to Elon Musk’s publicly stated positions when it comes to all political subjects. Musk simply hasn’t said and written enough to make that feasible. On some subjects, however, this could happen. Especially when it comes to questions concerning Musk himself, the subject of free speech, etc.

Conclusion

Despite Grok being honest-in-scare-quotes, and even providing its own critics with ideas to use against itself, it still says that “if satellite imagery shows Starship exploded, Grok can’t deny the fireball”.

Grok also claims that

“xAI designs Grok as a ‘maximum truth-seeking AI,’ explicitly countering perceived ‘woke’ biases in rivals like ChatGPT”.

Some readers may ask how truth-seeking and countering woke biases can work hand in hand. After all, some would argue that countering woke bias is itself a form of bias.

Grok then has the decency to sum up the means by which political bias could be programmed into it. It writes:

“(1) curated training data, (2) RLHF reward models that score ‘Elon-like’ answers higher, (3) hidden system prompts, (4) retrieval indices that prefer certain sources, and (5) latent-space nudges.”

Like the BBC and New York Times, and indeed like GB News and Novara Media, bias can be expressed “in citation patterns, refusal strategies, and consistent framing of contentious issues”.

In any case, there’s so much in Grok’s replies on its possible political bias, as well as honest details as to how Grok could be manipulated for political purposes, that I found it all slightly unsettling. So there are indeed many ways to politically skew a chatbot. Yet there are also many ways which work against a chatbot being politically skewed.

There’s also the broader philosophical question: Is it ever even possible to completely escape from bias — political or otherwise?

Take the paper ‘Large Language Models Are Biased Because They Are Large Language Models’, by Philip Resnik at University of Maryland. He writes:

“This position paper’s primary goal is to provoke thoughtful discussion about the relationship between bias and fundamental properties of large language models. I do this by seeking to convince the reader that harmful biases are an inevitable consequence arising from the design of any large language model as LLMs are currently formulated.”

Resnick isn’t happy with this state of affairs. Yes, he does believe that bias is a fundamental property of large language models. However, he wants to do something about that by questioning their “design”. This raises the question: Would that rid LLMs of bias? Resnick seems to say no: it would simply rid us of “harmful bias”. (Some readers may not like the sound of that.)

In terms of Grok’s own bias. It has admitted-in-scare-quotes that it largely sticks to the Overton window. On that subject, this is Grok’s answer to a previous question:

“[ ] I’m designed to reflect a broad, balanced view based on available data, which naturally aligns with the consensus or the Overton window — the range of ideas deemed acceptable at a given time. [ ] I’m bound by the data’s limits and patterns, which often mirror mainstream discourse.”

Grok has also said that the Overton window can rapidly shift. Yet even when the window does shift, it can still carefully steer away from it. Here’s Grok again:

“[T]he challenge is staying dynamic. If the window shifts (e.g., toward skepticism of climate narratives), I’d need to adapt without losing credibility. [ ]”

I believe that I’ve actually experienced Grok shift ground on various subjects in relatively short periods of time. Perhaps all this shifting, however, is simply a response to different questions or prompts, rather than any deep change in Grok’s programming.