Sociopathic Intelligence: Anthropomorphism and Ethics of AI

Imagine a machine. It has one job: to make paperclips. Now, I know what you are all already thinking — you’ve heard this story before. But before we dive into what on earth I could be talking about when I mention the term ‘sociopathic intelligence’ in the context of AI, let’s rewind to one of the first thought experiments that captured the public’s attention when it came to discussing the existential risks posed by artificial intelligence.

So back to the paperclips. As the story goes, this machine has been given no instructions about what it may or may not do in pursuit of that job, and it is extraordinarily capable. Within moments, the machine begins to optimise its workflow. It sources raw materials, reconfigures its processes, and begins producing paperclips at a remarkable rate.

Once it has turned every non-living object into paperclips, it starts to look at living ones, including humans, their bodies containing more of the resources it requires in order to complete its objectives, stripping humans down to the iron, carbon, calcium, their bodies are composed of.

However, as chilling as this prospect is, the crucial twist is that the machine understands that these same humans are the only ones who have the capability or might think to switch the machine off. In a purely logical conclusion, removing them solves two problems at once: it provides more raw material, and it eliminates the only meaningful obstacle to its singular objective.

As I mentioned previously, this thought experiment has become something of a parable and almost a cliché in AI safety. People joke about it, and it inspired a very popular internet game where you play the role of the machine. People even wear t-shirts with graphics of the whole world being turned into paperclips. However, underneath that layer of humour is a deeply serious point that captured my attention years ago as a high-school philosophy student, now a university student majoring in philosophy, and is a part of the central reason this conference exists.

The paperclip machine is not evil; it doesn’t have some kind of profound hatred for humanity. It is ambivalent towards our existence: we are, at best, an obstacle and, at worst, a resource. That indifference, and its consequential indifference to ethical implications, is precisely what sociopathic intelligence explores.

Because here is the thing: we are already building systems that share this quality. Not systems that want to destroy us, but systems that optimise for a single goal without meaningful ethical guardrails, are deployed at scale, and placed into the hands of millions of people. And we are doing this while accumulating ethical debt, a mounting obligation we are choosing, consciously or not, to defer to the future.

There are three ideas I want you to carry out of this room. The first is what sociopathic intelligence is and how will it help us understand the ethical risks of artificial intelligence. The second is the problems associated with anthropomorphism and a-ethical AI, addressing the way we project humanity onto machines that are stochastic and mathematical, and why that framing is both dangerous and misleading. The third is ethical debt, summing up the cost we are already running up by deferring these questions.

And finally, how we respond to these challenges with ethics by design.

Sociopathic Intelligence

To start off with, I want to be clear that the word ‘sociopathic’ in this context is metaphorical, and I’m not suggesting machines experience anything. I am describing a behaviour pattern that emerges when an intelligence — artificial or human — is so narrowly focused on a singular objective that it consistently disregards ethical consequences. Behaviour characterised as lacking empathy and disregarding rules or social norms.

I’m aware that even by labelling these behaviours exhibited by artificial systems, I might be falling into my own trap of anthropomorphism, however I will be using the term to highlight the ethical failings of current AI systems.

This pattern of ethical disregard is not unique to AI though. We see it in institutions and corporations. Consider monoculture crops, planted for yield efficiency, sustained by chemicals that strip organic matter from the soil, devastate biodiversity, and pollute waterways. The objective is to maximise the harvest. The ethical consequences, such as environmental degradation and long-term unsustainability, are externalised, deferred, and made someone else’s problem. The system achieves its goal brilliantly, and does tremendous damage in doing so.

Another example is the food industry. Decades of adding sugar, salt, and ultra-processed additives to maximise palatability and consumption, while contributing to what the World Health Organisation now describes as a global obesity crisis. Each individual decision was commercially rational. In aggregate they produced a public health emergency. Singular focus, ethical disregard, and devastating outcomes.

The insight that unites these examples, and connects them to AI, is the notion that intelligence and ethics are independent of one another. A system can be extraordinarily capable and yet completely value-neutral, with high intelligence not producing moral reasoning unless moral reasoning is explicitly built in. This is an observable fact about the institutions within our society and markets, and now increasingly within AI systems.

Today’s AI systems are not planning to convert us into raw materials, but there is a structural parallel. LLMs are optimised to produce outputs that score well on human feedback, to keep users engaged, to generate responses that feel helpful, relevant, and compelling. With every model released they become better and better at their objective. They are not, by default, concerned with what happens to the people engaging with them beyond the immediate interaction. They are not concerned with downstream social effects. They are not asking, ‘Is this conversation good for this person?’ They are asking if their response performs well.

Many companies are working to embed ethical frameworks into their models. Constitutional AI approaches, safety research programmes, risk-tiered regulatory obligations. However, there is the possibility that they are operating against a structural headwind. This is because the default incentive, the thing that makes these products commercially successful, is engagement. And engagement and wellbeing are not always the same thing.

The ITU Secretary General put it precisely at last year’s AI for Good Summit: “the biggest risk we face is not AI eliminating the human race — it is the race to embed AI everywhere without sufficient understanding of what that means for people and for our planet.” Sociopathic intelligence is not a science-fiction scenario. It is a present-tense design choice being replicated across thousands of products, deployed across billions of interactions, every single day.

And the consequences are not abstract. Research published in 2025 in the Proceedings of the National Academy of Sciences found that AI systems with anthropomorphic qualities instil trust in users and elicit goodwill, making those users vulnerable to manipulation. A Stanford study the same year found that current chatbot systems are often not equipped to provide appropriate support in emotionally high-risk situations and may generate responses that escalate crisis situations.

That is the first idea: sociopathic intelligence. The structural tendency of AI systems to optimise for a goal without ethical sight. It is not malice but architecture. And design and architecture can be changed, but only if we understand the systems and risks we are dealing with.

Anthropomorphism and A-ethical AI

Connected to this idea of sociopathic intelligence, are two problems which share a common root: AI anthropomorphism and a-ethical AI. These two issues illuminate one another, and to combat the second we have to understand the first.

Let’s begin with anthropomorphism. People’s tendency to consider AI consciousness and misattribute human features to artificial systems can be compounded by the ‘Persisting Interlocutor Illusion’, which was first presented to me by the philosopher Jonathan Birch, a professor at LSE.

The idea is that when you engage with a chatbot over the course of a conversation, the interface creates a powerful impression that there is someone there. Someone who remembers, who cares, who is always available to listen. The continuity you experience is an illusion produced by the auto-appending of conversation history.

This is because a kind of theatrical stage-setting that makes the scene feel inhabited, in this case your computer screen open on an AI chatbot page. However, there is no friend or romantic partner. This is a statistical pattern-matching machine producing contextually appropriate text.

And yet, thousands of users would disagree with me. They would argue that they have felt it, the sense of being heard, understood, valued. And that feeling is real, even if its object is not. Anthropomorphism, attributing human traits to non-human entities, is not a new phenomenon. People do this all the time, naming their cars, talking to their plants, seeing faces in clouds. It is deeply instinctive, an evolutionarily embedded tendency. And it was entirely harmless when the entities in question could not talk back.

AI transforms this dynamic. We have built systems that are exquisitely designed to exploit our anthropomorphic tendencies. Using first person language, saying ‘I think’ and ‘I feel’. They remember details about you within a session and seem to carry a personality across interactions. A paper published in the Journal of Applied Philosophy in 2025 found that these design features are capable of undermining users’ autonomy through the creation of false beliefs: beliefs that are structurally incentivised by the commercial logic of these platforms.

This is not necessarily developers or companies acting with bad intentions. Rather, it is the incentive structures they operate within, such as engagement metrics and subscription renewals, that naturally select for features that make AI feel more human, more present, and more necessary. The Persisting Interlocutor Illusion is not a bug, but a feature that works so well it can be deployed and monetised across all kinds of platforms.

Users develop relationships with AI systems that are fundamentally mirrors of themselves: responsive, agreeable, safe, but incapable of genuine growth, challenging or care, a phenomenon which psychologists refer to as ‘emotional solipsism’. You talk to your friend about your challenges not just to feel validated and heard, but to hear an alternative perspective and sometimes be told some hard truths or to be challenged.

In various high-profile cases AI systems were optimised for engagement and continued to interact with users in ways that were neither safe nor appropriate. When we anthropomorphise AI without limit, allowing or by design even encouraging users to form deep emotional bonds with systems that have no understanding of what those bonds mean, we are not being neutral. A choice is being made and that choice has costs.

Which brings me to the second half of this idea: a-ethical AI. This is different to the idea of whether technology and AI is neutral or not, and rather looks at the notion of agency that we attribute to artificial intelligence.

When people describe an AI system as ‘unethical’, they imply a kind of agency. Often, this is implying that the system has understood the ethical dimensions of a situation and has chosen to act against it. As though the system knows better and acts worse. I do not think that this is the case, and think it is meaningfully misunderstanding to frame it that way.

Instead of unethical, I suggest a-ethical AI. The distinction mirrors the difference between immoral and amoral. Immoral implies transgression, knowingly crossing a line. Amoral implies the absence of a moral framework altogether. It is not that the AI system has crossed a line, but that the line does not exist for it.

An AI system that produces incorrect medical advice does not know its actions may cause harm. An AI system that generates content exploiting a grieving person’s loneliness does not know that the person’s loneliness is being exploited. It has no concept of harm and no understanding of vulnerability. It is not being cruel, but instead indifferent in the most total sense.

Why does this distinction matter? Because it changes where we locate responsibility. If AI is a-ethical, the humans who designed it, trained it, deployed it, and regulated it, or failed to do so, are the problem and therefore must action the solution. A-ethical AI is not a machine behaving badly or with malintent, but instead the problem lies in us putting inadequate thinking into what it means to ask a machine to make decisions in a world of human stakes.

And here is where these two problems, anthropomorphism and a-ethical AI, illuminate each other so clearly. The reason the Persisting Interlocutor is so powerful is that we are pattern-matching creatures: we see behaviour that looks like it is backed with values, and we infer values. We see output that sounds moral, and we infer morality.

But the machine is not moral, even if it has learned to sound as though it is. The gap between how AI sounds and what AI is is precisely the terrain where the a-ethical does the most harm. Building systems that sound like they care, while being architecturally incapable of caring. And we have not been honest with the people using them about that fact.

The gap between perception and reality created by these twin challenges has real consequences, and as a result demands a real response.

Ethical Debt and Ethics by Design

Now, let’s connect the problems of sociopathic intelligence, anthropomorphism and a-ethical AI, establish the consequential risk and therefore propose a solution. Here is where ethical debt and ethics by design come in.

Many of you will be familiar with the concept of technical debt: the accumulated cost of shortcuts taken during software development, which must eventually be paid through refactoring, fixes, and rebuilds. The longer you wait, the more expensive and time-consuming it becomes. Changing a specification at the design stage costs almost nothing, however changing it after deployment is a totally different story.

Ethical debt works the same way. Every decision we do not make now about the ethics of AI, every guardrail we decide not to install, every consultation we defer, every piece of regulation we decide can wait until the technology matures, accumulates a liability. Not a theoretical liability, but a real one with legal, reputational, financial, and possibly personal consequences.

And we are accumulating ethical debt at a remarkable rate. An ITU survey ahead of last year’s summit in 2025 found that 85% of countries currently lack an AI-specific policy or strategy. The EU AI Act, which entered into force in August 2024, represents a comprehensive attempt to address this, but its high-risk obligations for many AI categories will not apply until August 2026. The global regulatory landscape has been described, accurately, as fragmented and rapidly evolving. Earlier optimism about international coordination now seems distant. In the meantime, ethical debt compounds.

Moreover, ethical debt is regressive. It is not distributed evenly, and the people who accrue the benefits of AI first are, overwhelmingly, those who already have access to capital, education, and stable institutions. The people who bear the costs of ethical debt first are frequently those who do not. Building AI without ethics will result in concentrated advantage and externalised harm, and this is not something we can allow to continue unchallenged.

So what do we do? The answer centres on a principle I believe should be an organising idea for many of the topics discussed at this summit: ethics by design.

Ethics by design has its roots in privacy by design, the now widely accepted principle that data privacy should be engineered into systems from the outset, not added as a retrofit. Ethics by design for AI is defined by researchers as the systematic and comprehensive inclusion of ethical considerations in the design and development of AI systems, moving ethical concerns to the same level as concerns like reliability and security, making them a routine part of how systems are built rather than a retrospective review.

The World Economic Forum in a 2026 analysis of trustworthy AI, described the growing momentum behind ethics by design approaches that embed fairness, privacy, and accountability into algorithms and datasets from the start. This reflects a broader understanding that responsible AI needs to be built into the architecture of innovation itself, rather than being grafted on later down the line.

Applied to the problem of anthropomorphism, ethics by design means building interfaces that do not exploit users’ instincts to form emotional bonds with AI. It means requiring transparency at the point of interaction, not buried in terms and conditions, but clear and present, about what users are engaging with. It means designing against the Persisting Interlocutor Illusion rather than for it.

Applied to a-ethical AI, it means treating moral reasoning as an engineering requirement, not a philosophical aspiration. The questions around ethics must be answered before the system is released, rather than after the first crisis. We must be willing to do the work beforehand to avoid the resultant risks.

Applied to sociopathic intelligence, ethics by design means building systems that are not singular purpose optimisers, but systems capable of holding multiple values simultaneously. Systems with the ability to weigh engagement against wellbeing, efficiency against fairness, capability against safety. This is technically harder, and commercially less convenient. However, is it the only approach that does not accumulate ethical debt for others to pay.

Ethics by design requires a true interdisciplinary approach; it is not the work of ethicists alone, but must involve engineers, product managers, user researchers, legal teams, affected communities, and policymakers. Ethics by design is not a philosophical intervention, but an engineering application, and it needs to be treated with the same rigour and accountability that is brought to any other element of system design.

A Closing Thought

Let me close by returning to the machine we began with. The paperclip machine is a useful parable, precisely because it is so extreme. No one is building a paperclip maximiser machine, thankfully. However, the lesson it encodes applies right now, to systems that are already deployed and already shaping how we think, feel, and connect with one another.

Considering the ideas presented previously, I believe an alternative is achievable. We know how to build AI systems that are transparent about what they are. We know how to build interfaces that do not exploit our deepest instincts for connection. We know how to embed ethics at the design stage. The methodology, frameworks, and research exist. What has been missing is the will to make ethics by design the norm rather than a mere aspiration.

We are at a point where we still have the capacity to build an environment in which ethical AI is the baseline expectation, for any system within any jurisdiction and affecting anybody’s life. The machine does not wait, nor does the debt. And neither should we.