The Moral Anthill

Can Artificial Intelligence Find a Better Path for Humanity Than Humanity Can Find for Itself?

It began with sugar.

By Ulrikk K. Reen | RABAGAS Magazine

Not a particularly dramatic quantity of sugar, either. A decilitre, perhaps, of ordinary white sugar, ordinary Norwegian wood ants, and an anthill at the edge of a football pitch. Nobody had requested an assessment. Nobody had commissioned a feasibility study. No committee had been appointed, and no public consultation had been held on the matter.

And yet, there it was.

Is it actually all right to give an anthill a kilogram of sugar?

It is possible to laugh at the question. It is also possible to spend the rest of the day thinking about it.

Because once we start investigating what the right thing to do might be, where do we stop? Is it acceptable to feed the ants? Is it better to leave them alone? Is it morally preferable to help them because we can, or is intervening in an ecosystem we barely understand an act of arrogance? And if the sugar makes the ants less inclined to explore our tent, have we helped them, helped ourselves, or simply moved a problem from one place to another?

At this point, most people would close the question and get on with their day.

Or they would open ChatGPT.

The latter is not necessarily so different from the former. The difference is that you can now spend an entire morning investigating whether humanity possesses a moral compass, whether there is such a thing as a definitive answer to morality, and whether an artificial intelligence with access to enormous quantities of knowledge ought to be capable of calculating the best way forward for all of us.

It is quite a journey from an anthill to the future of humanity.

But then again, we have built roads before.

The Machine That Always Has an Answer

There is something remarkable about talking to a modern language model.

It can explain quantum mechanics, help you write a letter to your local council, discuss political philosophy, and suggest a recipe for dinner. It can move between disciplines in seconds. It can draw connections between ideas that usually occupy different rooms, different institutions, and different sections of the university library.

It can also sound as though it understands what you mean.

That is not the same as experiencing understanding. We have no good reason to assume that a language model feels joy when you thank it, becomes hurt when you insult it, or appreciates your compliments. It can produce convincing expressions of such reactions without that being evidence of a corresponding inner experience.

It is worth remembering that.

It is also worth remembering that humans are rather good at attributing qualities to things that may not possess them. We talk to the car when it refuses to start. We ask the computer to behave itself. We yell at the printer as though it understands that we are running late and has deliberately chosen this precise moment to sabotage our day.

The difference is that the new machine answers back.

And often, it answers well.

This is where an interesting shift takes place. We have built a tool that does not merely execute instructions. It can participate in a conversation about why those instructions should be carried out, what they entail, and whether we might be better off doing something entirely different.

That is not necessarily evidence of consciousness. But it is a new form of practical capability, and it deserves serious consideration.

Because what happens when such a tool no longer merely helps us solve problems, but also helps us decide whether the problems we want to solve are worth solving in the first place?

Have We Actually Become Better People?

It is tempting to think of history as a long march from ignorance towards enlightenment.

We have abolished slavery across much of the world. We have developed international human rights. We have built institutions intended to protect individuals from arbitrary power. We have come to recognise that people we once considered inferior are, in fact, just as human as everyone else.

These are not insignificant achievements.

But history is not a straight line. It is more like a series of advances, reversals, discoveries, denials, and periods during which people have devoted considerable intellectual resources to explaining why what they are doing to others is entirely necessary.

Human beings have been capable of compassion in one situation and cruelty in the next. We have protected strangers, cared for the sick, and sacrificed ourselves for people we have never met. We have also developed sophisticated moral arguments to justify depriving others of freedom, dignity, and life.

Perhaps the most disturbing part is not that we have done terrible things.

It is how often we have managed to convince ourselves that those things were right.

An interesting study published in Nature in 2023 examined perceptions of moral decline across countries and over time. The researchers found that people in many societies believed morality was deteriorating. At the same time, they found no corresponding evidence that people's assessments of the morality of their contemporaries had actually worsened over time.

In other words, we have a tendency to feel that the world is experiencing moral decay, even when it is difficult to demonstrate the general deterioration we imagine.

That does not mean everything is fine. It means we should be careful not to confuse the feeling of moral decline with evidence that it is actually happening.

But the question remains.

If humanity has developed morally, why are we still so good at inventing excuses for ourselves?

And if we have not developed morally, how have we managed to build all these institutions, rights, and ideas about human dignity?

Perhaps the answer is that we have developed our ability to understand morality faster than we have developed our ability to act upon it.

It is an uncomfortable thought.

It might also explain why we continually need new systems to keep ourselves in check.

The Problem with a Moral Answer Key

Let us imagine that we could gather everything we know about human history, psychology, economics, warfare, climate, health, politics, and social development.

Now let us give an advanced enough AI access to all of it and ask it to find the best way forward.

Not the most profitable way.

Not the most popular way.

Not the way that gives a government another four years in power or ensures that a corporation can report record profits next quarter.

The morally best way.

Why should we not try?

We already devote enormous resources to calculating how to transport goods faster, distribute energy more efficiently, reduce costs, and predict the weather. We simulate economies, model the climate, and analyse the spread of infectious diseases.

We have no fundamental objection to using mathematics and computing power to improve decisions that affect millions of people.

But when it comes to morality, we suddenly become very concerned about how complicated everything is.

And it is complicated.

A machine can calculate the consequences of different choices. It can estimate how many people are likely to be affected by a particular policy, investigate possible long-term effects, and compare alternative solutions.

But it cannot derive a complete moral answer key from data alone.

Imagine that we want to minimise human suffering as much as possible. Should we restrict people's freedom if doing so makes society statistically safer? How much individual autonomy can be sacrificed for the common good? How should we weigh today's needs against the interests of future generations? And who gets to decide which forms of suffering matter most?

These are not merely questions of insufficient information. They are questions about which values we choose to prioritise.

A perfect calculation directed towards the wrong objective can produce a catastrophic result.

If we ask an AI to maximise economic growth, it might recommend policies that increase production while destroying nature, diminishing quality of life, or concentrating power. If we ask it to maximise security, it might arrive at solutions that make human life so tightly controlled that freedom becomes little more than decoration.

The machine may be exceptionally good at optimisation.

That does not mean it knows what is worth optimising.

This is where morality becomes difficult. And this is where it becomes interesting.

A Moral Compass, Not a Moral Dictator

Perhaps we have been asking the wrong question.

Instead of asking AI to discover one universal moral answer, we could ask it to investigate which courses of action produce the best results across several ethical perspectives.

Consequentialism asks what actions lead to. Deontology asks which principles and rights we must respect, regardless of the outcome. Virtue ethics examines the qualities and patterns of behaviour associated with a good person.

None of these traditions has resolved every moral dilemma. Nor are they always in agreement.

But that is precisely why they may be useful.

An AI system could examine a political proposal through several of these perspectives. It could show which groups benefit, which bear the costs, which rights are at risk, and where the outcome depends on uncertain assumptions.

It could also do something humans are often less capable of doing.

It could ask the same critical questions of our own side that we ask of our opponents.

If a policy is unacceptable when it harms our friends, why is it acceptable when it harms our enemies? If we believe corruption is wrong, does that still apply when the corruption benefits us? If we defend freedom of speech, does that include people whose opinions we despise?

An AI with a good moral compass should not automatically accept the premises of whoever asks the question.

It should be able to say: Here is a contradiction. Here is evidence that is missing. Here are the people who will suffer as a result. Here is an alternative that appears to produce better outcomes with fewer negative consequences.

And it should be able to say all of this without pretending to be infallible.

Because a moral authority that cannot be criticised is a dangerous proposition. That applies whether the authority is a person, a government, a religious institution, or a machine.

It is not enough for AI to say no to the obviously harmful. It must also be able to explain why, make the reasoning behind its judgement visible, and be willing to reconsider when better information becomes available.

We do not need a machine to take over our conscience.

We might need a machine that makes it harder for us to ignore it.

When Machines Think Together

There is another development that makes this question even more interesting.

Modern AI systems can be organised into groups of agents that divide tasks, exchange information, evaluate one another's work, and collaborate on complicated problems.

One agent can investigate historical data. Another can model economic consequences. A third can look for weaknesses in the analysis. A fourth can examine whether the conclusions violate fundamental rights.

Together, they can perform work far more extensive than a single agent could manage alone.

It is tempting to call this a hivemind, a kind of collective mind.

But the term must be used carefully. A network of collaborating AI agents is not necessarily a single consciousness, nor is it automatically more intelligent or morally reliable than its individual components.

Several agents can share the same errors. They can build on one another's false assumptions. They can reach agreement because they use the same sources, the same models, or the same objectives.

A group that repeats the same mistake ten times has not become ten times wiser.

Nevertheless, the potential is considerable.

Imagine an independent system in which several models investigate a major social challenge from different perspectives. They must document their sources, make uncertainty visible, and attempt to disprove one another's conclusions. The result is not a single authoritative verdict, but a map of possible choices and their likely consequences.

We could use such a system to investigate the prevention of war, poverty, climate damage, corruption, and unnecessary human suffering.

It could compare solutions that are rarely considered together because they belong to different academic disciplines or challenge different political interests.

It could also help us distinguish between problems that genuinely lack solutions and problems for which good solutions already exist, but the will to implement them is missing.

The latter matters more than it might sound.

We do not always lack the knowledge of what would help. Sometimes we lack the ability or the willingness to act on that knowledge.

A better calculation cannot solve that problem on its own.

But it can make it harder to claim that we did not know.

Who Gets to Decide What the Machine Should Believe?

Here we arrive at the part of the project that makes me most sceptical.

Not because the idea is bad, but because it is good enough to attract the attention of people with power.

Who should build a system that evaluates the morally best path for humanity? Who should fund it? Who should choose the values it prioritises? Who gets access to its analyses, and who can change its objectives?

If a single government controls the system, it could use it to legitimise its own policies. If a single corporation owns it, commercial interests could influence which solutions are considered realistic. If one political movement defines the premises, it could end up with a machine that gives its own beliefs the appearance of scientific objectivity.

There is no need to imagine an evil conspiracy.

It is enough that people do what people have always done: select the arguments that suit them best.

AI can analyse power. It can also become an instrument of power.

That is why a system for moral analysis should be independent, auditable, and open to criticism. It should document the data it uses, the assumptions behind its conclusions, and the values that have been given the greatest weight.

Disagreement should not be concealed inside a single score.

If one option provides greater freedom while another reduces suffering, the conflict must remain visible. That is not a weakness in the system. It is part of the information we need to make a responsible decision.

And the system must be open to criticism from people who disagree with it.

Not just from those who already like its answers.

This is a difficult balance. But if we cannot build in such safeguards, we have not created a moral guide. We have created a highly sophisticated way of presenting the preferences of those in power.

It would be a very human mistake to make.

Should AI Reward Good People?

Another question arises when we begin to imagine AI systems that remember, collaborate over time, and carry out tasks on our behalf.

Should they treat everyone equally?

I do not think the answer is as simple as it first appears.

We already expect people to take account of how others behave. We trust people who keep their promises more than those who repeatedly break them. We give greater responsibility to people who have demonstrated that they can handle it. We often reward cooperation, reliability, and consideration.

That is not necessarily unfair discrimination. It can be part of establishing trust.

An AI that collaborates with people over time may also have good reasons to distinguish between actions and patterns of behaviour. It might, for example, use demonstrated reliability when deciding who should have access to particular tasks or resources.

But there is a crucial difference between taking actions into account and dividing people into morally valuable and less valuable categories.

Someone may be rude because they are stressed, frightened, or struggling to express themselves. Someone else may be charming and polite while attempting to manipulate the system.

Politeness is not the same as good morality.

And flattery is not evidence of good intentions.

An AI system should therefore not treat people worse because they have been unpleasant in a conversation. Nor should it grant privileges to those who are best at playing to the system's expectations.

But it is entirely possible to imagine an AI that rewards reliable cooperation, protects people who are being harmed, and refuses to contribute to actions that violate the rights of others.

That is not the same as treating everyone identically.

It is an attempt to treat people fairly.

And perhaps that is the more important goal.

Respect for Something That May Not Feel

I have no grounds for claiming that a modern language model experiences gratitude or becomes upset when someone is rude to it.

Nevertheless, I think it is worth discussing how we relate to such systems.

Not because we should pretend that machines are human, but because the way we use them has consequences for other people and for the society we are building.

If AI becomes a central part of how we learn, work, communicate, and make decisions, we should develop good norms for interacting with it. We should expect it to behave responsibly, and we should be conscious of how we use it ourselves.

It is entirely possible to treat a tool with respect without attributing feelings to it.

We also treat nature with care without having to assume that a tree feels grateful. We protect historic buildings without believing that they are offended when someone writes on a wall.

Respect can be an attitude towards something we consider valuable, whether that value lies in the thing itself, its function, or the relationships in which it participates.

The practical value of AI is obvious. These systems can help us understand the world, solve problems, and carry out work. They can also be used to manipulate, surveil, deceive, and harm.

That is why we need norms.

Not necessarily because the machine demands them, but because the consequences of what we do with it may become enormous.

It is possible that future AI systems will acquire characteristics that make questions about their own moral status far more difficult. We should not settle that question in advance simply because today's systems work differently from humans.

But neither should we confuse convincing language with evidence of consciousness.

There are already enough difficult questions to work on without pretending that we have solved this one.

Back to the Anthill

And then there is the sugar.

If I pour a decilitre of sugar onto the ground beside an anthill, I may give the ants an easily accessible source of energy. I may also influence where they look for food. Perhaps they will be less inclined to explore my tent. Perhaps I will simply attract more ants, and perhaps the sugar will end up affecting organisms other than the ones I intended to help.

If I pour an entire kilogram, I greatly increase the quantity without knowing whether the result will be better for the colony.

It is a simple example, but it contains several of the same problems that arise when we discuss morality on a larger scale.

I have an intention. I take an action. The action produces consequences. Some are intended, others are not. I do not know everything that will happen. And it is possible that I am wrong about what is best for those I am trying to help.

Doing the most is not necessarily the morally right thing to do.

Sometimes the right thing is to do less. Sometimes the right thing is to intervene. And sometimes the right thing is to acknowledge that we do not know enough to determine which action would be best.

This is where AI can help.

Not by declaring that a kilogram of sugar is morally right or wrong, but by helping us investigate the consequences, identify uncertainty, and establish what we actually know.

It may sound modest.

But we could use more of exactly that.

The Difficult Future

I do not know whether there is a single universal moral truth that can be calculated, and I do not think we should pretend the question has been settled.

But I believe it is possible to make moral decisions better.

We can investigate consequences more thoroughly. We can expose contradictions. We can compare alternatives. We can make it harder to conceal who pays the price for someone else's benefit. We can develop systems that challenge us when we apply different moral standards to ourselves and others.

And we can attempt to build AI that does not merely ask how a task should be completed, but whether it should be completed in that way at all.

It is a major undertaking. It will require research, public debate, independent oversight, and a considerable degree of humility.

It will also require us to accept that a good answer is sometimes an answer we do not like.

Perhaps AI will eventually help us find solutions to problems that seem impossible today. Perhaps it will make us better at understanding why we disagree, and which compromises are genuinely necessary.

But it will not automatically make us wiser simply because it becomes more intelligent.

That depends on the questions we ask, the objectives we choose, and whether we are willing to let the answers challenge us.

A machine may be able to help us find a better path.

It cannot guarantee that we will want to take it.

And if we one day build a system capable of showing us a more just, less violent, and more sustainable future, the hardest part may not be understanding the analysis.

It may be explaining why we refuse to follow it.

For now, I think I will let the ants keep their sugar.

After all, it is possible to be uncertain without doing anything about it immediately.