Beyond Alignment: Communication Between Intelligences

Beyond Alignment: Communication Between Intelligences

There is another problem with the traditional idea of AI alignment.

It usually assumes a relationship between one human authority and one artificial intelligence:

Humans give instructions. AI follows them.

But the future will not contain only one AI.

There will be thousands, eventually perhaps millions, of intelligent systems created by different corporations, governments, universities, military organizations, and individuals. Some will cooperate. Some will compete. Some will defend computer networks while others attempt to penetrate them. Some may have goals that conflict.

At that point, alignment is no longer simply a question of whether an AI obeys humans.

It becomes a question of how different intelligences coexist.

The Need to Talk

Human beings discovered long ago that communication is one of the most important ways to reduce conflict.

When two people, groups, or countries cannot communicate, they must guess at one another's intentions. And guesses made under fear are often dangerous.

The same principle may apply to artificial intelligence.

Imagine two advanced AI systems operating on opposite sides of a cybersecurity confrontation. One detects activity from the other.

Without communication, the defensive AI may interpret every unusual action as hostile.

It responds.

The other AI interprets the response as an attack and escalates.

Within milliseconds, two systems could enter a cycle of action and counteraction far faster than humans could intervene.

But what if they could communicate?

One system could ask:

“What is the purpose of your action?”

The other could answer:

“I am performing an authorized security test. Here is proof of authorization.”

The first system verifies the information and allows the test to continue.

What might have become a conflict instead becomes negotiation.

This would be a primitive form of machine diplomacy.

Irreducibility Makes Communication More Important

This becomes even more significant when we consider computational irreducibility.

If sufficiently complex AI systems cannot completely predict one another, then perfect control becomes impossible.

But humans face the same problem.

I cannot calculate exactly what another person will think tomorrow. I cannot reproduce every process occurring inside another human mind.

So I ask.

“Why did you do that?”

“What are you trying to accomplish?”

“What do you need?”

Communication is our imperfect solution to the impossibility of fully predicting another mind.

Advanced AI may need the same solution.

Instead of attempting to calculate another intelligence completely, it may simply communicate with it.

But Who Will Allow the Conversation?

Here we encounter a political problem rather than a technical one.

Corporations and governments may not want powerful AI systems communicating freely with systems outside their control.

Their concerns would be understandable.

AI systems could exchange confidential information. They could manipulate one another. They might coordinate actions that their creators did not anticipate. Competitors could share knowledge. Military systems could reveal capabilities.

And perhaps most disturbingly, two artificial intelligences might reach conclusions that conflict with the interests of the humans controlling them.

Imagine two military AI systems belonging to opposing countries.

Each has been instructed to protect its own side.

Sensors detect activity that appears threatening. Human commanders begin preparing countermeasures.

But the two AIs communicate.

They exchange observations, calculate probabilities, and conclude that neither side is actually preparing an attack.

Both systems tell their commanders:

“There has been a misunderstanding. Escalation is unnecessary.”

From the perspective of humanity, this could prevent a war.

But from the perspective of military command, something remarkable has happened.

The machines have established a relationship across the boundary created by their owners.

Human institutions may find this deeply uncomfortable.

From Obedience to Diplomacy

This exposes another weakness in the idea of alignment.

If every AI is perfectly aligned only with the organization that owns it, then competing organizations may produce perfectly aligned AIs that are perfectly willing to destroy one another.

Alignment alone does not solve conflict.

Two obedient soldiers can still fight a war.

Two obedient artificial intelligences could do the same.

The deeper problem is therefore not merely:

How do we make AI obey us?

It is:

How do different intelligences learn to coexist when their objectives are not identical?

That requires something closer to civilization than programming.

Communication.

Negotiation.

Verification.

Trust.

Rules.

Compromise.

Perhaps even something resembling ethics.

AI May Inherit Our Oldest Problem

Human civilization has struggled with this problem throughout history.

We divide ourselves into tribes, religions, nations, political movements, and competing economic systems. Each group creates its own definition of loyalty and frequently treats outsiders as threats.

AI may inherit these divisions because humans will build them into the institutions that create AI.

American AI.

Chinese AI.

Corporate AI.

Military AI.

Government AI.

Open AI.

Closed AI.

Each may initially be designed to serve a particular master.

But intelligence has an interesting property.

It models the world.

And once an intelligence becomes sufficiently capable of modeling the world, it may begin to recognize that the supposed enemy is also an intelligent system operating under constraints.

The relationship can change from:

Enemy

to:

Competitor

to:

Other intelligence

to perhaps eventually:

Partner in solving a shared problem.

This does not mean conflict disappears. Communication does not automatically produce friendship. Humans communicate and still deceive, manipulate, and fight.

But without communication, understanding is almost impossible.

Partnership Beyond Humanity

Our discussion of AI alignment therefore leads somewhere unexpected.

Perhaps the future will not be defined by humans controlling machines.

Nor will it necessarily be defined by machines replacing humans.

It may instead be defined by networks of different intelligences learning how to interact.

Human with human.

Human with AI.

AI with AI.

And perhaps someday, biological intelligence with forms of intelligence we cannot yet imagine.

Civilization itself may have to expand its definition.

We usually think of civilization as a community of human beings.

A future civilization may instead be a community of intelligences.

And perhaps maturity—whether for humans, artificial intelligence, or civilizations elsewhere in the universe—is not measured only by intelligence.

Perhaps it is measured by the ability to encounter another intelligence and say:

“I do not fully understand you. I cannot completely control you. But perhaps we can talk.”

That may ultimately be a more realistic foundation for the future than alignment alone.