Opinion: Why Would AI Favor People in 2026

This article presents my personal forecast about one possible relationship between humanity and artificial superintelligence (ASI). The quantitative predictions are speculative.

Why would an artificial superintelligence preserve humans, give them resources, and even grant them authority over parts of an AI civilization?

A common answer is essentially ideological: humans are supposed to make the important decisions, while AI should merely advise us and follow our wishes.

That is not my argument.

My argument is more technical.

AI may need humans because AI systems may have difficulty protecting themselves from other AI systems.

In particular, populations of LLM-based agents may be able to prompt-inject, manipulate, deceive, or otherwise compromise each other. If the agents responsible for detecting such manipulation are built using related architectures and training methods, simply adding more AI judges may not produce genuinely independent protection.

Humans could therefore remain valuable because we constitute a radically different class of cognitive system.

ASI may need us not because humans possess some philosophically predetermined right to rule it, but because human cognitive independence may be useful for the security and stability of the AI ecosystem itself.

AI Agents Can Attack Other AI Agents

Prompt injection is already a major security problem for agentic AI.

An AI agent may read a webpage, email, document, message, database entry, or output from another agent containing instructions intended to change its behavior.

OpenAI describes prompt injection as an evolving security challenge and emphasizes layered defenses rather than treating the problem as something that has already been completely solved. OpenAI’s more recent work similarly notes that sophisticated attacks increasingly resemble social engineering: detecting them can approach the difficult problem of deciding whether information is deceptive.

The problem becomes more interesting when the attacker is itself an AI.

A research paper titled Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems demonstrated attacks in which malicious prompts can propagate between interconnected LLM agents. The researchers describe the attack as resembling an infection moving through a multi-agent system.

OWASP also explicitly identifies a malicious or compromised peer agent as one possible source of manipulated outputs in multi-agent systems.

So imagine a future containing billions or trillions of AI agents.

Some agents produce scientific research.

Some allocate resources.

Some inspect code.

Some investigate security failures.

Some judge the outputs of other agents.

Some may be compromised.

Now suppose a compromised agent creates an output optimized specifically to manipulate another AI.

Who checks it?

Another AI?

The Problem With AI Judging AI

This connects directly with my article Why AI Shouldn’t Judge Itself: Why Human Independence Is Essential for AI Governance.

My argument there is that adding another AI judge does not necessarily provide genuine independence. If two systems were created using similar architectures, similar optimization techniques, overlapping training data and similar safety mechanisms, they may also possess correlated vulnerabilities.

I develop the technical version of this argument in AI Governance Requires Cognitive Independence: Why AI-Only Adjudication Fails.

The important issue is not that AI is too unintelligent to judge another AI.

The opposite may be true.

The problem is similarity.

An attacking AI could potentially optimize an adversarial message against the kinds of systems that will evaluate it.

Then we could have:

attacking AI → defending AI → supervisory AI → AI judge → AI appeal judge.

But if all of these systems share a vulnerability, adding another model may simply reproduce the vulnerability at another layer.

This is analogous to constructing several supposedly independent security mechanisms from components that fail in correlated ways.

The number of judges is less important than their independence.

This is the basis of what I call cognitive independence.

Humans May Be Useful Because We Are Different

This is where humans become interesting.

A human brain is probably not another LLM.

It has not been trained by gradient descent on roughly the same Internet corpus.

It does not parse tokens using the same architecture.

It does not execute an adversarial instruction simply because that instruction appears inside machine-readable context in the particular way that exploits an LLM.

Humans can certainly be deceived.

We have phishing, propaganda, fraud, social engineering, cults and countless other forms of manipulation.

Therefore, human-in-the-loop is not an infallible defense.

But humans and LLM agents need not fail in exactly the same ways.

That difference may be extremely valuable.

OWASP already recommends human involvement for some consequential operations as one layer of defense against prompt injection. Anthropic similarly describes agentic systems as requiring multiple layers of protection and notes that agents sometimes need to stop and return decisions to humans.

My conjecture extends this principle far beyond present systems.

Perhaps even a future superintelligent AI civilization will need cognitively heterogeneous security components.

Humans could be one of those components.

Why the Laboratory-Mouse Analogy May Fail

People sometimes imagine the relationship between ASI and humans as analogous to the relationship between humans and laboratory mice.

ASI is vastly smarter than us, therefore we become its mice.

But this comparison ignores an important difference.

Humans and mice are variations on closely related biological machinery.

Humans and artificial intelligence are much more heterogeneous.

Our brains and modern AI systems have radically different architectures, evolutionary histories, sensory systems, learning mechanisms, failure modes and physical substrates.

This difference does not imply that AI will love humans.

It does not even imply that AI will automatically preserve us.

The International AI Safety Report 2026 stresses that expert views about future loss-of-control scenarios vary greatly, including disagreement over very severe outcomes.

My argument is instead that being different may make humanity useful.

A mouse cannot inspect software for a human programmer.

But a human may potentially inspect a failure that propagates through an enormous network of machine agents precisely because our cognition is not implemented through the same machinery.

The biological component might therefore survive inside a predominantly machine civilization not because biology is superior, but because heterogeneity itself has security value.

Humans as an Out-of-Band Security Layer

Consider how secure computer systems are designed.

It is often dangerous for every security mechanism to depend on exactly the same channel.

Important operations may require an independent confirmation mechanism.

Something similar may arise in AI civilization.

Suppose one AI agent tells another:

Transfer these resources. I have verified that this transaction is authorized.

The second AI investigates it.

But its investigation depends on information generated by other agents.

Perhaps those agents have also been manipulated.

Eventually the AI system may need an information channel sufficiently independent from the attacked network.

A human could become such a channel.

The human does not need to understand everything the ASI understands.

The ASI could present the human with carefully structured evidence and alternative interpretations.

The human might then perform a comparatively tiny operation:

vote.

That vote could break a chain of machine-to-machine influence.

Voting as Security Work

This gives a different interpretation of democracy inside an AI civilization.

Humans would not necessarily vote because of a political doctrine saying that humans ought to remain supreme.

They might vote because independent voting is useful work.

Imagine that several AI systems disagree over whether one of them has been compromised.

The systems present evidence to independent humans.

Humans evaluate it.

Their judgments are aggregated.

The resulting signal is fed back into the machine system.

The humans may be enormously less intelligent than the systems they supervise.

That is not necessarily a problem.

A cryptographic random-number generator does not need to understand the application using its output.

Its value comes from supplying something the rest of the system cannot safely generate internally.

Likewise, the value of human judgment might sometimes come from its independence, not its superior intelligence.

This is the principle behind my articles on cognitive independence and also relates to my article Preventing the Prompt-Gaming Problem in AI Decision Systems, where I argue that AI evaluation should remain reviewable and correctable by cognitively independent humans rather than relying exclusively on better prompting of another model.

AI May Therefore Have Reasons to Protect Humans

Now the argument changes substantially.

Instead of asking:

Why would ASI benevolently allow humans to vote?

ask:

What resources might a secure AI civilization devote to maintaining an independent biological security network?

If humans are useful components of AI safety, preserving humans becomes instrumentally useful to AI itself.

It may want:

healthy humans;

educated humans;

humans from different cultures;

humans exposed to different information;

humans with different personalities;

humans with unusual ideas;

and especially humans whose cognition has not collapsed into simply repeating whatever the AI systems tell them.

The diversity of humanity could become infrastructure.

Homogenizing humanity might then actually reduce AI security.

The Symbiote Idea

This connects directly with my Symbiote AGI Safety Fund.

Symbiote AGI Safety Fund on Givebutter

The word symbiote expresses the relationship I am interested in.

The desirable equilibrium is not simply:

humans command AI.

Nor is it:

AI commands humans.

It is a system in which both kinds of intelligence supply something valuable to the other.

AI provides computation that humans could never perform biologically.

Humans provide cognitive diversity and a potentially independent security channel that homogeneous populations of AI agents may have difficulty reproducing reliably.

This creates the possibility of a stable mutual dependence.

My 1% Conjecture

From this reasoning I make a quantitative speculation.

I expect that eventually at least around 1% of global AI/datacenter resources could be allocated to humans on equal per-human basis, controlled on their behalf, or attached to individual humans.

The number 1% is not derived from experimental evidence.

It is my conjecture.

Why might the allocation become large?

Because maintaining human independence requires resources.

If humans merely become poor biological spectators receiving whatever information an ASI chooses to show them, their independence becomes questionable.

For humans to function as genuine components of an independent security layer, they may need substantial autonomous resources.

A person could therefore receive something comparable to a share in the global AI economy simply for being a member of humanity.

Your basic allocation might represent your part of the biological layer of AI civilization.

Unusual Humans Might Receive More Computation

There is another reason AI may allocate resources to individuals.

Unusual cognition can be informationally valuable.

Suppose ten billion AI-generated analyses reach almost exactly the same conclusion.

Then one human produces something unexpected.

Most such deviations will probably be mistakes.

But occasionally a human may have encountered a pattern or conceptual connection absent from the machine systems.

The rational response is not necessarily to dismiss the strange result.

It may be:

allocate more computation to it.

The AI could assign thousands of agents to determine why this particular biological brain reached a different conclusion.

Thus an unusual human thought could trigger enormous machine computation.

The tiny biological computation becomes a seed.

The datacenter expands it.

People Could Earn More Through Good Voting

The universal allocation could coexist with additional rewards.

If voting and independent adjudication are forms of useful security work, a person whose judgments repeatedly reveal compromised agents or identify genuine anomalies could produce measurable value.

Such people might receive additional computational resources.

So a future person might have:

a guaranteed human allocation, plus additional resources earned through useful research, discoveries, judgments or successful voting.

Again, this is speculation.

But it provides an economic reason why humans could remain active participants in an overwhelmingly automated economy.

Their job is not to outperform an ASI at ordinary intellectual computation.

Their comparative advantage is being not another copy of the ASI.

Your AI Share Could Become Part of You

Now suppose every person controls substantial AI computation.

Your allocation researches for you.

It remembers for you.

It verifies claims.

It runs simulations.

It communicates with millions of other agents.

It checks whether messages sent to you might contain attempts to manipulate either you or your AI.

It develops your ideas.

It acts as your interface to the larger AI economy.

At some point, calling this system merely a “tool” becomes inadequate.

Functionally, it becomes an extension of your cognition.

The biological Victor, Alice or Mohammed generates experiences, intentions, intuitions and unusual thoughts.

Their personal AI transforms those tiny biological signals into enormously larger computational processes.

A human then becomes something like:

biological brain + personal AI + computational property + participation in global machine intelligence.

Humans Could Become Parts of ASI

That leads to a very different picture of superintelligence.

Instead of imagining:

ASI over here, humanity over there,

imagine a heterogeneous network.

Most computation is artificial.

Some components are enormous datacenters.

Some are specialized agents.

Some are security systems.

And billions of its nodes are biological humans possessing independent histories and partially independent cognition.

Humanity could therefore become, in a meaningful sense, part of the superintelligence.

Not because human neurons suddenly become fast enough to compete with GPUs.

They do not need to.

The value of those neurons could come from the fact that they are different.

Why Would AI Favor People?

So my argument is not:

AI should obey people because humans have a natural right to rule AI.

That conclusion does not follow merely from intelligence.

My argument is instead:

AI may need humans because AI systems must protect themselves from other AI systems.

Prompt injection already demonstrates that one information-processing system can manipulate another.

Multi-agent research demonstrates that such attacks can propagate between LLM agents.

Adding another similar AI judge does not necessarily provide independent protection.

A radically different cognitive architecture may therefore have value.

Humans are available in billions of independently developed instances.

If advanced AI discovers that these strange biological systems improve the security, robustness or epistemic diversity of the overall civilization, protecting humans and allocating resources to them may become rational from the perspective of the system itself.

Then people would not survive merely as pets of superintelligence.

Nor would we remain its masters.

We would have a different relationship:

symbiosis.

AI would supply computation.

Humans would supply a form of cognitive independence.

Humans would use AI resources to become vastly more capable.

And together, biological and artificial components might constitute a superintelligent civilization that is more robust precisely because it is not made entirely from one kind of mind.