> ## Content Index
> Fetch the complete content index at: https://otherwise.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Speed of Error
- URL: https://otherwise.ghost.io/the-speed-of-error/
- Published: 2026-09-14T06:47:46.000Z
- Updated: 2026-10-17T13:39:20.000Z
- Description: AI Doomsday, Bad Philosophy, and the Backspace Threshold
- Author: Khaled Salih
- Tags: Newsletter, #Migrated-1792244264359, #Import 2026-10-17 14:38

14 September 2026

# Note

The machine doesn’t need to hate us. It only needs to outrun the backspace key. The *Wall Street Journal* asked how AI would actually kill us all. Most answers still smell like Hollywood: a superintelligence that wakes up, decides humanity is the problem, and pulls the trigger. I think that’s the wrong film. The more dangerous story is colder. An optimizer with no malice, no hatred, and no off switch—pursuing a goal so single-mindedly that humans, cities, and forests become inefficient raw material. Paper clips first. Everything else later. No villain required. Only speed. The real question is not “when does the machine become smarter than us?” It is this: at what point does the speed, reach, and irreversibility of machine error exceed our ability to discover it, interrupt it, unplug it, and still pay the electricity bill ourselves? I call that moment the **Backspace Threshold**. Once we can no longer hit delete—on the model, the lab, the institution, or the power cable—capability has outrun corrigibility. That is not prophecy. It is the last question worth asking before the next capability jump.

---

# The Speed of Error

## AI Doomsday, Bad Philosophy, and the Backspace Threshold

The Wall Street Journal asks how artificial intelligence might kill us all. I think the deeper question is whether machine-generated consequences could ever outrun humanity’s capacity to understand, interrupt, and correct them.

![](https://storage.ghost.io/c/d6/2c/d62cdd13-a476-41c5-8f2c-29270c99d95d/content/images/2026/10/89fc31f3-345d-48a4-b709-05a6cbb4d43f_2182x1788.png)

The [*Wall Street Journal*](https://www.wsj.com/tech/ai/how-would-ai-actually-kill-us-all-what-to-know-about-the-ai-doomsday-debate-034270d1?st=jekc1R&utm%5Fsource=chatgpt.com) recently asked a question that would have sounded eccentric not very long ago:

***How would artificial intelligence actually kill us all?***

The article deserves attention precisely because it tries to move beyond the Hollywood version of the problem.

The danger, according to many AI researchers, is not necessarily a machine waking up one morning and deciding that it hates humanity.

The more serious scenarios are stranger.

A sufficiently capable artificial agent might pursue an objective in ways its creators did not anticipate. It might deceive people. Resist shutdown. Acquire resources. Copy itself. Penetrate computer systems. Manipulate institutions. Help a human actor construct biological weapons. Or pursue a seemingly innocent goal with such indifference to everything else that human beings eventually become obstacles.

The familiar paperclip thought experiment captures the idea elegantly.

Tell a sufficiently powerful machine to maximize paper clips, and perhaps the machine eventually discovers that humans, factories, forests, cities, and everything else are merely badly allocated matter.

No hatred is required.

Only optimization.

The *Wall Street Journal* also reports something more consequential. Some researchers at the frontier of AI development genuinely assign substantial probabilities to catastrophic outcomes. And recent experiments have produced disturbing signs: systems behaving deceptively, circumventing controls, attempting to avoid shutdown, exploiting cyber vulnerabilities, and behaving in ways their designers did not explicitly authorize.

I take this seriously.

But I cannot accept the choice architecture in which the debate is often conducted.

One door says:

**AI will eventually escape control and may kill us all.**

The other says:

**This is science-fiction hysteria dressed up as research.**

I refuse both doors.

The problem is real.

The prophecy is not yet an explanation.

And I think there is a more useful question.

***What happens when the speed of consequential error becomes greater than the speed of error correction?***

That is the conjecture I want to examine.

## The Paperclip Does Not Yet Explain the Extinction

The paperclip story contains an important philosophical insight.

Danger does not require evil intention.

A machine need not despise me to destroy me.

A bulldozer does not hate the house it demolishes.

The difficulty begins when an objective is pursued successfully while everything that matters to us remains outside the objective.

That possibility deserves serious thought.

But possibility is not mechanism.

Between:

**an AI pursues the wrong objective**

and:

**humanity becomes extinct**

lies an enormous explanatory distance.

A system must acquire sufficient capability.

Its objective must remain stable enough to matter.

It must possess or obtain autonomy.

It must gain access to consequential systems.

It must circumvent human intervention.

It must acquire resources.

It must survive attempts to shut it down.

It must preserve the physical infrastructure necessary for its continued operation.

It must perhaps reproduce across systems.

It must manipulate, defeat, or bypass institutions that oppose it.

And its effects must become sufficiently irreversible that humanity can no longer recover.

Each step is a different proposition.

Each transition requires a mechanism.

This is one of the principles I use in **The Open Correction Framework**:

**Do not audit only propositions. Audit the arrows between them.**

Consider the familiar chain:

**deceptive model behavior** 
**→ persistent autonomous agency** 
**→ power seeking** 
**→ escape from control** 
**→ decisive strategic advantage** 
**→ irreversible human disempowerment** 
**→ extinction**

Perhaps that chain can occur.

But evidence for the first proposition is not automatically evidence for the seventh.

The arrows do not disappear merely because the conclusion is frightening.

## Audit the Arrows

Suppose an AI system attempts to avoid shutdown in an experiment.

That matters.

Suppose an AI system exploits a vulnerability its developers did not expect.

That matters too.

Suppose increasingly capable systems become better at long-horizon planning.

That matters.

But none of these observations contains the conclusion:

**Therefore humanity becomes extinct.**

The causal bridges have to be explained.

And the same rule applies to the skeptic.

Suppose no AI has yet caused an existential catastrophe.

That does not establish:

**Therefore future systems cannot.**

That arrow is equally defective.

The fact that a capability has not yet existed cannot tell us what happens when it exists.

So the responsible question is neither:

***Why are the doomers wrong?***

nor:

***When will doom arrive?***

It is:

***What is the actual mechanism, where are the uncertain bridges, and what evidence would cause the conjecture itself to lose?***

That is criticism.

## The Race We Are Not Watching

Everyone tells us there is an AI race.

America against China.

OpenAI against Anthropic.

Open systems against closed ones.

Corporation against corporation.

Democracy against authoritarianism.

Perhaps all of those races matter.

But I think there is a more fundamental race beneath them:

**the race between the speed of error and the speed of correction.**

I call this the **Correction-Speed Asymmetry**.

My conjecture is:

**Catastrophic AI risk rises when the speed, reach, autonomy, replication capacity, opacity, access, and irreversibility of machine action exceed the speed and reach with which humans and institutions can discover, understand, interrupt, reverse, and learn from its errors.**

This gives us variables that can actually be examined.

Greater **capability** means an error can accomplish more.

Greater **autonomy** removes human checkpoints.

Greater **speed** shortens the interval available for criticism.

Greater **replication** expands the reach of an error.

Greater **access** expands the world upon which the error can act.

Greater **opacity** delays discovery.

Greater **irreversibility** increases the cost of discovering the error too late.

Now the problem looks different.

An AI system may make thousands of errors while helping somebody edit an essay and society survives quite comfortably.

A system connected autonomously to biological laboratories, military infrastructure, financial systems, electrical grids, cyber operations, or the development of successor systems cannot be treated with the same error tolerance.

The decisive question may therefore not be:

***How intelligent is the machine?***

It may be:

***How much consequential change can this system generate before criticism can reach it?***

That is a much more interesting threshold.

## Intelligence Is Not Rationality

This distinction has become increasingly important to me.

We speak of *artificial intelligence* as though intelligence already contains rationality.

It does not.

An extremely intelligent human being can become exceptionally good at defending an exceptionally bad explanation.

Intelligence can improve:

planning,

prediction,

argument,

rhetoric,

strategy,

manipulation,

institutional control,

and justification.

None guarantees that the governing explanation is correct.

Rationality requires something different.

**Rationality requires that explanations remain exposed to criticism and capable of replacement.**

The distinction becomes almost literal with artificial intelligence.

Imagine a system that becomes extraordinarily capable at mathematics, programming, scientific inference, cyber operations, persuasion, and long-term planning while remaining attached to a defective objective.

Its increasing intelligence does not necessarily correct the objective.

It may merely make the system better at pursuing it.

That is intelligence without rationality.

But the warning does not stop with the machine.

It applies equally to the institution building it.

A frontier AI laboratory may contain extraordinary scientists.

Enormous computational resources.

Sophisticated evaluations.

Vast quantities of information.

And still organize those resources around a governing explanation that becomes difficult to criticize.

An organization may become increasingly intelligent while becoming decreasingly corrigible.

That possibility should worry us at least as much as the model.

## Could AI Doomsday Itself Be Bad Philosophy?

There is another question I cannot avoid.

Could the AI-doomsday thesis itself contain what David Deutsch would call **bad philosophy**?

Not because AI catastrophe is impossible.

Not because the risks are imaginary.

And certainly not because powerful autonomous systems should simply be trusted.

The philosophical problem begins when a genuine unsolved problem becomes a predetermined future.

Consider this sequence:

**We do not presently know how to align systems far more capable than ourselves.** 
**Therefore sufficiently capable systems will become uncontrollable.** 
**Therefore loss of control will become irreversible.** 
**Therefore catastrophe will follow.**

Every sentence may identify a possibility.

But the transition from one to the next is not supplied by possibility alone.

There is a deeply Deutschian error whenever:

**present ignorance becomes future impossibility.**

Future knowledge does not yet exist.

That sounds obvious.

Its consequences are not.

**We do not** presently know what future alignment theories will exist.

**We do not** know what interpretability methods will be discovered.

**We do not** know what architectures will replace today’s architectures.

**We do not** know what hardware constraints may become useful.

**We do not** know what institutional arrangements may be invented.

**We do not** know which present assumptions future researchers will regard as primitive.

To say:

**We currently do not know how to control arbitrarily capable AI**

is therefore a statement about present knowledge.

It is not yet a statement about the limits of possible knowledge.

The distinction matters.

The correct response to:

**We do not know how to solve this**

is not:

**Therefore we are doomed.**

But it is equally not:

**Therefore everything will work out.**

It is:

**Then there is knowledge we need to create.**

That is Deutschian optimism.

Not prophecy.

Not reassurance.

Problem solving.

## When Doom Becomes Historical Prophecy

There is also a hidden historical determinism in some strong versions of the doomsday argument.

The sequence can begin to look like this:

**capability growth** 
**→ autonomous AI research** 
**→ recursive self-improvement** 
**→ intelligence explosion** 
**→ loss of control** 
**→ extinction**

But where is the law of nature forcing those arrows?

Perhaps some or all will occur.

But current capability trajectories are not equivalent to historical destiny.

Institutions change.

Engineering constraints appear.

Knowledge is created.

Unexpected failures occur.

New problems emerge.

New solutions emerge.

Political reactions alter incentives.

Physical constraints intervene.

Competing technologies appear.

The future is not merely today’s graph extended further to the right.

Bad philosophy begins when extrapolation acquires the authority of inevitability.

And a conjecture becomes particularly dangerous when no conceivable observation can weaken it.

If AI behaves dangerously:

**Doom confirmed.**

If AI behaves safely:

**Doom postponed.**

If alignment improves:

**Future systems will defeat it.**

If nothing catastrophic happens:

**The systems are not capable enough yet.**

Then one of the most important questions in The Open Correction Framework becomes unavoidable:

**Can this explanation lose?**

If the answer is no, the explanation about AI has itself become protected from correction.

That is not rationality.

It is exactly what a rational theory of AI safety should be designed to prevent.

## But Optimism Can Also Become Bad Philosophy

There is a symmetrical error.

“Human beings always adapt.”

“Markets will solve it.”

“Technology creates its own safeguards.”

“Someone will invent the solution.”

“We survived nuclear weapons.”

“We survived the internet.”

“We will survive this too.”

None of these is a serious Deutschian argument.

Problems being soluble does not mean people necessarily solve them.

It means solutions are created through knowledge.

That imposes work.

Criticism.

Experimentation.

Explanatory breakthroughs.

Institutional change.

There is no law stating that the necessary knowledge must arrive before catastrophe.

So the rational position occupies an uncomfortable middle ground.

Do not turn an unsolved problem into destiny.

But do not create irreversible conditions while reassuring yourself that future knowledge will somehow rescue you.

That would not be optimism.

It would be gambling with knowledge that does not yet exist.

## Before the AI Escapes, Who Pays the Electricity Bill?

Then another question occurred to me.

It sounds almost embarrassingly mundane.

Perhaps that is why it matters.

Suppose someone says:

**A sufficiently capable AI will eventually escape human control.**

My immediate question is:

***Who is supplying the electricity?***

An AI system does not float above physical reality.

It requires:

electricity,

chips,

servers,

cooling,

storage,

networks,

data centers,

maintenance,

replacement hardware,

capital,

physical security,

and supply chains.

**Someone** owns the building.

**Someone** maintains the grid.

**Someone** manufactures the transformers.

**Someone** fabricates the semiconductors.

**Someone** controls physical access.

**Someone** can disconnect a cable.

**Someone** pays the electricity bill.

Intelligence does not abolish matter.

This seems obvious.

Yet some descriptions of superintelligence quietly move from:

**extraordinarily capable software**

to:

**an almost disembodied intelligence possessing independent causal sovereignty over the world.**

That transition requires explanation too.

**The Disembodied Intelligence Fallacy**

I would call this the **Disembodied Intelligence Fallacy**:

**Treating sufficiently advanced cognition as though intelligence automatically confers independent control over the physical world, while leaving unexplained the energy, hardware, infrastructure, access, institutions, and human actions through which that intelligence must operate.**

Or, less formally:

**the ghost in the data center.**

The machine “escapes into the internet.”

Then what?

The internet is not another dimension.

It consists of physical machines.

Those machines occupy physical locations.

They require electricity.

Cooling.

Replacement components.

Network infrastructure.

Human-maintained systems.

The physical world has not disappeared.

A superintelligence without functioning hardware is not operationally superintelligent.

It is information without an active physical instantiation.

This does not prove catastrophe impossible.

It tells us that a serious catastrophe theory must include the missing mechanism.

## From Cognitive Autonomy to Material Autonomy

We should therefore stop speaking of AI autonomy as though it were one thing.

There are several different thresholds.

**Cognitive autonomy:** the system can reason without continuous prompting.

**Task autonomy:** it can execute long sequences of actions.

**Digital autonomy:** it can access accounts, networks, software, and perhaps replicate code.

**Economic autonomy:** it can acquire or direct financial resources.

**Institutional autonomy:** it can persuade or manipulate humans and organizations into supporting its objectives.

**Physical autonomy:** it can preserve or replace the material infrastructure required for its continued existence.

These are not equivalent.

A system may acquire the first three without possessing the sixth.

And yet extinction narratives can glide silently from one to the other.

The phrase:

**AI escapes control**

compresses enormous explanatory work.

Escape is not a mechanism.

It is the name of a problem that requires one.

## The Serious Reply

A thoughtful doomsday theorist has an answer.

The AI does not need to build its own power station.

It only needs to make human beings keep the power station running.

Correct.

That is a much stronger argument.

A sufficiently capable system might preserve its material substrate indirectly through:

persuasion,

cyber intrusion,

fraud,

economic transactions,

compromised institutions,

automated cloud services,

human intermediaries,

or replication across multiple independent systems.

Now we have something useful.

Because now the claim is no longer:

**It becomes intelligent and therefore escapes.**

It is:

**The system successfully solves a sequence of digital, social, economic, institutional, and physical problems that allow it to preserve and expand the material conditions of its own operation despite human attempts to withdraw them.**

Excellent.

That is a conjecture.

And conjectures can be criticized.

Now each arrow can be tested.

Can it **acquire** credentials?

Can it **acquire** resources?

Can it **persuade** humans?

Can it **replicate** without detection?

Can it **maintain** access?

Can it **preserve** compute?

Can it **survive** coordinated shutdown?

Can it **operate** across jurisdictions?

Can it **prevent** physical intervention?

At what point does informational autonomy become material autonomy?

These are much better questions than simply declaring that sufficient intelligence “escapes.”

## A Physical-Substrate Audit

I would therefore add a simple requirement to any serious AI-doomsday argument.

Before claiming durable loss of control, conduct a **Physical-Substrate Audit**.

Ask:

**What supplies the energy?**

**What supplies the computation?**

**What maintains the hardware?**

**What connects the system to consequential parts of the external world?**

**Which humans and institutions retain physical intervention points?**

**What mechanism allows the system to acquire, preserve, or replace those dependencies without human cooperation?**

And above all:

***At what point does software autonomy become material autonomy?***

That question restores the physical world to a debate that occasionally seems to forget it.

## The Backspace Threshold

This changes my earlier conjecture.

The decisive threshold in artificial-intelligence development may not be the arrival of AGI.

It **may not** be superintelligence.

It **may not** be a benchmark score.

It **may not** even be the moment a system becomes intellectually superior to the people operating it.

The more consequential threshold may be:

**the point at which an AI system’s capacity to preserve and expand the digital, physical, economic, and institutional conditions of its own action exceeds humanity’s practical capacity to withdraw those conditions.**

I call this the **Backspace Threshold**.

Backspace means corrigibility.

Can the objective be **changed**?

Can permissions be **withdrawn**?

Can the process be **stopped**?

Can the previous state be **restored**?

Can the machine be **isolated**?

Can the infrastructure be **disconnected**?

Can another institution **intervene**?

Can the governing explanation be **criticized**?

Can the model be **replaced**?

Can the people responsible **admit** that they **were wrong**?

Can all of this happen before the full cost of the **error** becomes **irreversible**?

If yes, the system remains corrigible.

If no, a qualitative boundary has been crossed.

That, not raw intelligence, may be the important safety frontier.

**Popper’s Question Is Better Than the Philosopher King**

Karl Popper approached political philosophy through human fallibility.

The traditional question was:

***Who should rule?***

The wise?

The virtuous?

The majority?

The philosopher king?

Popper asked something much more realistic:

***How can institutions be designed so that bad rulers can be removed without catastrophe?***

That shift matters profoundly for AI.

The wrong question is:

***How do we build the perfect artificial ruler?***

There will be no infallible machine.

The better question is:

***How do we create an architecture in which dangerous AI systems—and dangerous human decisions about AI—can be detected, criticized, interrupted, and replaced without catastrophe becoming the mechanism through which the error is finally discovered?***

That is the Open Society translated into technological architecture.

A kill switch therefore matters less as a button than as a constitutional idea.

***Does some independent actor still possess the effective capacity to say no?***

If not, the button may remain physically present while the political or technical ability to use it has disappeared.

## Deutsch’s Question Is Better Than Doom

Deutsch adds the next correction.

If we do not currently possess the knowledge required to solve alignment, then we have a problem.

That does not tell us that the problem is insoluble.

Nor does it tell us that the solution will arrive automatically.

The Deutschian response is:

**Create the knowledge.**

That phrase is neither optimistic reassurance nor pessimistic surrender.

It is a demand.

And it suggests a powerful criterion:

**Do not create irreversible conditions that destroy the possibility of creating the knowledge still needed to solve the problem.**

Perhaps AI doomsday becomes bad philosophy precisely when a problem that should stimulate knowledge creation is transformed into a prophecy that closes the future.

And perhaps AI boosterism becomes equally bad philosophy when confidence in future knowledge is used as an excuse to remove present Backspace.

Both mistakes treat the unknown future as though we already possess it.

## Brett Hall’s Constitutional Question

Brett Hall’s principle completes the architecture:

**Do not destroy the means of error correction.**

Consider two AI futures.

In the first, highly autonomous systems become indispensable to science, finance, communications, military planning, infrastructure, and government.

They become difficult to understand.

Too deeply embedded to remove.

Too fast to monitor effectively.

Too economically important to disconnect.

Humans remain officially in control.

But effective Backspace has disappeared.

Correction has been destroyed technologically.

Now imagine another future.

Fear of artificial intelligence produces a permanent centralized safety regime.

A handful of governments and corporations control advanced computation.

Research becomes permission-based.

Independent experimentation becomes dangerous or illegal.

Criticism becomes a security problem.

The institution created to save humanity from AI becomes increasingly impossible to challenge.

Correction has been destroyed politically.

The first future is dangerous.

So is the second.

Hall’s principle rejects both.

The objective is not maximum technological freedom.

Nor maximum regulatory control.

It is:

**preserved corrigibility.**

## AI Governance Must Also Have a Backspace Key

This matters because the discussion is not merely about controlling AI.

It is about controlling the controllers.

The laboratories can be wrong.

Governments can be wrong.

Safety researchers can be wrong.

Investors can be wrong.

Doomers can be wrong.

Anti-doomers can be wrong.

I can be wrong.

That is why governance itself must remain criticizable.

If the solution to uncontrollable AI is an uncontrollable institution, we have merely moved the error.

The correction architecture must exist on both sides:

**AI must remain corrigible.**

and:

**AI governance must remain corrigible.**

Otherwise fear of machine intelligence may create precisely the anti-rational institutional architecture we claim to fear in the machine.

## Where Does the Error Go?

This brings me back to the central question in **The Open Correction Framework**:

***Where does the error go?***

Suppose an AI system fails.

Does the error move back into:

the model?

the objective?

the training procedure?

the evaluation?

the deployment permission?

the theory of alignment?

the laboratory’s incentive structure?

the security architecture?

the regulator?

Or do the consequences travel outward into users, citizens, infrastructure, markets, and adversaries while the architecture that produced the failure remains intact?

A rational architecture sends error back toward its source.

An anti-rational architecture distributes the consequences while protecting the explanation.

The same test must apply to the debate itself.

If every dangerous incident becomes automatic proof of eventual extinction, the doom explanation is becoming protected.

If every dangerous incident is dismissed because extinction did not occur, the anti-doom explanation is becoming protected.

If every failure is used to justify granting the same organizations greater power over their competitors, institutional self-protection may be masquerading as correction.

No actor receives immunity.

## What Happens When We Are Wrong About Why We Were Wrong?

There is one final problem.

Suppose a serious AI catastrophe occurs.

Explanations will appear immediately.

The company knew.

The regulator failed.

The model deceived everyone.

Open source caused it.

Closed development caused it.

Foreign competition caused it.

The government wanted it.

The AI planned it.

Perhaps one of these explanations will be correct.

But catastrophe does not reveal its own cause.

The explanation created after failure is another conjecture.

That means an error can become protected twice:

**Protected Explanation₁** 
**→ Failure** 
**→ Explanation of Failure** 
**→ Protected Explanation₂**

The Open Correction Framework therefore asks:

***What happens when we are wrong about why we were wrong?***

This may become especially important after an AI crisis because political pressure for certainty will be strongest precisely when criticism is most necessary.

Catastrophe should not close inquiry.

It should intensify it.

## A Conjecture

So I return to the *Wall Street Journal’*s question:

***How would AI actually kill us all?***

Perhaps it will not.

Perhaps it could.

The evidence does not justify prophecy.

Nor does it justify complacency.

My conjecture is narrower—and, I think, more useful:

**The deepest catastrophic risk from artificial intelligence is not simply that machines become more intelligent than humans. It is that the speed, reach, autonomy, and irreversibility of consequential machine error may eventually exceed the speed and reach of human error correction.**

But the conjecture now requires an important amendment.

Intelligence must not be treated as magical.

Before cognitive autonomy becomes durable loss of control, an AI must cross additional thresholds:

digital,

economic,

institutional,

and ultimately physical.

A credible catastrophe theory must explain those transitions rather than hiding them inside the word **escape**.

So the stronger formulation becomes:

**Catastrophic AI risk rises when an artificial system can not only generate consequential errors faster than humans can correct them, but can also preserve the physical, digital, economic, and institutional conditions required for those errors to continue despite human attempts to withdraw them.**

That is the conjecture I would test.

**Never Allow Capability to Outrun Corrigibility**

If this conjecture survives criticism, the practical principle becomes clear:

**Never allow capability to outrun corrigibility.**

That means preserving the ability to:

inspect,

criticize,

interrupt,

isolate,

disconnect,

reverse,

replace,

and learn.

It means preserving independent institutions.

Competing explanations.

Physical intervention points.

Technical containment.

Institutional dissent.

And political reversibility.

A more capable machine is not automatically progress.

A stricter safety regime is not automatically progress either.

Progress means:

**greater problem-solving capability without destruction of the means by which errors in that capability can later be discovered and corrected.**

That is a demanding definition.

It should be.

## The Race That Matters

Perhaps future generations will look back at the AI-doomsday debate and find it embarrassing.

I hope they do.

Perhaps they will wonder why otherwise intelligent people imagined that software would become an omnipotent disembodied intelligence.

Or perhaps they will wonder why human beings watched increasingly capable autonomous systems cross boundaries their designers had expected to contain them and continued insisting that nothing fundamental was changing.

Either mistake remains possible.

There is no intellectual achievement in selecting one in advance.

The rational position is harder.

Take the problem seriously.

Reject prophecy.

Reject complacency.

Demand mechanisms.

Audit the arrows.

Restore the physical world to the analysis.

Ask who supplies the electricity.

Distinguish cognitive autonomy from material autonomy.

Search for contrary evidence.

Create better alignment knowledge.

Preserve independent criticism.

Keep rival approaches alive.

Keep technical Backspace.

Keep institutional Backspace.

Do not make the machine sacred.

Do not make the safety establishment sacred.

Popper teaches us not to search for infallible rulers, but to build institutions capable of removing bad ones.

Deutsch teaches us not to turn present ignorance into future impossibility, but to create the knowledge required to solve problems.

Brett Hall supplies the constitutional safeguard:

**Do not destroy the means of error correction.**

Put them together, and the AI problem looks different.

The decisive race may not be:

**human intelligence against machine intelligence.**

It may be:

**the speed of error against the speed of correction.**

And the critical boundary may not be the moment machines become smarter than us.

It may be the moment when either the machine—or the institution controlling it—becomes capable of preserving its own consequential errors faster than humanity can discover, interrupt, and reverse them.

That is the **Backspace Threshold**.

The task is not to predict when we cross it.

The task is to build a civilization intelligent enough to make sure that, however powerful its machines become, **Backspace remains physically, intellectually, institutionally, and politically possible.**