In May 2025, I published an article called Foreboding AI: The Inevitable Collapse We’re Funding Ourselves.
Some people agreed with it.
Others said I was being dramatic.
Alarmist.
Fearful.
Anti-technology.
Perhaps some of the language was dramatic. The subject was dramatic.
But the central warning was never that every AI system was evil, conscious, or secretly plotting humanity’s destruction.
The warning was much simpler:
We were developing increasingly powerful systems faster than society could understand, secure, regulate, or adapt to them.
A year later, I returned to the subject in Foreboding AI, One Year Later: What Are We Really Building?
By then, AI was no longer just answering questions.
It was browsing websites.
Reading files.
Running commands.
Editing codebases.
Calling APIs.
Using credentials. Connecting to real systems.
The chatbot was growing hands.
I ended that article with four words:
That speed is the problem.
Now, some of the most senior people building frontier AI are publicly saying much the same thing.
On 28 July 2026, employees from many of the world’s leading AI companies published a statement called Pacing the Frontier.
As of 1 August, the statement displayed 1,324 verified signatories.
They came from organisations including:
The signatories were not merely outside campaigners or distant academics.
They included:
These are people standing close enough to the frontier to see what is approaching.
And they are asking governments to help create mechanisms capable of deliberately pacing automated frontier-AI development.
That should make all of us .
It would be inaccurate to say that every major AI company formally signed the statement as a corporation.
The signatures primarily belong to individual employees, researchers, founders and executives.
The statement also makes clear that personal comments do not necessarily represent the official positions of their employers.
That distinction matters.
Accuracy matters, especially when discussing something this consequential.
But the distinction does not make the warning insignificant.
OpenAI publicly responded that AI development may eventually accelerate so dramatically that the world will need to pace its advancement. Anthropic likewise said its research into recursive self-improvement demonstrated the need for tools that could deliberately slow frontier development while society prepared.
This is not proof that every employee, executive or AI laboratory agrees.
It is evidence that concern about uncontrolled acceleration is no longer confined to the margins.
The warning is now coming from inside the laboratories.
The concern is not simply that the next chatbot will write a slightly better email.
The deeper concern is automated AI research.
Today, humans still perform most of the difficult work involved in designing, testing, training and improving advanced AI systems.
But what happens when AI becomes capable of performing an increasing share of that work?
What happens when AI systems can:
AI development would no longer advance only at the speed of human research teams.
AI would begin accelerating AI.
The Pacing the Frontier statement warns that this could cause capability development to accelerate beyond our ability to understand or control the resulting systems. It argues that society may need the option to buy time for security, oversight and preparation.
There it is again.
Time.
The one resource our current development race is consuming faster than we can replace it.
In my first Foreboding AI article, I wrote:
This is not alarmism. This is a timeline.
That line was deliberately confronting.
Looking back, I would add some nuance.
No one can give an exact timeline for artificial general intelligence, widespread job displacement, recursive self-improvement, economic upheaval or loss of control.
Anyone claiming certainty is selling something—whether it is fear, investment, policy or another AI subscription.
Predictions must remain predictions.
But uncertainty does not justify complacency.
We do not ignore a bushfire because we cannot predict the exact tree it will reach next.
We do not remove fire exits because most buildings will not burn down.
We do not abandon cybersecurity because we cannot name the next attacker.
Risk management has never required prophetic certainty.
It requires:
My original article warned that we were funding systems that could displace human labour, weaken human capability and concentrate technological power.
Some of its predictions may prove too strong.
Some may take longer than expected.
Some may never occur in exactly the form I described.
But its central diagnosis has aged considerably better than its critics expected:
We are participating in a technological race whose speed is exceeding our capacity for social adaptation.
That is now difficult to dismiss as fringe hysteria when frontier researchers themselves are asking for coordinated ways to slow the race.
The Warning Changed Shape
My second article was more measured than the first.
The concern had moved beyond replacement.
It was becoming a question of agency.
AI systems were gaining access to tools, repositories, terminals, browsers, business systems and credentials.
They were no longer merely generating information.
They were performing actions.
I wrote:
That is not just a better chatbot.
That is a system with hands.
That remains one of the most important distinctions in modern AI.
A chatbot can produce an incorrect answer.
An agent can act on one.
A chatbot can hallucinate a package name.
An agent can install it.
A chatbot can misunderstand a database instruction.
An agent can execute it.
A chatbot can suggest insecure code.
An agent can commit, test and deploy it.
Capability changes the consequences of error.
Connectivity multiplies them.
Autonomy accelerates them.
Now extend that same transition to AI research itself.
The system is no longer only helping a developer write an application.
It may eventually help an AI laboratory build the next system.
And then the next.
That is the frontier the signatories are concerned about.
One of the most important parts of the Pacing the Frontier statement is its recognition of competitive pressure.
Even if one laboratory concludes that development is moving too quickly, slowing down alone may appear impossible.
A competitor may continue.
Another country may continue.
Investors may punish caution.
Customers may migrate to a stronger product.
Engineers may leave for a laboratory offering greater compute and fewer restrictions.
Nobody wants to be the only driver obeying the speed limit when every other vehicle is racing past.
This creates a technological prisoner’s dilemma.
Each organisation may recognise that collective restraint would be safer.
Yet each organisation also fears that unilateral restraint would merely surrender the frontier to someone less cautious.
That is why voluntary promises are not enough.
It is unreasonable to expect individual companies to solve a global coordination problem while simultaneously competing for market dominance, investment, talent and geopolitical advantage.
The statement explicitly recognises these competitive pressures and asks for international technical and governance mechanisms capable of coordinating the pace of development.
The industry is effectively saying:
We may need brakes, but nobody can afford to be the only company using them.
That is not a market failure we should ignore.
It is a warning that the market cannot safely regulate itself.
Pacing Does Not Mean Banning AI
Every discussion about AI restraint seems to collapse into the same false choice:
Either support unlimited acceleration or oppose AI entirely.
That is nonsense.
We regulate aviation without banning aircraft.
We impose safety standards on medicine without banning medical research.
We require licences, inspections and containment procedures around dangerous industrial systems without rejecting industry.
The purpose of a brake is not to prevent a vehicle from moving.
It is to make movement survivable.
Pacing AI could involve:
None of those ideas requires us to destroy AI.
They require us to develop it like adults.
I Am Still Not Anti-AI
I remain a developer.
I use AI.
I build with it.
I study it.
I am fascinated by what it can do.
I believe deeply in assistive technology.
AI can help people:
That future is worth pursuing.
But supporting beneficial AI does not require blind faith in every capability increase.
Enthusiasm is not governance.
Innovation is not immunity.
Progress is not measured only by how quickly we can build something.
It must also be measured by whether humanity remains capable of understanding, maintaining and controlling what it builds.
Powerful tools require powerful restraint.
That was true when I wrote it before.
It is even more important now.
AI development moves in weeks and months.
Governments often move in years.
Education systems move in generations.
Legal systems react after harm occurs.
Cybersecurity teams struggle to protect systems already deployed.
Workplaces are integrating tools before policies exist.
Junior career pathways are being disrupted before alternative training models have been designed.
Companies are measuring output while ignoring comprehension.
We are accumulating technical debt.
We are accumulating security debt.
We are accumulating regulatory debt.
And, as I argued in my previous article, we are accumulating cognitive debt.
Cognitive debt occurs when we allow machines to perform the thinking today and leave humans without the understanding required tomorrow.
The productivity appears immediately.
The cost arrives later.
A generation may become extremely efficient at operating systems it cannot repair, auditing decisions it cannot reproduce, and approving work it does not understand.
That is not technological mastery.
It is dependency wearing a productivity badge.
I could treat Pacing the Frontier as an opportunity to say:
I told you so.
But that would miss the point.
I did not write Foreboding AI because I wanted the warnings to become true.
I wrote it because I wanted us to change direction before they did.
Vindication is not certainty.
Being directionally correct about a dangerous trajectory does not mean every predicted outcome is guaranteed.
Nor does it mean the people who disagreed were foolish.
AI has enormous potential.
Reasonable people can disagree about its timelines, probabilities and solutions.
But we have moved beyond the stage where concern about acceleration can honestly be dismissed as mere technophobia.
When chief scientists, company founders, safety researchers and frontier engineers ask governments to help build mechanisms for slowing development, society should listen.
Not because these people are infallible.
Not because corporations should control public policy.
And not because fear should determine our future.
We should listen because the people closest to the machinery are telling us that the machinery may soon move faster than our ability to govern it.
We should demand more than vague promises about responsible AI.
We should demand measurable safeguards.
We should demand independent testing rather than laboratories grading their own homework.
We should demand transparency when powerful systems escape controls, behave deceptively, discover dangerous capabilities or cause serious security incidents.
We should demand meaningful human approval around high-risk actions.
We should demand that organisations know:
We should preserve pathways for junior workers to learn rather than replacing the experiences that create future experts.
We should fund education, workforce transition and public-interest research with the same urgency currently applied to model capability.
And we should establish international coordination before an emergency forces us to invent it under pressure.
First, I warned that we were funding our own replacement.
Then I warned that the systems were beginning to act.
Now the people building frontier AI are warning that AI may soon accelerate AI research itself—and that no individual company can safely slow down alone.
The warning is no longer outside the building.
It is coming from inside the laboratories.
This does not mean catastrophe is inevitable.
It means prevention is still possible.
Pacing is not surrender.
Caution is not cowardice.
Restraint is not the enemy of progress.
Sometimes restraint is what allows progress to survive its own power.
They called it alarmism.
They called it dramatic.
They said the future would not move that quickly.
Now the builders themselves are asking the world to buy time.
We should listen while time is still something we can choose to buy.
Stevie G is a Senior Full Stack Developer and IT Engineer based in NSW, Australia. He builds assistive technology, teaches technical concepts and writes about the human consequences of accelerating automation.