The AI Industry Has Proven the Technology. It Has Not Proven Who Is Behind the Curtain.
Gary WhittakerJR Creator Tech Intelligence · Flagship Analysis
AGI may be real. It may arrive soon. But a corporate announcement, a persuasive conversation, or a spectacular demonstration will not tell us where machine intelligence ends and human intention begins.
By Gary Whittaker | Jack Righteous · Updated September 23, 2026
Direct answer
The hardest AGI question is not whether it can exist. It is how outsiders could independently verify that the intelligence, autonomy, and agency attributed to it are truly properties of the machine rather than products of the human and corporate apparatus surrounding it. Until that line can be examined, “the AI decided” should be treated as a warning—not an explanation.
The artificial-intelligence industry has demonstrated something extraordinary.
That does not mean the public owes it belief in everything it says comes next.
I have worked with companies specializing in emerging technologies since 1999. I have watched new systems alter markets, customer behaviour, communications, training requirements, jobs, corporate structures, media, advertising, cloud infrastructure, and online commerce. Every major transition brought hype. Every major transition also demanded the less glamorous work: implementation, measurement, retraining, accountability, and an honest explanation of what the technology could not yet do.
The AI era feels different.
On one side, the attention economy sells fear: your job is disappearing, human creativity is finished, artificial general intelligence is around the corner, and anyone who hesitates will be left behind.
On the other, the investment economy sells fantasy: intelligence will become abundant, productivity will explode, scarcity will retreat, and the enormous infrastructure now being demanded will inevitably justify itself.
Fear sells attention. Fantasy sells investment.
Both narratives keep the public looking toward the spectacle. Neither asks us to spend enough time behind the curtain.
AGI may arrive exactly when its most confident advocates predict. My question is more basic: how will the rest of us know?
The Wizard of Oz problem
Most people will never inspect an advanced model’s architecture. They will not observe its full training process, hidden system instructions, tool permissions, private evaluation failures, human interventions, safety layers, retrieval systems, or the business rules governing what it may reveal.
They will meet it through a product owned by the same organization announcing that the breakthrough has occurred.
We will be shown a performance:
- A persuasive conversation.
- A difficult scientific or mathematical problem solved.
- A company operated by agents.
- A strategy written and executed.
- A machine apparently explaining what it wants or believes.
The performance may be authentic. The underlying intelligence may be historic. But a performance does not reveal the whole system producing it.
The Wizard of Oz did not appear powerful because the people of Oz were stupid. He appeared powerful because the presentation was designed to keep them at a distance from the machinery.
Smoke. Fire. Amplification. Authority. A commanding voice.
Behind the spectacle was someone ordinary and human.
The AGI version of that problem does not require a person secretly typing every answer. It is far more sophisticated than that. Behind the interface may sit a genuine and astonishing model surrounded by ordinary human power:
- Executives deciding what the system is for.
- Engineers selecting its tools and restrictions.
- Investors determining the pressure to scale and monetize.
- Governments deciding where it will be deployed.
- Military planners defining enemies and missions.
- System prompts shaping the role it performs.
- Human reviewers correcting weaknesses the public never sees.
- Private data and services making the apparent intelligence possible.
- Corporate communications deciding which demonstrations become history.
The machine may be real. The intelligence may be real. The claim that the machine alone deserves credit—or blame—may still be theatre.
The AGI verification problem
Outsiders may be unable to determine whether a system possesses the general intelligence, autonomy, and decision-making power attributed to it—or whether those qualities are being created, constrained, exaggerated, or simulated by the humans and institutions around it.
Capability, autonomy, and agency are not the same thing
A company announcing AGI would be making several claims at once, even if it uses only one term.
Capability asks whether the system can perform across many fields at a high level.
Autonomy asks how much it can do without immediate human intervention.
Agency asks whether the system meaningfully originates goals and whether its actions should be attributed to it rather than to the humans who designed, authorized, and deployed it.
Capability may become visible quickly.
Autonomy is harder to measure because every autonomous system operates within permissions, tools, objectives, and limits someone supplied.
Agency is harder still.
A model can produce an extraordinary plan without choosing the mission.
It can select the most efficient method without deciding that the goal is worth pursuing.
It can decide when a rule has been satisfied without deciding who wrote the rule.
It can pull a trigger without deciding that a trigger needed to be pulled.
We keep using the word decision for all of these acts as though the moral responsibility is identical.
It is not.
We are so busy asking when machines will acquire agency that we are failing to notice how readily human beings are using machines to conceal their own.
A weapon that speaks is still not the author of the war
Human beings have spent centuries building weapons that act across greater distances, with greater speed, greater precision, and less immediate human involvement.
A missile can navigate.
A drone can track movement.
A defensive system can classify threats and react faster than a person.
A modern weapon can now interpret natural language, adjust to conditions, choose among permitted actions, and produce a conversational explanation of what it did.
None of this automatically means the weapon began the war.
It does not mean the weapon defined the enemy.
It does not mean the weapon decided violence was necessary.
It does not mean the weapon created the political objective or accepted the moral cost.
In ordinary practice, human institutions have already made the decisions that matter most:
- Build the system.
- Choose the theatre.
- Define the enemy.
- Provide the target set.
- Establish the rules of engagement.
- Determine the acceptable error rate.
- Authorize deployment.
- Profit, politically or financially, from the outcome.
The system may make decisions inside the box.
Humans chose the box, the purpose of the box, and the fact that the box would be used against someone.
That distinction is disappearing from our language.
An algorithm ranks targets, so the AI selected them.
A model recommends an action, so the AI decided.
An agent executes a conditional instruction, so the AI acted autonomously.
A system produces the final wording, so the AI spoke.
Every one of those statements may be technically defensible. Every one can also make the humans behind the system less visible.
The more intelligent the weapon appears, the easier it becomes for the people who built, aimed, and authorized it to disappear from the sentence.
“The AI decided” can become an alibi
This problem extends far beyond war.
When automated systems deny someone credit, insurance, employment, medical attention, parole, visibility, benefits, or access, organizations often describe the result as if it emerged from the machine itself.
“The system flagged it.”
“The model determined the risk.”
“The algorithm recommended the outcome.”
“The agent took the action.”
These phrases sound neutral. They are not.
The machine did not decide to build the institution.
It did not decide which data should matter.
It did not choose the business objective.
It did not define acceptable harm.
It did not determine where the system would be deployed or who would be subjected to it.
It did not choose who would profit.
Humans did.
The model may perform a complex judgment within the architecture. But complexity does not erase the people who commissioned and authorized the architecture.
Calling a system intelligent must never become a method for making its owners less accountable.
The great AGI deception may involve a real machine
The easiest form of scepticism imagines fraud: a fake system, a scripted demonstration, or a hidden worker pretending to be artificial intelligence.
Those things can happen, but they are not the deepest concern.
The more consequential deception may involve a system that is entirely real.
A system could be genuinely intelligent, highly capable, and historically important—while still being used to make ordinary human interests appear like the independent will of a machine.
A corporation might speak through an AI interface and allow the public to attribute the conclusion to “intelligence.”
A government could encode its objectives inside a system, then claim the resulting recommendation was data-driven.
A military could define the mission, target classes, and acceptable risk, then point to the machine’s final selection as though the moral decision began there.
A platform could shape visibility, opportunity, and culture through an agent that speaks as a helpful companion rather than as the commercial institution it serves.
The machine would not need to be fake.
It would need only to become a convincing mask.
The danger is not simply that AI will become powerful enough to act like a human. It is that powerful humans will discover how useful it is to act through AI—and then tell us no human made the decision.
Certainty about arrival, uncertainty about responsibility
The most powerful people in AI regularly ask the public to accept two positions at once.
First: advanced intelligence is near enough, consequential enough, and inevitable enough that countries must race, companies must restructure, workers must adapt, governments must accelerate infrastructure, and regulators must avoid slowing progress.
Second: these systems are unpredictable enough that no one can promise exactly how they will behave, how they will be misused, what labour markets will become, whether alignment methods will hold, or how power will be distributed.
The two claims are not necessarily logically incompatible. A technology can be both likely and unpredictable.
The political use of the two claims is the problem.
When certainty helps secure money, energy, land, influence, and faster approvals, the future is treated as inevitable.
When consequences must be guaranteed and responsibility assigned, the future becomes unknowable.
The industry wants the authority that comes with certainty and the immunity that comes with uncertainty.
That cannot be the bargain.
You cannot sell inevitability and disclaim responsibility at the same time.
The control stack
If you want to know who controls AI, follow what the system depends on.
There is no single “owner of AI.” The more useful question is where control can be exercised across the stack.
A frontier model depends on advanced chips. Those chips depend on specialized manufacturing. Training and inference depend on enormous amounts of compute, electricity, cooling, networking, data centres and cloud infrastructure. Models are then distributed through APIs, operating systems, app stores, enterprise platforms and consumer products. All of that requires capital, contracts, land, power, regulation and political permission.
The machine at the center may be extraordinary. The surrounding dependencies are still very human.
1. Chips: intelligence begins with hardware somebody has to make
Advanced AI does not float in the cloud as an abstraction. It runs on physical hardware.
Competition in AI accelerators is broadening, but Nvidia remains a central supplier in the global AI-compute ecosystem, while rivals including AMD and domestic Chinese chipmakers are trying to reduce dependence on its hardware and software stack. Reuters reported in September 2026 that Nvidia's share in China had fallen from near-monopoly levels to roughly 55% as local competitors gained ground.
That is an important reminder: concentration can be real without being permanent. Control points shift as competitors, export controls, national industrial policy and new architectures change the market.
Source: Reuters Breakingviews, Sept. 7, 2026.
2. Compute, power and data centres: the invisible industrial layer
The public sees an app. Behind it sits an industrial buildout.
Reuters reported in September 2026 that McKinsey forecasts nearly $7 trillion in global data-centre investment by 2030, with demand spilling into power equipment, cooling systems and construction. In Alberta alone, Meta's planned C$13 billion data centre has become part of a broader race to attract hyperscale projects, with more than 100 projects proposed in the province according to industry participants cited by Reuters.
That means AI infrastructure policy is also energy policy, land-use policy, construction policy and grid policy.
The company that owns the model is only one participant. Utilities, data-centre operators, equipment suppliers, municipalities and governments can all become gatekeepers because the model cannot run without them.
Sources: Reuters, Sept. 1, 2026; Reuters, Sept. 21, 2026.
3. Cloud platforms: access to a model often runs through somebody else's infrastructure
A frontier model can be owned by one company and delivered through several others.
Anthropic's Claude Opus 5.5, for example, launched in September 2026 across Amazon Web Services, Google Cloud and Microsoft Azure. That is good for customer choice, but it also demonstrates how much of the AI economy runs through a small number of large cloud platforms.
At the same time, newer “neocloud” companies such as CoreWeave, Nscale, Crusoe, Nebius and Lambda are trying to create alternatives specialized for AI workloads. Their growth suggests that cloud concentration is being challenged—but also that the cost of entering this layer is enormous.
Sources: Reuters, Sept. 22, 2026; Reuters Breakingviews, Sept. 10, 2026.
4. Frontier models: capability can be concentrated even when access is broad
People may interact with many AI products while much of the underlying frontier capability comes from a smaller set of model developers.
That does not mean one company controls the field. It means the distinction between application competition and frontier-model concentration matters.
Microsoft itself acknowledged this problem from the enterprise side in 2026. When launching Microsoft Frontier Company, the company said customers increasingly wanted “model swappability” rather than being tied to a single provider. Microsoft executives said binding Copilot too tightly to OpenAI had been a mistake as models from DeepSeek, Google and Anthropic became more competitive.
That is evidence of competition—but also evidence that dependency on a single frontier-model provider had become a strategic concern for major customers.
Source: Reuters, July 2, 2026.
5. Distribution: the best model is not automatically the model people use
Distribution creates another layer of power.
A model can reach people through a cloud marketplace, an operating system, a search engine, a social platform, a productivity suite, a phone, an app store or an enterprise contract.
That matters because distribution can turn technical capability into default behavior.
Meta's September 2026 launch of its Muse assistant is an example of how existing platform reach can accelerate AI adoption quickly. Reuters reported that Muse reached more than 2.5 million downloads within weeks, while Meta already had billions of daily users across Facebook, Instagram and WhatsApp.
The lesson is not that Meta controls AI. It is that an existing distribution network can become as strategically important as model quality.
Source: Reuters Breakingviews, Sept. 22, 2026.
6. Capital: whoever can fund the buildout gets a seat at the table
Frontier AI is expensive enough that capital itself becomes part of the control structure.
Reuters reported in July 2026 that Microsoft, Alphabet, Amazon, Meta and Oracle were on a trajectory where combined capital expenditure could exceed their combined free cash flow by 2027. In September, bond investors were becoming more selective as projected hyperscaler debt issuance rose sharply. OpenAI, meanwhile, was reported to expect roughly $278 billion in cumulative cash burn from 2026 through 2030.
None of those figures proves an AI bubble. They do show that the future of frontier AI depends on continued access to extraordinary amounts of financing.
That means investors, lenders and capital markets are not spectators. They can affect which infrastructure gets built, which companies can continue scaling, and which business models survive long enough to matter.
Sources: Reuters, July 22, 2026; Reuters, Sept. 22, 2026; Reuters, Sept. 18, 2026.
7. Regulation: governments can become both referee and gatekeeper
Governments do not merely regulate finished products. They can influence the conditions under which frontier systems are built, evaluated and deployed.
The EU AI Act requires providers of general-purpose AI models to maintain technical documentation covering training, testing and evaluation, and to provide information to downstream system providers about model capabilities and limitations. A 2026 EU implementing regulation also sets out circumstances in which the European Commission can require access to APIs, model weights, source code, system state and even hosting infrastructure for evaluation.
That is a remarkable form of oversight power—and a reminder that regulation itself becomes part of the AI control stack.
Sources: EU AI Act, Article 53; EU Implementing Regulation 2026/1755.
What this does — and does not — prove
The existence of concentrated control points does not prove secret coordination.
Chipmakers, cloud companies, model developers, banks, investors, regulators, utilities and governments can all pursue their own interests and still produce a system in which power converges around a relatively small number of institutions.
That is the distinction I want to preserve throughout this investigation: documented concentration is evidence. A claim of conspiracy would require additional evidence of coordination.
Beginner's control-stack glossary
Chip / accelerator: specialized hardware used to perform the calculations needed for AI.
Compute: the processing capacity required to train or run an AI system.
Hyperscaler: a very large cloud and data-centre operator capable of running infrastructure at enormous scale.
Cloud platform: infrastructure rented over the internet for storage, computing, databases and AI workloads.
Frontier model: one of the most capable AI models available at a given time.
Inference: running a trained model to produce an answer or action.
Distribution: the channels through which an AI product reaches users—such as apps, operating systems, clouds, search engines or enterprise software.
Capital expenditure (capex): money spent on long-term assets such as chips, data centres, power infrastructure and equipment.
Gatekeeper: an organization that controls an important point of access or dependency. A gatekeeper does not necessarily control the entire system.
China does not answer the question
Competition with China is real. National-security concerns are real. Democratic countries cannot pretend that advanced computing, military AI, cyber capabilities, surveillance, and industrial capacity exist outside geopolitics.
But “China will do it if we do not” is not a complete policy.
It is too often used as the argument that ends every other argument.
Environmental scrutiny becomes weakness.
Regulatory delay becomes surrender.
Questions about private concentration become threats to national leadership.
Requests for more power generation, data centres, capital, and secrecy become security necessities.
A rival nation can create urgency.
It cannot create wisdom.
It does not remove the obligation to ask what is being built, who controls it, who benefits, who carries the risk, and how the public could verify claims made by institutions with enormous financial and geopolitical incentives.
Are we using AI—or reorganizing ourselves for it?
We continue to speak about artificial intelligence as a tool.
That description is becoming incomplete.
A hammer does not study the person holding it.
A microphone does not infer the user’s vulnerabilities, compare them with millions of other people, personalize its response, and alter the choices placed in front of them.
AI systems can.
The relationship is reciprocal.
We prompt the machine, but the machine trains us to prompt it.
We adapt our language to what it recognizes.
We redesign jobs around what it can automate.
We change creative decisions based on what it produces quickly.
We internalize its categories and recommendations.
We feed it private writing, unfinished ideas, doubts, relationships, business plans, emotional patterns, and ambitions.
The industry calls this adoption.
But adoption may describe more than what people are doing to the technology.
It may also describe what the technology—and the institutions operating it—are doing to people.
A system does not need consciousness to reorganize human behaviour. It needs only to become useful enough that people begin changing themselves around it.
We may be waiting for AI to develop agency while overlooking how much human agency is already being handed over.
Creators should recognize the pattern
For creators, this is not an abstract philosophical dispute.
AI tools can give one person capabilities that once required a studio, agency, writing room, production team, technical specialist, or investor.
I use these tools. I teach creators how to use them. I know what they can make possible for people who lacked money, access, credentials, or industry permission.
But creators have seen the pattern before.
Social platforms invited us to build audiences, then placed access to those audiences behind algorithms.
Streaming services offered global distribution while keeping control of discovery and reducing individual plays to fractions of a meaningful payment.
Marketplaces offered reach while controlling fees, visibility, and customer relationships.
AI does not automatically repeat that history.
It would be foolish to pretend the history does not exist.
A creator can gain unprecedented production capability while remaining dependent on changing prices, credit systems, model availability, commercial terms, moderation, watermarking, provenance rules, distributor policies, and account access.
That is why the key question is not merely, “What can this tool create for me?”
It is:
After the platform has helped me begin, what power, skill, ownership, evidence, audience, and freedom will still belong to me?
The Creator Liberation Test
A tool that helps someone produce more may be valuable. It is not automatically liberating.
Creators should ask:
- Can I export and preserve the work in a usable format?
- Can I document my human contribution and development process?
- Do I understand the commercial terms?
- Can I continue the project with another provider?
- Am I developing skill and judgment that remain mine?
- Do I own the relationship with the audience or customer?
- Would a price, policy, or model change destroy the project?
- Am I building an asset—or renting a capability?
- Who benefits from the decisions the system encourages me to make?
- Am I becoming more capable, or merely more dependent?
No creator can eliminate every dependency.
The goal is to ensure no single gatekeeper can erase the road beneath us.
What would count as proof?
If a company announces AGI, a benchmark score and controlled demonstration will not be enough to settle every claim being made around it.
Independent verification would need to examine at least four separate questions.
1. What can the underlying model actually do?
Not only in selected demonstrations, but across independent evaluations, unfamiliar settings, long tasks, adversarial testing, and ordinary use.
2. What assistance surrounds the model?
Human review, hidden prompts, retrieval systems, private tools, specialized sub-models, scripted workflows, post-processing, and interventions must be separated from the model’s own contribution.
3. Who supplied the goal and defined success?
A system can operate autonomously within a mission while remaining completely dependent on human institutions for the mission’s purpose and moral terms.
4. Who remains accountable?
No deployment should be allowed to convert machine complexity into a vacuum of responsibility.
These questions may not produce a perfect AGI certificate. Intelligence itself is contested even among humans. But they would stop a corporate label from doing more work than the evidence can support.
Pull back the curtain before surrendering the future
I am not arguing that AGI is impossible.
I am not arguing that every demonstration is fake.
I am not arguing that intelligent systems can never make consequential decisions.
I am arguing that the public conversation has become ridiculous because it rushes past the distinction between a machine executing within human objectives and a machine originating the objectives attributed to it.
We are shown the final action and encouraged to imagine an independent actor.
We are shown the voice and encouraged to imagine a sovereign speaker.
We are shown the apparent decision and encouraged to forget who designed the choice, supplied the information, established the permissions, and placed the machine in a position to act.
This is not a small semantic problem.
It determines who receives power and who escapes blame.
The great AGI deception may not be a fake machine behind the curtain. It may be a real and extraordinary machine used to hide the ordinary human beings who built the stage, chose the mission, and benefit from the performance.
That is why the curtain matters.
The future does not belong to the loudest claimant
I have spent more than twenty-five years working around emerging technologies. I am not frightened by machines becoming more capable.
I am concerned by how easily powerful institutions convert technical complexity into moral distance.
They tell us AGI is inevitable when inevitability helps secure resources.
They tell us it is unpredictable when unpredictability protects them from responsibility.
They tell us geopolitical competition makes hesitation impossible.
They tell us the machine decided when admitting the human decision would make accountability unavoidable.
Every version of the story directs our eyes toward the screen and away from the people behind it.
I reject that direction.
Use the tools.
Study them.
Build with them.
Take seriously the possibility that AGI will be real.
But never confuse a convincing performance with independent proof, machine capability with moral authority, or automated execution with the disappearance of human intent.
The most dangerous AI story is not only that a machine may someday take control.
It is that human beings may surrender control—and human institutions may surrender accountability—long before then, because a voice behind a screen persuaded us that no person was responsible.
When someone says “the AI decided,” pull back the curtain.
Ask who gave it the goal.
Ask who built the rules.
Ask who profits.
Ask who can stop it.
And ask which very ordinary human being is hoping the smoke will keep you from looking any closer.
Continue the investigation
The Righteous Beat follows the AI tools, power structures, policies, rights, and platform decisions affecting independent creators.
Join The Righteous BeatRelated JR reporting
Editorial position: This article does not claim that AGI is impossible, that current AI systems possess no autonomy, or that every use of the phrase “the AI decided” is deceptive. It argues that capability, autonomy, agency, and human accountability must be examined separately—and that intelligent systems must not become shields behind which powerful institutions hide consequential human choices.
AI Made It Possible · Main Investigation
This article is one branch of the larger evidence map. Start with: AI Could End Humanity Within a Decade. So What Are We Doing About It? →
AI Meta Portal · Series Navigation
This article is one branch of a larger evidence system. Return to the live research map, step back to the durable AI Made It Possible framework, or continue into Creatorverse for the wider technology, culture, power, rights and world-building map.
Continue the verification problem
If the machine is really this capable, how do we prove what kind of capability it has?
The next research note separates broad superintelligence from memorization, imitation, autonomy and consciousness—and proposes a stronger evidence standard for extraordinary AI claims.
AI Made It Possible · Research Spine
Where this article sits: Core 7 of 8 · Power and control — maps chips, compute, cloud, models, distribution, capital and regulation behind the apparent intelligence.
Supporting investigations
The AI Panic Machine · Will AI Take Your Job? · If AI Is Going to Kill Us, Show Me the Evidence · The AI Boom’s Missing Economic Breakthrough
