Artificial Intelligence
and Humanity

This matters to every person
on Earth right now

The debate about artificial intelligence is usually framed as a future problem — something to worry about when the robots arrive. That framing is wrong. The decisions that will determine AI's relationship with humanity are being made today, in boardrooms and server farms and government offices, mostly out of public view.

Right now, five private companies — Amazon, Google, Meta, Microsoft, and Oracle — control 71% of the computing infrastructure that runs the AI systems the world increasingly depends on.^[1] That number was 63% eighteen months ago. The trend is accelerating. By the time most people become aware of this concentration, the structural decisions will already be locked in.

This case study does not pretend there is a simple answer. AI has already saved lives, accelerated science, and made previously impossible things ordinary. It has also concentrated power in ways that history tells us always produce harm unless deliberately corrected. Both of these things are true at the same time.

What follows is an honest account of both sides — the genuine good that AI is already doing, the genuine risks that are already unfolding, and the structural question about who controls the infrastructure that most mainstream coverage still avoids.

When a company that has raised billions of dollars to build AI systems says, publicly, that the technology threatens human control, it is worth paying attention. This is not a fringe concern. It is the considered judgment of people who are building these systems from the inside.

The real benefits are
already documented

Before examining the risks, it is important to be honest about what AI has already genuinely delivered. Some of these achievements represent decades-long scientific problems solved in months. They are not hypothetical future benefits — they are documented, peer-reviewed, and happening now.

Medicine: the protein folding breakthrough

For fifty years, one of biology's hardest problems was predicting how a protein folds into its three-dimensional shape. The shape determines its function. Knowing the shape is essential for drug discovery. Human researchers could determine perhaps a few hundred protein structures per year with enormous effort.

In 2021, DeepMind's AlphaFold2 solved this problem at scale. By 2022, the system had predicted the structures of over 200 million proteins — essentially every known protein in biology — and made the database freely available to any researcher in the world.^[S1] Researchers who had spent careers on single proteins suddenly had access to the entire biological library. Drug discovery for diseases including Parkinson's, malaria, and antibiotic resistance was transformed overnight.

Cancer detection: outperforming human specialists

In a study published in Nature in January 2020, Google Health's AI system detected breast cancer with an 11.5% reduction in false negatives and a 5.7% reduction in false positives compared to radiologists — even when the AI was working from scans alone, without the patient history that human doctors use.^[S1]

AI-powered diabetic retinopathy screening has been deployed at scale in Thailand and India, detecting vision-threatening conditions with over 90% sensitivity in settings where specialist ophthalmologists are too few to meet demand. The technology brings the diagnostic capability of a specialist to a clinic that cannot afford one.

Drug discovery: compressing timelines

The typical journey from molecule to clinical trial takes four to five years. Insilico Medicine, using AI to design drug candidates, moved a treatment for idiopathic pulmonary fibrosis — a fatal lung disease — from initial design to Phase 2 clinical trials in approximately 30 months, while reducing costs by an order of magnitude. Recursion Pharmaceuticals is currently scanning 2.2 million compound-disease combinations per week, a rate no human team could approach.^[S1]

Climate and energy

DeepMind's AI reduced the energy consumed by Google's data center cooling systems by 40% — verified and ongoing since 2016. Applied across the world's data centers, this represents billions of kilowatt-hours of electricity per year.^[S1]

In 2023, Google and NOAA released GraphCast, an AI weather forecasting model that outperformed the European Centre for Medium-Range Weather Forecasts — humanity's best numerical weather prediction system — on more than 90% of the 1,380 atmospheric variables it was tested on. More accurate weather forecasting at a fraction of the computational cost.

Accessibility: what was impossible is now ordinary

For a blind person, the world was largely inaccessible text. Microsoft's Seeing AI app now reads text, identifies people by face, describes physical scenes, and narrates currency denominations in real time. Google's Project Euphonia creates synthetic voices for people with ALS and motor neuron disease, trained on recordings of their own voice made before the disease progressed. People who were losing the ability to speak now have a voice that sounds like themselves.

Live captions in more than 20 languages are now standard features in video conferencing. Real-time translation across 7,000+ language pairs has made cross-language communication available to anyone with a smartphone, in contexts that previously required professional interpreters.

Productivity: the honest picture

A randomized controlled trial with 95 software developers found that those with AI coding assistance completed tasks 55.8% faster.^[S1] In legal practice, AI document review reduces the time to review large document sets by 85%, at roughly one-tenth the cost per document. These are not marginal gains.

The concerns are real,
not hypothetical

The same technology producing those benefits is also producing genuine harms and genuine risks. An honest account requires looking at both with equal seriousness. The following are not worst-case scenarios invented for dramatic effect — they are documented, sourced, and in several cases already underway.

Job displacement: what the numbers actually say

Goldman Sachs estimated in 2023 that generative AI could automate the equivalent of 300 million full-time jobs globally. The IMF's 2024 assessment found that 40% of jobs worldwide are exposed to AI automation — and in advanced economies, the figure rises to 60%.

The honest nuance here is important. AI also creates jobs — it always has, and the net employment effect of previous technology revolutions has been positive over long periods. The problem is the transition. The Industrial Revolution displaced skilled artisans faster than factory employment grew. The period of net job loss lasted decades. For the people living through a transition period, "it will be positive in the long run" is cold comfort.

What is different this time: previous automation replaced physical labor. AI is replacing cognitive labor — the category of work that historically absorbed the workers displaced by physical automation. The retraining pathways are less clear than they were in past transitions.

Power concentration: a historical comparison

Standard Oil, at the height of John D. Rockefeller's monopoly, controlled approximately 65% of US oil refining. The federal government eventually broke it up because that concentration was judged incompatible with a democratic economy. Five companies currently control 71% of the computing infrastructure that runs global AI — and that share is growing.^[1]

The comparison is not perfect, but it is instructive. Oil was physical infrastructure. What these companies control is cognitive infrastructure — the systems that increasingly mediate how people access information, make decisions, and interact with institutions. The power differential is arguably greater, not lesser.

Surveillance: what is already deployed

China's surveillance infrastructure, built on centralized AI facial recognition and behavioral tracking, monitors over one billion people. The Uyghur detention system used AI to flag individuals for surveillance based on patterns of movement and behavior. These are not hypothetical capabilities — they are documented deployments, extensively reported and verified.

Surveillance is not only an authoritarian government problem. Amazon's Ring doorbell network has created what civil liberties organizations have described as a de facto distributed surveillance infrastructure across American neighborhoods — with footage routinely shared with law enforcement without warrants or user notification. Commercial data brokers in the US maintain profiles of over 250 data points per adult. AI makes these profiles actionable for targeting in ways that were not previously possible.

Algorithmic decisions affecting real lives

The COMPAS recidivism algorithm used in US courts to predict whether a defendant would re-offend was found by ProPublica in 2016 to be nearly twice as likely to falsely flag Black defendants as high-risk compared to white defendants. Algorithmic systems are making or influencing decisions about bail, credit, hiring, and housing — decisions that have material effects on people's lives — with limited human oversight and often without the people affected knowing an algorithm was involved.

Microsoft discontinued its AI recruiting tool in 2018 after discovering it was penalizing applications that included the word "women's" — as in "women's chess club." The same bias patterns are present across AI hiring tools because they are trained on historical hiring data, which reflects historical discrimination.

Disinformation at industrial scale

In the New Hampshire primary in January 2024, a robocall using an AI-synthesized voice mimicking President Biden told Democratic voters not to vote. This was not a sophisticated attack — it was done cheaply, quickly, and with consumer tools. AI-generated images, video, and audio of political figures circulated in the 2024 Indian, Bangladeshi, and US elections. The cost of producing disinformation has dropped from millions of dollars — the budget required for professional filmmaking — to near zero.

Five companies.
71% of the infrastructure.
Honestly assessed.

Centralized AI means AI development concentrated in large organizations — OpenAI (backed by Microsoft), Google DeepMind, Anthropic (backed by Amazon), xAI (Elon Musk), and Meta — with massive computing clusters, unified training pipelines, and models that hundreds of millions of people use through a single access point.

This is how most people currently experience AI. When you use ChatGPT, Google Search, or Gmail's AI features, you are using centralized AI. Understanding what that means — genuinely — requires looking at both why it works and what it costs.

What centralized AI genuinely does well

Raw capability. The most powerful AI systems in the world — GPT-4, Gemini Ultra, Claude — were built through centralized training runs on massive computing clusters. This is not an accident. The physical laws of how machine learning works currently favor large, coordinated training. You cannot get the same results by distributing the process across many small nodes. Centralized AI wins on capability, and that capability is what delivers the medical breakthroughs and productivity gains described earlier in this document.

Coordinated safety enforcement. When a safety researcher at Anthropic discovers that a model can be prompted to produce dangerous content, Anthropic can update the guardrails globally and push the fix to every user in a single release. One entity, one update, worldwide effect. This is genuinely useful for safety. It is also a structural argument for centralized control that deserves honest acknowledgment even in a critique of centralization.

Accountability. If OpenAI deploys a system that causes harm, there is a legal entity to sue, regulate, or sanction. Governments know who to call. Courts know who to hold responsible. This accountability surface matters — and it disappears when you distribute the same capability across thousands of anonymous nodes.

The real dangers: what the numbers mean in practice

Standard Oil at the height of its monopoly controlled approximately 65% of US oil refining. The government eventually broke it up because that concentration was deemed incompatible with a democratic economy. The current concentration in AI compute is higher. And oil is physical infrastructure — pipes and refineries. What these companies control is cognitive infrastructure: the systems increasingly mediating how people think, search, and decide.

The "industrial capture" scenario is not a conspiracy theory. It describes a structural problem: when the safety of AI systems depends on the incentives of the companies building them, and those companies have financial incentives to accelerate deployment, the safety measures depend on good intentions rather than structural constraints. Good intentions change. Incentives do not.

The whistleblower letters from inside these companies are important data points. The people who understand these systems best, who have access to internal safety evaluations and model capabilities that are not publicly disclosed, are saying they are worried and that they cannot speak freely. That is not evidence of imminent catastrophe. It is evidence of a structural problem in how safety information flows from the companies that have it to the public and regulators who need it.

The energy trajectory is worth noting as a concrete indicator of trajectory. xAI's Colossus supercomputer in Memphis, Tennessee, draws 280 megawatts — roughly the electricity consumption of a small city, and more than twenty times the power draw of the Oak Ridge National Laboratory's Summit supercomputer in 2019.^[1] AI supercomputer power requirements are doubling every thirteen months. The physical concentration of that infrastructure — in Memphis, in Northern Virginia, in Iowa — is a geopolitical fact as much as a technical one.

The alternative: distributed,
open, and honestly imperfect

Decentralized AI is AI development distributed across many independent actors with no single controlling entity. It includes open-source models whose weights anyone can download and run, federated learning systems where AI is trained on data that never leaves local devices, and blockchain-based AI networks where thousands of participants collectively provide computing power and AI services.

Decentralized AI is often discussed in utopian terms — the democratization of AI, power to the people, no central authority to shut it down. The reality is more complicated and more interesting.

What decentralized AI genuinely solves

Privacy by architecture, not by policy. When your phone's AI keyboard learns your typing patterns without sending data to a central server, your private communications cannot be breached in a data breach that doesn't exist. Google's Gboard keyboard has used federated learning for years — billions of devices, zero raw data transmitted. The multi-hospital federated model described earlier — AUC 0.85 for patient mortality prediction, across six hospitals, without any patient record crossing an institutional boundary — is proof that serious AI capability can be built without centralizing sensitive data.^[16]

Cannot be shut down by any single authority. Meta's Llama 3 model has been downloaded hundreds of millions of times. It runs on consumer laptops, hospital servers, and research institutions in over 100 countries. No government order, no corporate policy change, no board decision can remove it from the world. For people in authoritarian contexts, or for minority-language communities whose needs corporate AI filters ignore, this matters enormously.

Distributed value creation. Bittensor — a blockchain network with 128 specialized AI subnets — distributes the economic value of AI provision to the people who contribute models and computing power, not to the shareholders of one corporation. Subnet count grew 97% in 2025. The economic model is genuinely different: anyone with a graphics card and an AI model can participate and earn.^[16]

Transparent and auditable. When a model's weights are publicly available, security researchers, governments, and academics can examine them. Hidden capabilities cannot stay hidden. Bias patterns can be documented by parties with no financial stake in minimizing them. Community-discovered vulnerabilities get fixed in public view, with credit going to the discoverer rather than being silently patched by a corporate security team.

The risks that honest advocates acknowledge

Here is the part that advocates of decentralized AI often underemphasize: removing centralized control also removes centralized safety enforcement. When anyone can modify a model and redistribute it, the safety properties of the original model are not preserved. Researchers have documented that open-weight models can be fine-tuned in under an hour to remove all alignment constraints — the guardrails that prevent them from producing harmful content. The technical capability to do significant harm with an ungoverned open-weight model is real and documented. This is not a theoretical risk.^[16]

Multi-agent decentralized systems also produce emergent behaviors that are unpredictable from testing individual components. Research from the Cooperative AI Foundation in 2025 identified coordination failures in multi-agent systems as producing "novel and under-appreciated risks." Research from Penn State and Duke confirmed that failures in complex multi-agent systems are "not only common but incredibly difficult to diagnose." Without a central authority, misaligned or malfunctioning nodes cannot be corrected quickly.^[6]

The cryptocurrency dimension of blockchain-based AI networks deserves honest treatment. Gensyn's AI token surged 250% and then crashed 46% within days of its April 2026 launch. Much of what markets call "decentralized AI" is speculative infrastructure with genuine technical ambitions but years away from matching the capability of frontier centralized models. Bittensor is a serious project — but Bittensor's best models remain significantly behind GPT-4 class systems on general benchmarks. The narrative of decentralized AI often runs ahead of its technical reality.^[16]

An honest view
without false certainty

The honest conclusion of this research is that no one knows how this ends — and anyone claiming certainty is not being straight with you. What is knowable is the current direction of travel, what history suggests about similar moments, and what the people who understand these systems best are currently saying.

What we know with confidence

AI is already delivering genuine, documented benefits to human life. Protein folding, cancer detection, drug discovery, climate modeling, accessibility tools — these are real and they matter. The tendency in mainstream discourse to treat AI's benefits as speculative while treating its risks as concrete has things backwards. The benefits are documented. The catastrophic risks are also real, but they are structural and governance-dependent, not inherent in the technology itself.

Power is concentrating at a pace and scale that has no historical parallel. Five private companies controlling 71% of the infrastructure for global cognitive tasks — and growing — is not a normal market outcome. Every previous instance of comparable infrastructure concentration has eventually required regulatory intervention. There is no reason to believe AI will be different; there is reason to believe the stakes are higher than in previous cases.

The governance window is narrowing. Regulatory processes operate on timescales of years. AI capability advances on timescales of months. Anthropic's public statement in June 2026 — a frontier AI lab urging a pause on development — was not a public relations exercise. It was a signal from people who have access to internal capability assessments that the gap between what exists and what governance can address is becoming critical.

What is genuinely uncertain

Whether the net employment effect of AI automation will be positive, and over what timeframe, is genuinely unknown. History suggests yes, eventually. The transition period, and what happens to the people living through it, is less clear.

Whether decentralized AI can eventually match frontier capabilities — removing the trade-off between privacy and power — is an open research question. If efficient distributed training is achievable, the argument for centralized AI weakens significantly. If it is not, the world will likely maintain a landscape where the most capable (and most dangerous) systems remain centralized.

Whether the current regulatory frameworks — the EU AI Act, US executive orders, voluntary safety commitments — are sufficient is unknown, though the evidence from analogous technology monopolies suggests they are not.

What would actually change the trajectory

The interventions most likely to make a genuine difference are structural, not voluntary. Mandatory reporting on the compute used for frontier AI training runs — so regulators know what exists. International agreements on compute thresholds that trigger mandatory oversight, analogous to nuclear non-proliferation treaties. Federated learning requirements for sensitive domains where privacy is a legal right. Public compute infrastructure that gives researchers and smaller organizations access to the tools currently monopolized by five companies.

What is unlikely to work: voluntary safety commitments by companies with financial incentives to accelerate. Purely open models with no governance structure. National AI strategies that replicate the concentration problem at the nation-state level rather than addressing it.

The most important thing to understand

The question of who controls AI infrastructure is not a technical question. It is a political and economic one dressed in technical language. The decisions being made now — about ownership, governance, and access to AI systems — will determine the distribution of power and opportunity for the next generation as surely as the decisions about who controlled railroads and oil in the nineteenth century determined the distribution of wealth in the twentieth.

Most people are not participating in those decisions. They are not participating because the discussions are happening in forums they don't follow, in language designed not to be widely accessible, and with timelines that feel abstract. This document is an attempt to close that gap, even partially. The technology is too important for its governance to remain the exclusive concern of the people who profit from it.