Politicians love a good crusade, and right now, their favorite target is artificial intelligence. When Australian Industry Minister Ed Husic declared that letting AI companies self-regulate is a model doomed to fail, he captured the collective anxiety of governments worldwide. The narrative is simple: greedy tech giants are running wild with dangerous code, and only strict, state-mandated guardrails can save society from ruin.
It is a comforting story. It is also completely wrong.
By demanding heavy-handed government intervention, politicians are not reining in Silicon Valley. They are doing its bidding. The push for sweeping, preemptive AI legislation is the greatest gift we could possibly give to the world's most dominant tech companies. Under the guise of public safety, we are building a regulatory moat so deep and so wide that no startup will ever be able to cross it.
The Regulatory Capture Masterclass
If you want to understand who benefits from a law, do not look at who opposes it. Look at who lobbyist groups are pushing to write it.
The executives of trillion-dollar tech platforms are not hiding in fear from government regulation. They are practically begging for it. They fly to Washington, Brussels, and Canberra to testify in front of politicians, nodding solemnly as they agree that AI is too powerful to be left unchecked. They propose draft frameworks. They offer their own internal safety teams as consultants.
This is not corporate altruism. It is classic regulatory capture.
When a government creates a complex, bureaucratic licensing regime for software, it introduces massive compliance costs. A giant tech corporation can easily employ five hundred compliance lawyers, data scientists, and policy experts to fill out thousand-page impact assessments. A three-person startup operating out of a garage cannot.
By forcing AI companies to obtain government approval before deploying a model, we are effectively outlawing garage-stage innovation. We are ensuring that the only entities allowed to build the foundation models of tomorrow are the ones that already have billions of dollars in the bank today. The established players want to pull the ladder up behind them, and governments are handed the hammer to secure the rungs.
The Myth of the Omniscient Auditor
The core premise of government-led AI safety is that a state agency can accurately audit and control a neural network. This premise is built on a fundamental misunderstanding of how modern machine learning works.
Traditional software is deterministic. It runs on explicit, line-by-line code written by human programmers. If something goes wrong, you can inspect the code, find the bug, and fix it. State regulators can write rules for deterministic software because its behavior is predictable and auditable.
Deep learning is different. It is probabilistic. We do not write the rules; we write the algorithm that learns the rules from massive datasets. The resulting model is a black box containing hundreds of billions of numerical weights. Nobody, not even the engineers who built it, can predict exactly how a complex model will respond to every possible prompt.
How does a government bureaucrat audit a black box?
They cannot. Instead, they rely on proxy measures and theater. They demand compliance checklists:
- Did you document your training data?
- Did you perform a bias assessment?
- Did you run a government-approved red-teaming exercise?
These exercises do not make AI safer. They make it more expensive. A company can check every box, file every form, and still deploy a model that exhibits unexpected, harmful behaviors under real-world conditions. Conversely, a highly secure, heavily aligned open-source model might fail to launch simply because its creators did not have the administrative resources to navigate the state's labyrinthine approval pipeline.
The Ghost of GDPR
We do not have to guess what happens when governments attempt to regulate fast-moving digital technology. We have a massive, real-world case study: the European Union’s General Data Protection Regulation (GDPR).
Introduced in 2018, GDPR was supposed to break the stranglehold of big tech by giving users control over their data. It was hailed as a triumph of consumer protection.
The actual result? GDPR decimated the European startup ecosystem while cementing the dominance of the US tech giants.
Large platforms had the resources to implement complex consent tracking, hire data protection officers, and handle the legal overhead. Small European startups, faced with the threat of existential fines for minor compliance errors, simply shut down or chose not to launch in the European market. US venture capital shifted away from Europe because the regulatory risk was too high. Today, European consumers are greeted with annoying cookie banners on every website, while the market share of the dominant tech advertising platforms remains virtually untouched.
The EU AI Act is currently repeating this exact mistake on a much larger scale. By classifying foundation models based on arbitrary computing thresholds, it penalizes raw computational power rather than actual real-world application. It treats the tool itself as inherently dangerous, rather than focusing on how the tool is used.
The Real Threat is the Death of Open Source
The most dangerous casualty of the regulatory crusade is open-source AI.
Open-source models are the democratic counterweight to corporate monopolies. When researchers publish their weights publicly, they allow academic institutions, independent developers, and small businesses to inspect the code, find vulnerabilities, and build specialized applications without paying rent to a tech giant. Open source is why AI capability has advanced so quickly; it allows the global scientific community to collaborate and iterate openly.
Yet, under almost every proposed regulatory framework, open-source developers are treated as a liability.
Because open-source models can be downloaded and run locally, they cannot be easily controlled, recalled, or subjected to continuous government monitoring. Regulators view this decentralized nature as a security threat. They argue that open models could be used by bad actors to write malicious code or generate misinformation, and therefore, publishing open-source weights should be heavily restricted or outright banned.
This argument is highly flawed. Bad actors already have access to powerful tools, and they do not wait for regulatory approval. Restricting open-source AI does not stop malicious behavior; it merely strips defensive tools from the hands of the public.
Imagine a scenario where a small team of researchers builds a highly efficient, locally deployable AI model designed to identify cybersecurity vulnerabilities. Under a strict licensing regime, this team would be forced to undergo months of government vetting, pay exorbitant licensing fees, and potentially face liability if their model is ever misused. Unable to meet these demands, they scrap the project. Meanwhile, state-sponsored hacking groups continue to develop their own proprietary models in secret. The public is left less secure, not more.
By killing open source, we hand total control of the cognitive infrastructure of the future to a tiny cartel of corporate executives and government censors.
Focus on Harm, Not Technology
The solution is not to let AI companies do whatever they want. Self-regulation is indeed insufficient, but not for the reasons politicians think. The mistake lies in trying to regulate the technology rather than the action.
We do not need new, sweeping AI departments to police algorithms. We already have laws that govern human behavior, and those laws do not stop applying just because someone uses a computer.
- If someone uses an AI to commit fraud, they should be prosecuted under existing fraud laws.
- If a self-driving car causes an accident, the liability should fall on the manufacturer under existing product liability frameworks.
- If an AI-generated medical diagnostic tool misdiagnoses a patient due to negligence, it is a medical malpractice issue.
By shifting the focus from pre-market licensing to post-market liability, we create a system that encourages innovation while holding bad actors accountable. If companies know they will face crippling civil lawsuits and criminal liability if their products cause real, measurable harm, they will invest heavily in safety. They will do so because their survival depends on it, not because they are trying to satisfy a bureaucrat's checklist.
This liability-driven model allows startups to build, experiment, and compete freely. It does not require permission from a government board before writing a line of code. It simply holds everyone to the same standard: if you break it, you pay for it.
Politicians like Ed Husic want us to believe the choice is between wild-west chaos and wise government stewardship. That is a false dichotomy. The real choice is between a dynamic, competitive market driven by open innovation, and a stagnant, corporate oligopoly protected by state gatekeepers.
If we choose the path of preemptive regulation, we will get exactly what we deserve: slower progress, fewer choices, and a tech industry entirely controlled by the very giants we claimed we wanted to restrain.
