Lots of AI Regulations are Conceivable; Few Will Address Existential Threats

One problem with calls for “regulating artificial intelligence” is that it is not entirely clear what should be done, especially on the core issue of models becoming autonomous, in the context of nation state competition. 

The key idea of one approach known as “Mutual Assured AI Malfunction” is that nation state actors pursuing a monopoly on superintelligent AI should face the realistic prospect that rivals will detect and disable a destabilizing project before it matures. 

That is a specific problem, separate from general controls on autonomous AI use that might apply to all instances of autonomous action. 

MAIM sits within a larger conversation about whether AI competition more closely resembles nuclear arms races or something else. It rejects the idea that global pauses are possible or  politically unrealistic. Instead it treats great-power rivalry as inevitable.

The objective is creation of a default strategic environment in which no power can safely race to an unbridled advantage without expecting its own effort to be maimed. 

MAIM remains a theoretical and strategic proposal rather than formal policy in analyses of United States–China AI competition, for example. 

Critics will note limitations in the analogy to mutually assured destruction, which likewise prevented nuclear war by essentially ensuring that any use of strategic nuclear weapons would result in destruction of whichever nation chose to use it. So strategic nuclear weapons essentially become unusable. 

Nuclear weapons are passive instruments under human command chains.  Autonomous or superintelligent AI is an independent optimizing agent that may pursue goals orthogonal to its creators’ aims.

Nor are rogue actors limited to nation states. Once software exists, copying and iterating is cheap. This makes exclusive control or reliable sabotage harder. Non-state or rogue actors will not face the same mutual-vulnerability constraints that bind nuclear powers.

AI catastrophe scenarios can be harder to calculate in advance and may not leave clear “retaliatory” options once triggered. 

Incentives and actors are more heterogeneous. Nuclear deterrence primarily involved a small number of states with high stakes in regime survival. AI involves commercial firms, researchers, and diffuse national programs whose short-term competitive or profit motives can outweigh distant shared risks. 

So coordination problems are harder. Also, opaque model behavior can erode the signaling and crisis stability that nuclear deterrence relies on.

The point is that deterrence of this sort will be challenging. We won’t always agree on what moves are “destabilizing.” We will not have perfect transparency so there will always be some amount of ambiguity.

Sure, all sorts of “regulations” might also be proposed, but most of them do not directly involve controlling autonomous system misuse. 

That is not to say such other measures are unimportant. But they do not address concerns about existential AI threats posed by autonomous or super-intelligent systems. 

And the “easier” measures might not be directly effective in addressing autonomous system risks, either interstate or intrastate. 

So the “danger” is that politicians under pressure to “do something” will make a lot of noise about issues that arguably are useful, without addressing the “existential” dangers people worry about. 

In that sense, AI autonomy and its existential threats are no easier to solve than many other thorny social problems. 

Approach	What government would actually do	Main benefit	Main drawback / risk	Practicality
Apply existing laws	Use consumer protection, fraud, discrimination, privacy, securities, product-liability, employment and criminal laws when AI causes ordinary legally cognizable harms	Immediate and technology-neutral. Doesn't require government to predict which AI capabilities will matter	Existing laws may not fit novel harms; enforcement is often after the damage occurs	Very high
Mandatory disclosure, transparency	Require companies to disclose model capabilities, limitations, training-data summaries, evaluations, ownership, incidents, etc.	Gives customers, regulators and researchers information without directly controlling development	Disclosure can become expensive paperwork; sensitive information can reveal IP or security vulnerabilities; disclosures don't necessarily make systems safer	Very high
Independent testing, audits	Require frontier or high-risk systems to undergo standardized safety, cybersecurity, bias, privacy or performance testing before deployment	Creates an external check on company self-assessment; potentially catches problems before deployment	Hard to define meaningful tests for rapidly changing models; auditors can become a new bureaucracy; "passing" a benchmark may create false confidence	High
Risk-management standards	Require companies to maintain documented processes for identifying, measuring and mitigating AI risks—similar to financial, aviation or cybersecurity risk management	Regulates process rather than a particular technology; adaptable as AI changes	Companies can satisfy procedures without actually reducing risk; standards can become box-checking exercises	Very high
Incident reporting	Require developers/deployers to report serious AI failures, security breaches, dangerous behavior or loss of control to regulators	Creates a feedback system for regulators and researchers; particularly useful when risks are initially poorly understood	Companies may under-report; determining what constitutes a "serious" incident is difficult; reporting could expose security-sensitive information	High
Liability, duty of care	Make developers or deployers financially responsible when AI causes foreseeable harm, potentially through negligence, product liability or a specific AI duty of care	Uses economic incentives rather than prescriptive technical rules; encourages companies to internalize externalities	Liability can discourage experimentation; causation can be extremely difficult when many parties contribute to an AI system	High, especially sector-by-sector
Sector-specific regulation	Give FDA, FAA, financial regulators, healthcare regulators, employment regulators, etc. authority over AI used in their domains	Probably the most natural approach because the risk of an AI system depends heavily on what it does	Creates multiple regulatory regimes; can slow adoption and produce inconsistent requirements	Very high
Prohibit particular applications	Ban specified uses—e.g., certain forms of biometric surveillance, manipulation, autonomous weapons, discriminatory decision-making or fraud	Very clear red lines; prevents particularly unacceptable uses regardless of technical sophistication	Defining the prohibited category is difficult; legitimate and harmful uses can look technically similar; enforcement can be evaded	High for narrow uses
High-risk classification	Create categories such as "high-risk AI" and impose enhanced testing, documentation, human oversight and monitoring on them	Focuses regulation where potential harm is greatest rather than regulating every chatbot	Classification disputes; innovation may migrate around regulatory boundaries	High
Frontier-model requirements	Impose additional obligations on the largest/most capable models—risk evaluations, cybersecurity, incident reporting, red-team testing, safety plans, etc.	Concentrates regulation on the relatively small number of firms developing frontier systems	Capability thresholds quickly become obsolete; could protect incumbents by making it harder for startups to compete	Medium-high
Government licensing / pre-approval	Require a company to obtain government permission before training or releasing models above a specified capability/compute threshold	Gives government the strongest ability to stop genuinely dangerous systems before deployment	Extremely intrusive; government must determine what constitutes dangerous capability; potentially slows beneficial innovation and entrenches existing firms	Low-medium
Regulate compute	Require reporting/licensing for very large AI-training runs or data centers; track high-end chips and compute clusters	Targets the physical bottleneck underlying frontier AI rather than trying to regulate every model	Compute becomes cheaper/more distributed; companies can optimize algorithms or move overseas; difficult to distinguish benign from dangerous compute	Medium
Chip/export controls	Restrict advanced AI chips, semiconductor equipment or model weights from particular countries/entities	Useful when the objective is national security, because chips are tangible and controllable	Doesn't directly address domestic AI harms; encourages substitution, stockpiling and indigenous foreign capacity	High for national-security objectives
Competition, antitrust policy	Challenge acquisitions, exclusive cloud/model/chip arrangements, discriminatory access, tying and potentially vertical integration	Addresses the possibility that AI becomes concentrated among a handful of firms and preserves competitive access to compute/models	Antitrust remedies are slow; breaking up or restricting efficient integration could reduce economies of scale	High
Interoperability, portability	Require APIs, data portability, model portability or interoperability between AI services	Reduces switching costs and makes it easier for startups to compete	Standardization can freeze technology prematurely and potentially reduce security	Medium
Copyright, training-data rules	Require licensing, compensation, opt-outs, attribution or disclosure concerning copyrighted training material	Protects creators and establishes property rights around a major AI input	Could substantially increase model costs; licensing billions of works is technically difficult; may advantage firms with large legal budgets	Medium
Privacy, data regulation	Restrict collection and use of personal data for training and inference; require consent, minimization and deletion	Addresses a concrete existing harm without regulating AI as such	Reduces useful training data; enforcement and cross-border data issues are difficult	High
AI-generated-content labeling	Require disclosure/watermarking of synthetic images, video, audio or text	Helps people distinguish synthetic from human-created material and can improve provenance	Watermarks can be removed; detection becomes harder as models improve; doesn't solve persuasion/misinformation itself	Medium-high
Government procurement rules	Government agencies purchase only AI satisfying specified security, transparency, privacy or performance requirements	Government can use its enormous purchasing power without regulating the entire private economy	Primarily affects government suppliers; procurement standards can become politicized or overly restrictive	Very high
Taxes, fees on AI or compute	Tax AI usage, high-end compute, AI profits or automation; use revenue for worker transition, social insurance, etc.	Addresses distributional effects while allowing firms to continue innovating	Difficult to identify the tax base; could discourage productive investment and push activity offshore	Medium
Worker-transition policies	Expand unemployment insurance, retraining, wage insurance, portable benefits or education rather than restricting AI	Deals with the consequences of AI rather than trying to stop AI; preserves productivity benefits	Expensive; difficult to target assistance toward genuinely displaced workers	Very high
International agreements	Establish common standards for dangerous capabilities, testing, incident reporting, chip controls and military uses	Reduces regulatory arbitrage and makes some global risks manageable	Extremely difficult because countries have different economic and national-security interests	Low-medium, but potentially important

Diffusion Models Can Create New Images Without Infringing on Training Sources

A study of language models, especially those designed to create images and video, might suggest that copyright infringements by diffusion-based models might be impossible to prove. 

The authors say “locating a part of the training data that can be held responsible for a generated sample, can become impossible if a model is trained on a sufficiently large corpus of data.”

In other words, a large training set is all that is needed to render copyright traceable to any single source impossible. 

The study methodology relies on omitting an attributed unit from the training set (one image or all images by the same creator), the induced counterfactual sample cannot be directly attributed to that same unit. 

source: Nature Communications

In the above example, a generated image created from a smaller data set is more attributable to a single training image.

A large data set, on the other hand, creates a new image that shows no trace of a single training image.  

Model / family	Main generation	Diffusion?	Basic architecture / approach
GPT-4 / GPT-4o / GPT-5 family	Text, multimodal	No*	Primarily autoregressive Transformer for language
Claude 3/4 family	Text, code, multimodal	No*	Autoregressive Transformer
Gemini 1–4 family	Text, multimodal	No*	Primarily autoregressive Transformer
Llama family	Text, code	No	Autoregressive Transformer
Mistral / Mixtral family	Text, code	No	Autoregressive Transformer, including MoE
DeepSeek family	Text, code, reasoning	No	Autoregressive Transformer/MoE
Gemini Diffusion	Text/code	Yes	Text diffusion Transformer; generates blocks of tokens and iteratively refines them (Google DeepMind)
DiffusionGemma	Text/code	Yes	Non-sequential diffusion Transformer based on Gemma research (Google DeepMind)
Stable Diffusion 1.x/2.x/XL	Images	Yes	Latent diffusion, historically U-Net-based
Stable Diffusion 3/3.5	Images	Yes	Diffusion Transformer / MM-DiT
FLUX.1 / FLUX family	Images	Yes	Diffusion/flow-based Transformer
DALL·E 2	Images	Yes	Diffusion
DALL·E 3	Images	Yes	Diffusion-based image generation
Imagen 2/3/4	Images	Yes	Diffusion-based image generation; Google explicitly describes Imagen as a diffusion model (Google DeepMind)
Midjourney	Images	Yes*	Widely understood to use diffusion/related denoising techniques, though Midjourney does not disclose its architecture in detail
Adobe Firefly	Images/video	Yes*	Diffusion-based generation among its models; architecture varies by model
Sora	Video	Yes	Diffusion Transformer (DiT); starts from noisy video patches and denoises them (OpenAI)
Sora 2	Video	Yes*	Diffusion-based video generation; details of current implementation are less fully disclosed
Veo 2/3/3.1	Video	Yes*	Diffusion/Transformer-based video generation; Google describes Veo as drawing on its diffusion-model research (Google DeepMind)
Runway Gen-2/3	Video	Yes*	Diffusion-based video generation
Kling	Video	Yes*	Diffusion/Transformer-based video generation
Luma Dream Machine	Video	Yes*	Diffusion/Transformer-based video generation
MusicGen	Music/audio	No	Autoregressive Transformer over discrete audio tokens
AudioCraft / related models	Audio/music	Mixed	Different models use different architectures
Genie 2	Interactive worlds/video	Yes	Autoregressive latent diffusion—an interesting hybrid (Google DeepMind)

“Our findings provide a compelling case that attribution, the task of locating a unit of data within the training set that can be held responsible for a generated sample, is practically impossible on contemporary generative diffusion models,” say authors Zheng Dai and David Gifford.