The Obsolescence Crisis in Horizontal AI Regulation
Why It Matters
If regulation targets the wrong technical layer, it creates ineffective compliance burdens while failing to mitigate actual harms like fraud or cyber threats. This shift suggests a move away from 'one-size-fits-all' AI acts toward sector-specific, outcomes-based oversight.
Key Points
- AI capabilities are increasingly driven by post-training, tool use, and inference scaling rather than just the base model development.
- Safety mitigations like filters and oversight mechanisms are often applied at the software scaffold level, making model-centric regulation ineffective.
- A shift from 'horizontal' (cross-industry) to 'outcomes-based' regulation is necessary to allow diverse actors to negotiate appropriate safety interventions.
- The burden of proof should be on regulators to show that existing negligence laws are insufficient before passing new AI-specific legislation.
Tech policy expert Seb Krier has issued a critique of current AI legislative frameworks, arguing that they are increasingly decoupled from technical reality. The core of the argument centers on the fact that modern AI performance and safety mitigations are increasingly derived from post-training techniques, scaffolding, and inference compute scaling rather than just the base model weights. Consequently, regulations that focus exclusively on model developers fail to address the complex chain of actors responsible for how AI is deployed. Krier suggests that existing legal proposals, particularly those following the European 'entity-based' model, are ill-equipped to handle risks like AI-assisted fraud which cannot be 'tuned out' at the model level. The analysis calls for a shift toward outcomes-based regulation and deeper domain expertise within existing regulatory bodies to address specific vertical risks.
Imagine trying to regulate car safety by only looking at the engine, while ignoring the brakes, seatbelts, and the driver. That is what current AI laws are doing. Experts are warning that because AI is now built in layers—where a base model gets extra 'tools' and 'filters' added later by different companies—targeting just the original model builders doesn't work. The real risks happen in how the software is used, not just how it's trained. We need to stop trying to pass one giant AI law and instead focus on specific rules for specific problems like cyber-attacks or banking fraud.
Sides
Critics
Argues that current AI bills are obsolete because they ignore the importance of scaffolding and post-training in the AI value chain.
Defenders
Promotes a centralized, horizontal, and entity-based regulatory framework that Krier argues is too rigid and ineffective.
Noise Level
Forecast
Legislative bodies will likely face increasing pressure to amend 'omnibus' AI acts in favor of sector-specific guidelines as the technical gap widens. We can expect a rise in 'regulatory sandboxes' where domain-specific regulators, rather than general tech authorities, lead the enforcement for their respective industries.
Based on current signals. Events may develop differently.
Timeline
Policy Critique Published
Expert Seb Krier publishes a detailed thread highlighting the mismatch between AI law and the current technical state of post-training and tool-use.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.