Regulating Innovation: Legal Challenges and Governance Frameworks for AI-Driven Startups

artificial intelligence regulation; startups; regulatory incidence; algorithmic value chain; risk-based regulation; regulatory sandboxes; compliance capability; innovation policy

New models of artificial intelligence distribute responsibility on the provider of an AI system, differentiate the responsibilities based on the risk of the intended use, and impose the responsibilities through documentary conformity requirements that are evaluated prior to deployment. Each of these three design decisions plays a game-changing role and creates its own distinct set of issues with the structural nature of AI startups, and the poor outcomes are not the result of transitional friction, but rather foreseeable consequences of the way the regulated object has been designed. We make three assertions. The first is that there are obligations (that are part of the system provider) in the middle of the algorithmic value chain, where the firm does not have access to the information held by the customer and the control over the behaviour of the model is not what the model provider has, since the customer is the one who has the information. Second, "whatever a statute allocates a contract reallocates," by which I mean that the upstream provider, by the terms of service, and the downstream enterprise customer, by the terms of procurement, disclaim, thereby reallocating the incidence down through bargaining power—rather than the intent of the legislature—because startups are squeezed between. Third, the risk-tier classification assumes a fixed intended use, but early-stage companies change their use, which means a firm's risk-tier is not stable in the population that might not be able to handle re-classification. We also differentiate between compliance cost, which can be paid, and compliance capability (which requires organisational structures that are not available in young firms), and propose that this differentiation is why funding-based support instruments are not as effective. Six governance instruments are evaluated for these problems, and seven propositions with an analytical protocol are provided. No artifical statistics are stated.