Back to the MIT repository
6. Socioeconomic and Environmental3 - Other

Financial Costs

The estimated financial costs of training, testing, and deploying generative AI systems can restrict the groups of people able to afford developing and interacting with these systems.

Source: MIT AI Risk Repositorymit171

ENTITY

1 - Human

INTENT

1 - Intentional

TIMING

3 - Other

Risk ID

mit171

Domain lineage

6. Socioeconomic and Environmental

262 mapped risks

6.1 > Power centralization and unfair distribution of benefits

Mitigation strategy

1. Implement model right-sizing strategies, prioritizing the deployment of the smallest model that meets use-case specific accuracy and latency targets. This includes leveraging model optimization techniques such as quantization and batching to significantly reduce inference costs and increase throughput. 2. Establish a comprehensive Cloud FinOps framework to gain end-to-end visibility into the Total Cost of Ownership (TCO) for all AI use cases, coupled with a token-weighted chargeback mechanism to assign financial accountability and mitigate 'noisy neighbor' resource overuse. 3. Adopt open-source models and strategically deploy a hierarchical agentic architecture that utilizes cheaper CPU-based virtualization/containerization for smaller task-specific models, reserving specialized accelerators solely for large models that demand the lowest latency.

ADDITIONAL EVIDENCE

Concretely: sourcing training data, computing infrastructure for training and testing, and labor hours contribute to the overall financial costs. These metrics are not standard to release for any system, but can be estimated for a specific category, such as the cost to train and host a model.