Disparate access to benefits due to hardware, software, skills constraints
Due to differential internet access, language, skill, or hardware requirements, the benefits from LMs are unlikely to be equally accessible to all people and groups who would like to use them. Inaccessibility of the technology may perpetuate global inequities by disproportionately benefiting some groups.
ENTITY
1 - Human
INTENT
2 - Unintentional
TIMING
2 - Post-deployment
Risk ID
mit257
Domain lineage
6. Socioeconomic and Environmental
6.1 > Power centralization and unfair distribution of benefits
Mitigation strategy
1. Prioritize the development and deployment of Small Language Models (SLMs) and offline-capable solutions to mitigate infrastructure dependencies. This strategy directly addresses differential internet access and hardware constraints by enabling AI-powered functionality in resource-limited environments, thereby bridging the fundamental digital divide. 2. Enforce rigorous accessible design principles and comprehensive digital literacy programs. Designs must comply with standards like WCAG, ensuring optimal contrast, clear visual cues, and compatibility with assistive technologies (e.g., screen readers). Concurrently, invest in initiatives that enhance digital literacy and effective prompt utilization skills across all population groups. 3. Employ multilingual expansion and localized fine-tuning to ensure linguistic equity. Models should be systematically fine-tuned using native subject matter experts (SMEs) on diverse languages and regional dialects to reduce accuracy degradation in non-primary language contexts, ensuring model utility and culturally sensitive communication for a global user base.
ADDITIONAL EVIDENCE
Access to economic opportunities LM design choices have a downstream impact on who is most likely to benefit from the model. For example, product developers may find it easier to develop LM-based applications for social groups where the LM performs reliably and makes fewer errors; potentially leaving those groups for whom the LM is less accurate with fewer good applications (see Lower performance by social group).