Entrenchment and exacerbation of existing inequalities
The most serious access-related risks posed by advanced AI assistants concern the entrenchment and exacerbation of existing inequalities (World Inequality Database) or the creation of novel, previously unknown, inequities. While advanced AI assistants are novel technology in certain respects, there are reasons to believe that – without direct design interventions – they will continue to be affected by inequities evidenced in present-day AI systems (Bommasani et al., 2022a). Many of the access-related risks we foresee mirror those described in the case studies and types of differential access. In this section, we link them more tightly to elements of the definition of an advanced AI assistant to better understand and mitigate potential issues – and lay the path for assistants that support widespread and inclusive opportunity and access. We begin with the existing capabilities set out in the definition (see Chapter 2) before applying foresight to those that are more novel and emergent. Current capabilities: Artificial agents with natural language interfaces. Artificial agents with natural language interfaces are widespread (Browne, 2023) and increasingly integrated into the social fabric and existing information infrastructure, including search engines (Warren, 2023), business messaging apps (Slack, 2023), research tools (ATLAS.ti, 2023) and accessibility apps for blind and low-vision people (Be My Eyes, 2023). There is already evidence of a range of sociotechnical harms that can arise from the use of artificial agents with natural language interfaces when some communities have inferior access to them (Weidinger et al., 2021). As previously described, these harms include inferior quality of access (in situation type 2) across user groups, which may map onto wider societal dynamics involving race (Harrington et al., 2022), disability (Gadiraju et al., 2023) and culture (Jenka, 2023). As developers make it easier to integrate these technologies into other tools, services and decision-making systems (e.g. Marr, 2023; Brockman et al., 2023; Pinsky, 2023), their uptake could make existing performance inequities more pronounced or introduce them to new and wider publics.
ENTITY
1 - Human
INTENT
2 - Unintentional
TIMING
2 - Post-deployment
Risk ID
mit425
Domain lineage
6. Socioeconomic and Environmental
6.1 > Power centralization and unfair distribution of benefits
Mitigation strategy
1. Mandate the curation and use of training data sets that are statistically representative of the diverse user demographics to prevent the initial encoding of societal biases, supplementing with data augmentation or synthetic generation where necessary to address data voids. 2. Integrate technical fairness audits, utilizing advanced statistical and machine learning tools (e.g., disparate impact metrics, counterfactual fairness) during model development, testing, and post-deployment monitoring to detect and correct algorithmic disparities across sensitive attributes. 3. Establish a comprehensive governance framework that includes mandating cognitive and demographic diversity in AI development and audit teams, and instituting an independent AI ethics review board to mandate and verify the implementation of fairness principles across the AI lifecycle.