Departmental Research Seminar with Dr Brendan Kelters

Our departmental research seminar talks will take place on Wednesdays during term time from 15:00 to 16:30, in room 005, 48 Old Elvet, Durham, DH1 3HN.

This week's speaker is Dr Brendan Kelters from Durham University.

Abstract: In this talk I will argue that the general problem of AI beneficence is a specially serious problem and must be answered in terms of the selective conditions confronting an AI ecosystem. Within this framing, I argue that there is a robust trade-off between how much AI does and AI humanitarianism, since AI doing more things (relatively and in practice also absolutely) involves reducing human impacts on the shared environment and such impactful-ness on the part of prospective non-kin beneficiaries is needed to render beneficence fit behaviour. Hence, if we let AIs do too much, we may expect them to tend to lose even very minimal beneficence towards humans, with intolerable consequences. Sequestering AI power through disciplining AI systems cannot be expected to correct this problem because, even if effective disciplinary techniques are developed, incentive structures will encourage their disuse or misuse and countervailing ‘inverse alignment’ phenomena may cancel their effects. ‘Containing’ AI through limiting its use to some safe maximum cannot be expected to correct this problem because no such maximum is discoverable. The only apt response to this problem is actioning a general logic of AI curtailment; ‘the less AI, the better!’