Porsdam Mann, S., Earp, B. et al. (2023).
Nature Machine Intelligence.
The recent releases of large-scale language models (LLMs), including OpenAI’s ChatGPT and GPT-4, Meta’s LLaMA, and Google’s Bard have garnered substantial global attention, leading to calls for urgent community discussion of the ethical issues involved. LLMs generate text by representing and predicting statistical properties of language. Optimized for statistical patterns and linguistic form rather than for
truth or reliability, these models cannot assess the quality of the information they use.
Recent work has highlighted ethical risks that are associated with LLMs, including biases that arise from training data; environmental and socioeconomic impacts; privacy and confidentiality risks; the perpetuation of stereotypes; and the potential for deliberate or accidental misuse. We focus on a distinct set of ethical questions concerning moral responsibility—specifically blame and credit—for LLM-generated
content. We argue that different responsibility standards apply to positive and negative uses (or outputs) of LLMs and offer preliminary recommendations. These include: calls for updated guidance from policymakers that reflect this asymmetry in responsibility standards; transparency norms; technology goals; and the establishment of interactive forums for participatory debate on LLMs.
(cut)
Credit–blame asymmetry may lead to achievement gaps
Since the Industrial Revolution, automating technologies have made workers redundant in many industries, particularly in agriculture and manufacturing. The recent assumption25 has been that creatives
and knowledge workers would remain much less impacted by these changes in the near-to-mid-term future. Advances in LLMs challenge this premise.
How these trends will impact human workforces is a key but unresolved question. The spread of AI-based applications and tools such as LLMs will not necessarily replace human workers; it may simply
shift them to tasks that complement the functions of the AI. This may decrease opportunities for human beings to distinguish themselves or excel in workplace settings. Their future tasks may involve supervising or maintaining LLMs that produce the sorts of outputs (for example, text or recommendations) that skilled human beings were previously producing and for which they were receiving credit. Consequently, work in a world relying on LLMs might often involve ‘achievement gaps’ for human beings: good, useful outcomes will be produced, but many of them will not be achievements for which human workers and professionals can claim credit.
This may result in an odd state of affairs. If responsibility for positive and negative outcomes produced by LLMs is asymmetrical as we have suggested, humans may be justifiably held responsible for negative outcomes created, or allowed to happen, when they or their organizations make use of LLMs. At the same time, they may deserve less credit for AI-generated positive outcomes, as they may not be displaying the skills and talents needed to produce text, exerting judgment to make a recommendation, or generating other creative outputs.
