Openness and Safety in the Development of Large Language Models

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11230%2F24%3A10484333" target="_blank" >RIV/00216208:11230/24:10484333 - isvavai.cz</a>

Result on the web

—

DOI - Digital Object Identifier

—

Result language

angličtina

Original language name

Openness and Safety in the Development of Large Language Models

Original language description

The development of the state-of-the-art LLMs (Large Language Models) and other types of generative AI is increasingly carried out behind closed doors of private AI laboratories that belong to competing technological companies. Concerns over risks associated with LLMs led to proposals of regulatory models, many of them motivated by the technological companies. Some then argue that the reduction of risks associated with private, versatile, and powerful LLMs should remain in the hands of their owners because other actors are not equipped to handle the task. This approach makes regulation ineffective by failing to guarantee fair and unbiased representation of social preferences on the capabilities of LLMs and by denying accountability of private LLM owners in domains impacted by LLMs. We argue that alternative regulatory models are possible and outline what needs to change to make them viable. Instead of versatile and private LLMs, domain-specific and open models should be encouraged by regulation to allow oversight by domains regulators best equipped to reduce the risk. The decentralized regulation of domain-specific and open LLMs enables a better representation of values of marginalized groups of people. Something which is harder to achieve with opaque and versatile models developed by private companies.

Czech name

—

Czech description

—

Type

O - Miscellaneous

CEP classification

—

OECD FORD branch

50601 - Political science

Project

<a href="/en/project/LX22NPO5101" target="_blank" >LX22NPO5101: The National Institute for Research on the Socioeconomic Impact of Diseases and Systemic Risks</a><br>

Continuities

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Publication year

2024

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů