Experts at the University of Oxford and international collaborators have released a new study in Nature Machine Intelligence that addresses the ethical concerns related to the responsibility for outputs generated by large language models (LLMs). LLMs such as ChatGPT pose critical issues regarding the attribution of credit and rights for useful text generation, which differs from traditional AI responsibility debates centered on harmful consequences.
According to the study, while human users cannot take full credit for the positive results generated by an LLM, they are still responsible for harmful uses, such as generating misinformation or being careless in checking the accuracy of generated text. This divergence can lead to an “achievement gap” where useful work is being done, but people cannot gain as much recognition or satisfaction as they used to.
The study, which was co-authored by interdisciplinary experts in law, bioethics, machine learning, and related areas, delves into the potential impact of LLMs in critical areas such as education, academic publishing, intellectual property, and the creation of mis- and disinformation. LLMs can be helpful in education but are error-prone, and overuse might affect critical thinking skills. Institutions must adapt assessment styles, rethink pedagogy, and update academic misconduct guidance to handle LLM usage effectively.
The researchers recommend that guidelines for LLM use and responsibility must be established quickly, especially in education and publishing. The submission of articles should include a statement on LLM usage, along with relevant supplementary information. Disclosure for LLMs should be similar to human contributors, acknowledging significant contributions.
Rights in generated text, such as intellectual property rights and human rights, pose challenges since they rely on notions of labor and creativity established with humans in mind. The researchers recommend developing or adapting frameworks like “contributorship” to handle this fast-evolving technology while still protecting the rights of creators and users.
LLMs can be used to generate harmful content, including large-scale mis- and disinformation. Therefore, people must be accountable for the accuracy of LLM-generated text they use, alongside efforts to educate users and improve content moderation policies to mitigate risks.
In conclusion, LLMs such as ChatGPT bring about an urgent need for an update in our concept of responsibility. The authors recommend guidelines on authorship, requirements for disclosure, educational use, and intellectual property, drawing on existing normative instruments and similar relevant debates, such as on human enhancement. Norms requiring transparency are especially important to track responsibility and correctly assign praise and blame.