Understanding the Capabilities, Limitations, and Societal Impact of Large Language Models

By Alex Tamkin, Miles Brundage, Jack Clark, Deep Ganguli. 2021

333 citations recorded by the source corpus.

On October 14th, 2020, researchers from OpenAI, the Stanford Institute for Human-Centered Artificial Intelligence, and other universities convened to discuss open research questions surrounding GPT-3, the largest publicly-disclosed dense language model at the time. The meeting took place under Chatham House Rules. Discussants came from a variety of research backgrounds including computer science, linguistics, philosophy, political science, communications, cyber policy, and more. Broadly, the discussion centered around two main questions: 1) What are the technical capabilities and limitations of large language models? 2) What are the societal effects of widespread use of large language models? Here, we provide a detailed summary of the discussion organized by the two themes above.

Subfields: Artificial Intelligence, Health Informatics, Large Language Models, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Health Informatics

In Large Language Models

In Safety, Robustness & Interpretability

Citations in the corpus

Training language models to follow instructions with human feedback (2022, background): “These risks have been extensively documented (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021).”
Ethical and social risks of harm from Language Models (2021, background): “While some predict that it will be cheaper to hire humans to generate disinformation (Tamkin et al., 2021), it is possible that LM-assisted content generation may offer a cheaper way of generating diffuse disinformation at scale.”
TruthfulQA: Measuring How Models Mimic Human Falsehoods (2021, background): “This could lead to deception and distrust (Tamkin et al., 2021).”
Societal Biases in Language Generation: Progress and Challenges (2021, background): “With continuous technological advances, more organizations will turn to effective NLG techniques, making it imperative to start setting norms to reduce harmful allocational impacts (Tamkin et al., 2021).”
Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned (2022, background): “…social biases (e.g., [47, 28, 1, 33, 7]), generating offensive or toxic outputs [25], leaking personally identiﬁable information from the training data [13], aiding in disinformation campaigns [12], generating extremist texts [37], spreading falsehoods [35], and more [9, 10, 18, 57, 22, 51].”
Alignment of Language Agents (2021, background): “Another recent study (Tamkin et al., 2021) summarizes a discussion on capabilities and societal impacts of LLMs.”
Predictability and Surprise in Large Generative Models (2022, background): “Although we focus on scaling laws, many of our points complement existing views on the societal risks of deploying large models [7, 67, 9, 19, 72, 41].”
What Changes Can Large-scale Language Models Bring? Intensive Study on HyperCLOVA: Billions-scale Korean Generative Pretrained Transformers (2021, background): “To overcome these issues, many researchers argue the necessity of controllability when generating sentences such as ﬁltering and investigate how to more effectively reﬁne the data for debiasing (Tamkin et al., 2021).”
Taxonomy of Risks posed by Language Models (2022, background): “As one paper concluded, it is difficult to scope all possible (mis-)uses of LMs [180].”
A Survey of Large Language Models (2023, positive): “However, the use of LLMs in law also raises concerns about legal challenges, including copyright issues [414], personal information leakage [415], or bias and discrimination [416].”
CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society (2023, background): “Research in AI alignment focuses on discouraging AI models from producing false, offensive, deceptive, or manipulative information that could result in various harms [56, 112, 42, 37].”
Multimodal Large Language Models in Health Care: Applications, Challenges, and Future Outlook (2024, neutral): “Large language models (LLMs) are sophisticated machine learning algorithms designed to process, understand, and generate humanlike language, enabling key developments in applications such as automated conversation, text analysis, creative writing, and complex problem-solving [1].”
GPT-4o System Card (2024, background): “While this is a known issue for text in GPT models [18, 19], there was concern from red teamers that this information may be more persuasive or harmful when delivered through audio, especially if the model was instructed to speak emotively or emphatically.”
When Combating Hype, Proceed with Caution (2021, background): “education, medicine, law, and the service sector more broadly, as well as making mass surveillance and misinformation campaigns far more effective, and opening up additional new use cases that will be hard for us to foresee (Brundage et al., 2018; Tamkin et al., 2021; Bommasani et al., 2021).”
Towards a Standard for Identifying and Managing Bias in Artificial Intelligence (2022, background): “Early practice has shown that concerns about the use of LLMs are indeed valid, with preliminary experimental results showing LLMs exhibit significant bias [158, 165, 166].”
The Ghost in the Machine has an American accent: value conflict in GPT-3 (2022, background): “Amongst the recommendations in the aforementioned 2020 whitepaper, was a call for steering the model toward human values[4]: our work helps address this call.”
The Robots Are Coming: Exploring the Implications of OpenAI Codex on Introductory Programming (2022, methodology): “Anticipating this shift, we put the following sentence into GPT-3, “The robots are taking over.””
Opportunities and Challenges for ChatGPT and Large Language Models in Biomedicine and Health (2023, background): “The limitations, challenges, and risks associated with LLMs have been extensively discussed in previous research [19, 111, 112], and in this survey, we will specifically focus on those relevant to the context of biomedical and health domains.”
Large Language Models in Finance: A Survey (2023, background): “Two major challenges are the production of disinformation and the manifestation of biases, such as racial, gender, and religious biases, in LLMs [43].”