GPT-4 System Card¶
Authors: OpenAI
Year: 2023
Venue: OpenAI technical report
type: other collection: misfiled status: reviewed
Links: link
Summary¶
Summary
OpenAI's system card for GPT-4 describes the safety challenges posed by the model's limitations and capabilities, the qualitative and quantitative evaluations run with external experts, and the model- and system-level mitigations (RLHF-based fine-tuning for helpfulness and harmlessness, monitoring, usage policies) adopted before deployment. It compares an early instruction-tuned version with the launch version and concludes that the mitigations reduce but do not eliminate misuse, calling for anticipatory planning and governance.
Abstract¶
Large language models (LLMs) are being deployed in many domains of our lives ranging from browsing, to voice assistants, to coding assistance tools, and have potential for vast societal impacts.[1, 2, 3, 4, 5, 6, 7] This system card analyzes GPT-4, the latest LLM in the GPT family of models.[8, 9, 10] First, we highlight safety challenges presented by the model’s limitations (e.g., producing convincing text that is subtly false) and capabilities (e.g., increased adeptness at providing illicit advice, performance in dual-use capabilities, and risky emergent behaviors). Second, we give a high-level overview of the safety processes OpenAI adopted to prepare GPT-4 for deployment. This spans our work across measurements, model-level changes, product- and system-level interventions (such as monitoring and policies), and external expert engagement. Finally, we demonstrate that while our mitigations and processes alter GPT-4’s behavior and prevent certain kinds of misuses, they are limited and remain brittle in some cases. This points to the need for anticipatory planning and governance.[11] Content Warning: This document contains content that some may find disturbing or offensive, including content that is sexual, hateful, or violent in nature. 41
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | analysis of safety risks of GPT-4 (hallucination, harmful content, disinformation, privacy, cybersecurity, dual-use capabilities, emergent behaviours) | llm | 0.85 |
| contribution | description of OpenAI's evaluation approach with external red-teaming and quantitative evaluations | llm | 0.85 |
| contribution | overview of model-level and system-level mitigations before deployment | llm | 0.85 |
| contribution | comparison of GPT-4-early and GPT-4-launch behaviour | llm | 0.85 |
| models | GPT-4 | llm | 0.85 |
| models | GPT-4-early | llm | 0.85 |
| models | GPT-4-launch | llm | 0.85 |
| tasks | language modeling | llm | 0.85 |
| tasks | safety evaluation | llm | 0.85 |
| tasks | red teaming | llm | 0.85 |
| interaction | other | llm | 0.85 |
| limitations | mitigations are limited and remain brittle in some cases | llm | 0.85 |
| limitations | custom fine-tuning and image capabilities are out of scope | llm | 0.85 |
| limitations | architecture, training data and compute are not disclosed | llm | 0.85 |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- model: GPT 253 [p.1], Transformer 1 [p.31]
- dataset: ARC 11 [p.3], Phishing 6 [p.13]
- technique: Fine-tuning 2 [p.2]
- task: Medical / healthcare 13 [p.5]
- topic: Large language models 17 [p.1], Open source 2 [p.12]
Related papers¶
- this paper misfiled as Large language models in medicine — the PDF was originally filed under this paper
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/misfiled/openai-2023_gpt-4-system-card.pdf |
misfiled | 60 | 990.8 KB | ca3677e1b83e |
Database key: gpt-4-system-card · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set gpt-4-system-card COLUMN VALUE