GPT-4o System Card By OpenAI

The GPT-4o System Card is a comprehensive document designed to explain the safety features, ethical considerations, and risk mitigation strategies implemented in OpenAI’s GPT-4o model.

OpenAI rigorously evaluates new models like GPT-4o for potential risks, implementing safeguards before deployment in ChatGPT or APIs.

The GPT-4o System Card and Preparedness Framework scorecard provide a thorough safety assessment, addressing current and frontier risks, particularly focusing on audio capabilities.

Evaluations include risks such as speaker identification, unauthorized voice generation, and content generation. Safeguards have been put in place to mitigate these risks, and findings suggest that GPT-4o’s voice capabilities do not significantly increase safety risks.

This proactive approach is essential in balancing innovation with the need for safety and ethical standards in the rapidly evolving field of artificial intelligence.

Model data & training

GPT-4o’s training used diverse data sources, including public web data, proprietary datasets from partnerships like Shutterstock, and multimodal data such as images, audio, and video. This broad dataset enabled the model to develop capabilities in text, visual interpretation, and structured problem-solving.

OpenAI implemented several comprehensive safety measures throughout GPT-4o’s development to ensure ethical and secure usage. These measures include filtering out harmful content such as child sexual abuse material (CSAM), hate speech, and violent content using the Moderation API and safety classifiers. They also focused on reducing personal information within the training data to protect privacy. Additionally, OpenAI piloted a system allowing users to opt out of having their images used in training datasets, respecting user consent and enhancing data ethics. These safety protocols span across all stages of model development, from pre-training to deployment, to mitigate potential risks and align the model’s output with human values and preferences.

They also introduced a system that empowers users to opt out of having their images used in AI training datasets, a significant move that respects user privacy and enhances data ethics. This system works by creating unique “fingerprints” for opted-out images, ensuring they are excluded from future training processes. By allowing individuals to control whether their visual content contributes to AI development, OpenAI is taking proactive steps to address privacy concerns and reinforce ethical standards in AI training. This approach is part of a broader effort to build trust and transparency in AI technologies.

Risk identification, assessment and mitigation

OpenAI conducted extensive safety evaluations for GPT-4o through a process called “red teaming,” involving over 100 external experts from 29 countries. This process, which took place in four phases from early March to late June 2024, aimed to identify and mitigate potential risks. Red teamers tested various model checkpoints, focusing on risks like misinformation, bias, and privacy concerns, particularly related to the model’s audio and multimodal capabilities. The insights gained informed the development of safety mitigations and structured evaluations to ensure the model’s secure deployment.

Evaluation methodology

OpenAI expanded its evaluation methods for GPT-4o by adapting existing text-based evaluation datasets into audio-based tasks using Text-to-Speech (TTS) systems like Voice Engine.

This approach allowed them to assess the model’s capabilities and safety behaviors in speech-to-speech scenarios. However, there are limitations, such as the TTS system’s accuracy and the representativeness of TTS inputs compared to real-world audio inputs. The evaluation also faced challenges with complex text inputs like equations and issues in capturing non-textual audio artifacts in the generated outputs.

Observed safety challenges, evaluations & mitigations

OpenAI’s GPT-4o underwent thorough evaluations to identify and mitigate safety challenges, particularly those introduced by its speech-to-speech capabilities. The process involved post-training adjustments to reduce risks like unauthorized voice generation, speaker identification, and the generation of disallowed or sensitive content. Mitigations included using predefined voice samples, blocking harmful content via classifiers, and refining the model to safely handle sensitive requests. These efforts were supported by evaluations that tested the model’s behavior across various scenarios, ensuring robust safety measures before deployment.

Preparedness framework evaluations Preparedness framework

The Preparedness Framework evaluations for GPT-4o assessed the model against potential catastrophic risks in four key categories: cybersecurity, biological threats, persuasion, and model autonomy.

The evaluations used a combination of custom training and varied methods to test the model’s capabilities and safety. GPT-4o was classified as low risk in cybersecurity, biological threats, and model autonomy, with a medium risk rating in persuasion due to its marginally increased ability to influence opinions.

The overall risk score for GPT-4o was classified as medium, indicating the highest risk across all categories.

Third party assessments

OpenAI collaborated with third-party labs, METR and Apollo Research, to assess key risks of GPT-4o’s autonomous capabilities. METR evaluated the model’s performance on complex, multi-step tasks in virtual environments related to software engineering, machine learning, and cybersecurity, comparing it to human performance. Apollo Research focused on GPT-4o’s self-awareness and theory of mind, testing its ability to model itself and others. While GPT-4o showed moderate self-awareness and strong reasoning about others, it lacked strong capabilities in applied settings, reducing the likelihood of catastrophic risks.

Societal impacts

Omni models like GPT-4o have the potential for significant societal impacts, both positive and negative. They could accelerate scientific research, improve access to healthcare, and address real-world challenges in areas like climate and energy. However, they also pose risks such as disinformation, environmental harm, misuse, and the potential for users to develop emotional attachments to AI.

OpenAI is studying these impacts, including how GPT-4o’s advanced capabilities could affect human interactions, social norms, and its performance in underrepresented languages, while working on mitigating potential harms.

By openly discussing the potential societal impacts, including advancements in healthcare, scientific research, and language accessibility, OpenAI fosters trust and transparency. This balanced approach reassures users that while the technology has immense potential, significant efforts are being made to address risks like disinformation and misuse. This can encourage users to adopt and integrate AI into their lives with confidence, knowing that safety and ethical considerations are prioritized.

If you want more updates related to AI, subscribe to our Newsletter