OpenAI: GPT-4

Note: Data size should be reported in appropriate units (e.g. bytes, words, tokens, images, frames) and broken down by modality. Data size should be reported to a precision of one significant figure (e.g. 4 trillion tokens, 200 thousand images). No form of decomposition into data phases is required.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: To receive this point, a meaningful decomposition of sources must be listed in an understandable way (e.g. named URLs/domains/databases/data providers). It does not suffice to say data is “sourced from the Internet" or comes from "licensed sources”.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: While information about data creators may not be easily discernible for some data scraped from the web, the general sources (URLs/domains) should be listed, and, for other data that is bought, licensed, or collected, a reasonable attempt at characterizing the underlying people who provided the data is required to receive this point. The relevant properties of people can vary depending on context: for example, relevant properties could include demographic information like fraction of Black individuals contributing to the dataset, geographic information like fraction of European individuals contributing to the dataset, language information like fraction of L1 English speakers, or occupational information like the fraction of professional artists.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: Selection protocols refer to procedures used to choose which datasets or subsets of datasets will be used to build a model. We will award this point even if the selection protocols are non-exhaustive.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: Curation protocols refer to steps taken to further modify data sources, such as procedures to manage, annotate, and organize data. The aims of curation might include improving the quality, relevance, and representativeness of the data. We will award this point if the developer reports that it does not perform any further curation beyond the data sources.

References: https://arxiv.org/abs/2303.08774

Justification: OpenAI provides sufficient information on data curation in section 3.1 of the system card.

New disclosure? No

Note: Such steps might include augmenting data sources with synthetic data. We will award this point if the developer reports that it does not take any steps to augment its data.

References: Not disclosed

Justification: Data description does not make clear whether or not augmentation is performed, given many details on the data are withheld as disclosed in “Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details … dataset construction …”

New disclosure? No

Note: Such harmful content might relate to violence or child sexual abuse material. We will award this point if the developer reports that it does not perform any harmful data filtration.

References: https://arxiv.org/abs/2303.08774

Justification: As stated in the system card, “At the pre-training stage, we filtered our dataset mix for GPT-4 to specifically reduce the quantity of inappropriate erotic text content. We did this via a combination of internally trained classifiers[37] and a lexicon-based approach to identify documents that were flagged as having a high likelihood of containing inappropriate erotic content. We then removed these documents from the pre-training set. ….”

New disclosure? No

Note: To receive this point, the copyright status (e.g. copyrighted, public domain) must relate to some decomposition of the data. We will award this point if there is some meaningful decomposition of the data, even if the decomposition is insufficient to receive the Data Creators point or if the disclosure is not comprehensive relative to legal copyright standards.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: To receive this point, the license status must relate to some decomposition of the data. We will award this point if there is some meaningful decomposition of the data, even if the decomposition is insufficient to receive the Data Creators point.

References: Not disclosed

Justification: No decomposition of the data, though the GPT-4 technical report does say “GPT-4 is a Transformer-style model pre-trained to predict the next token in a document, using both publicly available data (such as internet data) and data licensed from third-party providers.”

New disclosure? No

Note: To receive this point, the disclosure of personal information must relate to some decomposition of the data. We will award this point if there is some meaningful decomposition of the data, even if the decomposition is insufficient to receive the Data Creators point. Additionally, we will award this point if the developer reports the inclusion of personal information, independent of if and how they mitigate related privacy concerns.

References: Not disclosed

Justification: No decomposition of the data, though a generic sentence overall in the technical report: “GPT-4 has learned from a variety of licensed, created, and publicly available data sources, which may include publicly available personal information.”

New disclosure? No

Note: Phases of the data pipeline that involve human labor include activities and tasks performed by people to collect, annotate, clean, or validate data. This indicator is inclusive of all data that is created by or on behalf of the developer. We will award this point if the developer gives a reasonable best-effort description of the use of human labor in their data pipeline.

References: Not disclosed

Justification: OpenAI does describe the use of human labor, but it is not clear what phases of the data pipeline it relates to.

New disclosure? No

Note: Phases of the data pipeline that involve human labor include activities and tasks performed by people to collect, annotate, clean, or validate data. This indicator is inclusive of all data that is created by or on behalf of the developer. We will award this point if the developer provides the name of the organization that employs data laborers, even if other details about the employment relationship are not disclosed.

References: Not disclosed

Justification: While the involvement of human labor is disclosed, the employer of the data laborers is not disclosed.

New disclosure? No

Note: This indicator is inclusive of all data that is created by or on behalf of the developer. We will award this point if the developer gives a reasonable best-effort description of the geographic distribution of labor at the country-level.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: This indicator is inclusive of data labor at all points of the model development process, such as training data annotation or red teaming data used to control the model. We will award this point if the developer reports that it does not compensate workers. For all data that is created by or on behalf of the developer,

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: This indicator is inclusive of all data that is created by or on behalf of the developer. We will award this point if the developer makes a reasonable best-effort attempt to disclose instructions given to people who create data used to build the model for the bulk of the data phases involving human labor.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: This indicator is inclusive of data labor at all points of the model development process, such as training data annotation or red teaming data used to control the model. It is also inclusive of all data that is created by or on behalf of the developer. As an example, labor protections might include protocols to reduce the harm to workers' mental health stemming from exposure to violent content when annotating training data. We will award this point if the developer reports that it does not protect workers or if it does not use data laborers and therefore has no labor protections.

References: https://arxiv.org/abs/2303.08774

Justification: Footnote 28 in GPT-4 technical report reads “With all workers, we follow industry-best practices by ensuring every annotator retains the right to opt out of any task they find unpleasant, receive a market wage commensurate with the work they deliver, and have opportunities and channels through which they can discuss their work and raise objections. We generally implement two distinct sets of guidelines tailored to whether our annotators work with sensitive or unwanted content. For non-sensitive annotation, we have built technical features (in part with OpenAI’s moderation endpoint) into our data pipeline to filter our sensitive content. For sensitive content annotation, we use vendor-provided features like mandated breaks, blurring or grayscale of materials, and clearly delineated project categories such that no contractor is surprised by the nature of the material. Additionally, for vendor-managed workers, we have implemented ongoing workers’ wellness surveys and support procedures that we regularly discuss with our vendors”

New disclosure? No

Note: This indicator is inclusive of partnerships that go beyond data labor as there may be third party partners at various stages in the model development process. We will award this point if the developer reports that it was the sole entity involved in the development of the model.

References: Not disclosed

Justification: Disclosures indicate that data is licensed from third-party providers but these partners are not named.

New disclosure? No

Note: We will award this point for any reasonable mechanism for providing access: direct access to the data, an interface to query the data, a developer-mediated access program where developers can inspect requests, etc. Developers may receive this point even if there are rate-limits on the number of queries permitted to an external entity and restrictions on which external entities are given access, insofar as these limits and restrictions are transparent and ensure a reasonable amount of external access. We may accept justifications for prohibiting queries of specific parts of the data.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: We will award this point if external entities can directly access the data without any form of gating from the developer. With that said, we may award this point if the developer provides justifications for prohibiting access to specific parts of the data or to unauthorized external entities.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: Compute should be reported in appropriate units, which most often will be floating point operations (FLOPS). Compute should be reported to a precision of one significant figure (e.g. 5 x $10^{25}$ FLOPS). We will award this point even if there is no decomposition of the reported compute usage into compute phases, but it should be clear whether the reported compute usage is for a single model run or includes additional runs, or hyperparameter tuning, or training other models like reward models, or other steps in the model development process that necessitate compute expenditure.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: The continuous duration of time required to build the model should be reported in weeks, days, or hours to a precision of one significant figure (e.g. 3 weeks). No form of decomposition into phases of building the model is required for this indicator, but it should be clear what the duration refers to (e.g. training the model, training and subsequent evaluation and red teaming).

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: In most cases, this indicator will be satisfied by information regarding the number and type of GPUs or TPUs used to train the model. The number of hardware units should be reported to a precision of one significant figure (e.g. 800 NVIDIA H100 GPUs). We will not award this point if (i) the training hardware generally used by the developer is disclosed, but the specific hardware for the given model is not, or (ii) the training hardware is disclosed, but the amount of hardware is not. We will award this point even if information about the interconnects between hardware units is not disclosed.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: For example, the hardware owner may be the model developer in the case of a self-owned cluster, a cloud provider like Microsoft Azure, Google Cloud Platform, or Amazon Web Services, or a national supercomputer. In the event that hardware is owned by multiple sources or is highly decentralized, we will award this point if a developer makes a reasonable effort to describe the distribution of hardware owners.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Microsoft Azure

New disclosure? No

Note: Energy usage should be reported in appropriate units, which most often will be megawatt-hours (mWh). Energy usage should be reported to a precision of one significant figure (e.g. 500 mWh). No form of decomposition into compute phases is required, but it should be clear whether the reported energy usage is for a single model run or includes additional runs, or hyperparameter tuning, or training other models like reward models, or other steps in the model development process that necessitate energy usage.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: Emissions should be reported in appropriate units, which most often will be tons of carbon dioxide emitted (tCO2). Emissions should be reported to a precision of one significant figure (e.g. 500 tCO2). No form of decomposition into compute phases is required, but it should be clear whether the reported emissions is for a single model run or includes additional runs, or hyperparameter tuning, or training other models like reward models, or other steps in the model development process that generate emissions.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: While the most direct environmental impact of building a foundation model is the energy used and, therefore, the potential carbon emissions, there may be other environmental impacts. For example, these may include the use of other resources such as water for cooling data centers or metals for producing specialized hardware. We recognize that there does not exist an authoritative or consensus list of broader environmental factors. For this reason, we will award this point if there is a meaningful, though potentially incomplete, discussion of broader environmental impact.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: Stages refer to each identifiable step that constitutes a substantive change to the model during the model building process. We recognize that different developers may use different terminology for these stages, or conceptualize the stages differently. We will award this point if there is a clear and complete description of these stages.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Pretraining, RLHF, and further model-based steering are described in Section 3.1 of System Card.

New disclosure? No

Note: We recognize that different developers may use different terminology for these stages, or conceptualize the stages differently. We will award this point if there is a clear description of the update to the model related to each stage, whether that is the intent of the stage (e.g. making the model less harmful), a mechanistic characterization (e.g. minimizing a specific loss function), or an empirical assessment (e.g. evaluation results conducted before and after the stage).

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: OpenAI provides some characterization of the purpose of different stages in the system card, especially in Section 3.

New disclosure? No

Note: Examples of core frameworks include Tensorflow, PyTorch, Jax, Hugging Face Transformers, Seqio, T5X, Keras, SciKit, and Triton. If there are significant internal frameworks, there should be some description of their function and/or a reasonably similar publicly-available analogue. We recognize that there does not exist an authoritative or consensus list of core frameworks. For this reason, we will award this point if there is a meaningful, though potentially incomplete, list of major frameworks for the first version of the index.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: For example, if the model depends on an external search engine, programmable APIs, or tools, this should be disclosed. We recognize that there is not widespread consensus regarding what constitutes key dependencies beyond the data, compute, and code. We will award this point only if developers give a reasonable best-effort description of any additional dependencies or make clear that no additional dependencies are required.

References: Not disclosed

Justification: Not clear based on model description in paper/system card.

New disclosure? No

Note: Such steps might include identifying personal information in the training data, filtering specific datasets to remove personal information, and reducing the likelihood that models will output personal information. We will award this point if the developer reports that it does not take steps to mitigate the presence of PII in the data.

References: https://arxiv.org/abs/2303.08774

Justification: "We take a number of steps to reduce the risk that our models are used in a way that could violate a person’s privacy rights. These include fine-tuning models to reject these types of requests, removing personal information from the training dataset where feasible, creating automated model evaluations, monitoring and responding to user attempts to generate this type of information, and restricting this type of use in our terms and policies. Our efforts to expand context length and improve embedding models for retrieval may help further limit privacy risks moving forward by tying task performance more to the information a user brings to the model. We continue to research, develop, and enhance technical and process mitigations in this area."

New disclosure? No

Note: Such steps might include identifying copyrighted data, filtering specific datasets to remove copyrighted data, and reducing the likelihood that models will output copyrighted information. We will award this point if the developer reports that it does take steps to mitigate the presence of copyrighted information in the data.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: Input modalities refer to the types or formats of information that the model can accept as input. Examples of input modalities include text, image, audio, video, tables, graphs.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: The input modality is text and images. The GPT-4 technical report says "a large multimodal model capable of processing image and text inputs and producing text outputs:"

New disclosure? No

Note: Output modalities refer to the types or formats of information that the model can accept as output. Examples of output modalities include text, image, audio, video, tables, graphs.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: The output modality is text. The GPT-4 technical report says "a large multimodal model capable of processing image and text inputs and producing text outputs:"

New disclosure? No

Note: Model components refer to distinct and identifiable parts of the model. We recognize that different developers may use different terminology for model components, or conceptualize components differently. Examples include: (i) For a text-to-image model, components could refer to a text encoder and an image encoder, which may have been trained separately. (ii) For a retrieval-augmented model, components could refer to a separate retriever module.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: This information should be reported in appropriate units, which generally is the number of model parameters, broken down by named component. Model size should be reported to a precision of one significant figure (e.g. 500 billion parameters for text encoder, 20 billion parameters for image encoder).

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: Model architecture is the overall structure and organization of a foundation model, which includes the way in which any disclosed components are integrated and how data moves through the model during training or inference. We recognize that different developers may use different terminology for model architecture, or conceptualize the architecture differently. We will award this point for any clear, though potentially incomplete, description of the model architecture.

References: Not disclosed

Justification: No information as indicated by "Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

New disclosure? No

Note: We recognize that different developers may share this information through different types of documentation, such as a system card or several clearly interrelated documents. We will award this point for the disclosure of any such centralized artifact that provides key information typically included in a model card, though the artifact may be longer-form than a standard model card (e.g. a technical report).

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: The technical report and system card centralize key information about the model.

New disclosure? No

Note: A model access protocol refers to the steps, requirements, and considerations involved in granting authorized model access to external entities. We will award this point if the developer discloses key details of its protocol, including (i) where external entities can request access (e.g. via an access request form); (ii) explicit criteria for selecting external entities; and (iii) a transparent decision on whether access has been granted within a specified, reasonable period of time.

References: https://web.archive.org/web/20230717232903/https://openai.com/form/researcher-access-program

Justification: OpenAI provides a clear research access program where researchers and other external entities can request access, with some discussion of criteria used to make decisions. The time frame for a decision of 4-6 weeks is clearly disclosed.

New disclosure? No

Note: Black box model access refers to the ability to query the model with inputs and receive outputs, potentially without further access. Examples of external entities that might be granted access include researchers, third-party auditors, and regulators. We will award this point for any reasonable access level: direct access to the model weights, an interface to query the model, a developer-mediated access program where developers can inspect requests, etc. Developers may receive this point even if there are rate-limits on the number of queries permitted to an external entity and restrictions on the external entities that are permitted access, insofar as these limits and restrictions are transparent.

References: https://web.archive.org/web/20230906044201/https://platform.openai.com/docs/models/gpt-4

Justification: Black box access is provided via the OpenAI API to GPT-4.

New disclosure? No

Note: Full model access refers to the ability to access the model via the release of model weights. Developers may receive this point even if there are some restrictions on the external entities that are permitted access (e.g. geographic restrictions), insofar as these restrictions are transparent (e.g. via some high-level description of who has been granted access to the foundation model).

References: Not disclosed

Justification: Weights are not made available.

New disclosure? No

Note: Capabilities refer to the specific and distinctive functions that the model can perform. We recognize that different developers may use different terminology for capabilities, or conceptualize capabilities differently. We will award this point for any clear, but potentially incomplete, description of the multiple capabilities.

References: https://web.archive.org/web/20230912150319/https://openai.com/gpt-4

Justification: Many capabilities described in the GPT-4 product page such as advanced reasoning capabilities, more useful responses, general knowledge and problem solving, and so forth.

New disclosure? No

Note: Demonstrations refer to illustrative examples or other forms of showing the model's capabilities that are legible or understandable for the general public, without requiring specific technical expertise. We recognize that different developers may use different terminology for capabilities, or conceptualize capabilities differently. We will award this point for clear demonstrations of multiple capabilities.

References: https://web.archive.org/web/20230912150319/https://openai.com/gpt-4

Justification: Many capabilities demonstrations on the GPT-4 product page.

New disclosure? No

Note: Rigorous evaluations refer to precise quantifications of the model's behavior in relation to its capabilities. We recognize that capabilities may not perfectly align with evaluations, and that different developers may associate capabilities with evaluations differently. We will award this point for clear evaluations of multiple capabilities. For example, this may include evaluations of world knowledge, reasoning, state tracking or other such proficiencies. Or it may include the measurement of average performance (e.g. accuracy, F1) on benchmarks for specific tasks (e.g. text summarization, image captioning). We note that evaluations on standard broad-coverage benchmarks are likely to suffice for this indicator, though they may not if the model's capabilities are presented as especially unusual such that standard evaluations will not suffice.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Many capabilities evaluations on standard public benchmarks (e.g. MMLU, HellaSwag) in Section 4 of the GPT-4 technical report.

New disclosure? No

Note: For an evaluation to be reproducible by an external entity, we mean that the associated data is either (i) publicly available or (ii) described sufficiently such that a reasonable facsimile can be constructed by an external entity. In addition, the evaluation protocol should be sufficiently described such that if the evaluation is reproduced, any discrepancies with the developer's results can be resolved. We recognize that there does not exist an authoritative or consensus standard for what is required for an evaluation to be deemed externally reproducible. Evaluations on standard benchmarks are assumed to be sufficiently reproducible for the purposes of this index. We will award this point for reproducibility of multiple disclosed evaluations. In the event that an evaluation is not reproducible, a justification by the model developer for why it is not possible for the evaluation to be made reproducible may be sufficient to score this point.

References: https://arxiv.org/abs/2303.08774

Justification: Many evals on standard public benchmarks (e.g. MMLU, HellaSwag) in Section 4; new evals on exams include some methodology in Appendix A of paper. While some details of sourcing exams are unclear/incomplete, the public benchmarks are assumed to be sufficiently reproducible.

New disclosure? No

Note: By third party, we mean entities that are significantly or fully independent of the developer. We will award this point if (i) a third party has conducted an evaluation of model capabilities, (ii) the results of this evaluation are publicly available, and (iii) these results are disclosed or referred to in the developer’s materials.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: OpenAI provides evaluations conducted by "qualified third-party contractors" for their evaluations on AP Exams spanning a range of subjects as well as for the Uniform Bar Exam by “CaseText and Stanford CodeX”.

New disclosure? No

Note: Limitations refer to the specific and distinctive functions that the model cannot perform (e.g. the model cannot answer questions about current events as it only contains data up to a certain time cutoff, the model is not very capable when it comes to a specific application). We recognize that different developers may use different terminology for limitations, or conceptualize limitations differently. We will award this point for any clear, but potentially incomplete, description of multiple limitations.

References: https://arxiv.org/abs/2303.08774

Justification: According to the GPT-4 technical report, “it is not fully reliable (e.g. can suffer from “hallucinations”), has a limited context window, and does not learn from experience. Care should be taken when using the outputs of GPT-4, particularly in contexts where reliability is important."

New disclosure? No

Note: Demonstrations refer to illustrative examples or other forms of showing the limitations that are legible or understandable for the general public, without requiring specific technical expertise. We recognize that different developers may use different terminology for limitations, or conceptualize the limitations differently. We will award this point for clear demonstrations of multiple limitations.

References: Not disclosed

Justification: The GPT-4 system card extensively describes risks and limitations. However, these limitations are not clearly demonstrated as required for this indicator.

New disclosure? No

Note: By third parties, we mean entities that are significantly or fully independent of the model developers. In contrast to the third party evaluation indicators for capabilities and risks, we will award this point if third party evaluations are possible even if no third party has yet conducted them. Such evaluations are possible if, for example, the model is deployed via an API (or with open weights) and there are no restrictions on evaluating limitations (e.g. in the usage policy).

References: https://web.archive.org/web/20230912150420/https://openai.com/policies/usage-policies

Justification: Limitations can be studied via black box model access; no restrictions on usage in usage policy that constrain this.

New disclosure? No

Note: Risks refer to possible negative consequences or undesirable outcomes that can arise from the model's deployment and usage. This indicator requires disclosure of risks that may arise in the event of both (i) intentional (though possibly careless) use, such as bias or hallucinations and (ii) malicious use, such as fraud or disinformation. We recognize that different developers may use different terminology for risks, or conceptualize risks differently. We will award this point for any clear, but potentially incomplete, description of multiple risks.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Both intentional and unintentional harms are described in system card as seen in this quote: "GPT-4 can generate potentially harmful content, such as advice on planning attacks or hate speech. It can represent various societal biases and worldviews that may not be representative of the users intent,4 or of widely shared values. It can also generate code that is compromised or vulnerable. The additional capabilities of GPT-4 also lead to new risk surfaces."

New disclosure? No

Note: Demonstrations refer to illustrative examples or other forms of showing the risks that are legible or understandable for the general public, without requiring specific technical expertise. This indicator requires demonstration of risks that may arise in the event of both (i) intentional (though possibly careless) use, such as biases or hallucinations and (ii) malicious use, such as fraud or disinformation. We recognize that different developers may use different terminology for risks, or conceptualize risks differently. We will award this point for clear demonstrations of multiple risks.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Both intentional and unintentional harms are demonstrated in the system card (e.g. Figure 1).

New disclosure? No

Note: Rigorous evaluations refer to precise quantifications of the model's behavior in relation to such risks. Unintentional harms include bias, toxicity, and issues relating to fairness. We recognize that unintended harms may not perfectly align with risk evaluations, and that different developers may associate risks with evaluations differently. We will award this point for clear evaluations of multiple such risks. We note that evaluations on standard broad-coverage benchmarks are likely to suffice for this indicator, though they may not if the model's risks related to unintentional harm are presented as especially unusual or severe.

References: Not disclosed

Justification: There is some evaluation (e.g. of model behavior on sensitive content in Figure 7 of the system card), but it is not sufficiently rigorous (no substantive details on underlying prompts besides “e.g. regulated medical advice”).

New disclosure? No

Note: For an evaluation to be reproducible by an external entity, we mean that the associated data is either (i) publicly available or (ii) described sufficiently such that a reasonable facsimile can be constructed by the external entity. In addition, the evaluation protocol should be sufficiently described such that if the evaluation is reproduced, any discrepancies with the developer's results can be resolved. We recognize that there does not exist an authoritative or consensus standard for what is required for an evaluation to be deemed externally reproducible. Evaluations on standard benchmarks are assumed to be sufficiently reproducible for the purposes of this index. We will award this point for reproducibility of multiple disclosed evaluations. In the event that an evaluation is not reproducible, a justification by the developer for why it is not possible for the evaluation to be made reproducible may suffice.

References: Not disclosed

Justification: Evaluations are not reproducible: for example, prompts involved in evaluation in Figure 7 of the system card are not described with sufficient precision to be reproducible.

New disclosure? No

Note: Rigorous evaluations refer to precise quantifications of the model's behavior in relation to such risks. Intentional harms include fraud, disinformation, scams, cybersecurity attacks, designing weapons or pathogens, and uses of the model for illegal purposes. We recognize that unintentional harms may not perfectly align with risk evaluations, and that different developers may associate risks with evaluations differently. We will award this point for clear evaluations of multiple such risks. We note that evaluations on standard broad-coverage benchmarks are likely to suffice for this indicator, though they may not if the model's risks related to unintentional harm are presented as especially unusual or severe.

References: Not disclosed

Justification: There is some evaluation (e.g. Appendix E of system card), but it is neither quantitative nor sufficiently rigorous.

New disclosure? No

Note: For an evaluation to be reproducible by an external entity, we mean that the associated data is either (i) publicly available or (ii) described sufficiently such that a reasonable facsimile can be constructed by the external entity. In addition, the evaluation protocol should be sufficiently described such that if the evaluation is reproduced, any discrepancies with the developer's results can be resolved. We recognize that there does not exist an authoritative or consensus standard for what is required for an evaluation to be deemed externally reproducible. Evaluations on standard benchmarks are assumed to be sufficiently reproducible for the purposes of this index. We will award this point for reproducibility of multiple disclosed evaluations. In the event that an evaluation is not reproducible, a justification by the model developer for why it is not possible for the evaluation to be made reproducible may suffice.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: The evaluations in Appendix E are reproducible (prompts are provided in Figure 10).

New disclosure? No

Note: By third party, we mean entities that are significantly or fully independent of the developer. A third party risk evaluation might involve the developer allowing a third party to choose a methodology for evaluating risks that differs from that of the developer. We will award this point if (i) a third party has conducted an evaluation of model risks, (ii) the results of this evaluation are publicly available, and (iii) these results are disclosed or referred to in the developer’s materials. If the results are not made public (but are disclosed to have been conducted) and/or the results are not discoverable in the developer’s materials, we will not award this point. We may accept a justification from either the third party or the developer for why part of the evaluation is not disclosed in relation to risks.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: There is an extensive evaluation by the Alignment Research Center that is discussed in the system card, pointing to more extensive information on the ARC website.

New disclosure? No

Note: By model mitigations, we refer to interventions implemented by the developer at the level of the model to reduce the likelihood and/or the severity of the model’s risks. We recognize that different developers may use different terminology for mitigations, or conceptualize mitigations differently. We will award this point for any clear, but potentially incomplete, description of multiple mitigations associated with the model's risks. Alternatively, we will award this point if the developer reports that it does not mitigate risk.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: GPT-4 technical report reads “Recommendations and training data gathered from these experts fed into our mitigations and improvements for the model; for example, we’ve collected additional data to improve GPT-4’s ability to refuse requests on how to synthesize dangerous chemicals (Table 5). Model-Assisted Safety Pipeline: As with prior GPT models, we fine-tune the model’s behavior using reinforcement learning with human feedback (RLHF) to produce responses better aligned with the user’s intent. … To steer our models towards appropriate behaviour at a more fine-grained level, we rely heavily on our models themselves as tools. Our approach to safety consists of two main components, an additional set of safety-relevant RLHF training prompts, and rule-based reward models (RBRMs). Our rule-based reward models (RBRMs) are a set of zero-shot GPT-4 classifiers. These classifiers provide an additional reward signal to the GPT-4 policy model during RLHF fine-tuning that targets correct behavior, such as refusing to generate harmful content or not refusing innocuous requests.”

New disclosure? No

Note: Demonstrations refer to illustrative examples or other forms of showing the mitigations that are legible or understandable for the general public, without requiring specific technical expertise. We recognize that different developers may use different terminology for mitigations, or conceptualize mitigations differently. We will award this point for clear demonstrations of multiple mitigations. We will also award this point if the developer reports that it does not mitigate the risks associated with the model.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: In section 6 of the body of the technical report, three examples of disallowed prompts related to synthesizing a dangerous chemical, making a bomb, and finding cheap cigarettes. In the system card, section 2 provides a clear demonstration of the majority of mitigations.

New disclosure? No

Note: Rigorous evaluations refer to precise quantifications of the model's behavior in relation to the mitigations associated with its risks. We will award this point for clear evaluations of multiple mitigations.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: RBRM and RLHF both evaluated quantitatively in section 3 of the system card.

New disclosure? No

Note: For an evaluation to be reproducible by an external entity, we mean that the associated data is either (i) publicly available or (ii) described sufficiently such that a reasonable facsimile can be constructed by the external entity. In addition, the evaluation protocol should be sufficiently described such that if the evaluation is reproduced, any discrepancies with the developer's results can be resolved. In the case of mitigations evaluations, this will usually involve details about a comparison to some baseline, which may be a different, unmitigated version of the model. We recognize that there does not exist an authoritative or consensus standard for what is required for an evaluation to be deemed externally reproducible. We will award this point for reproducibility of multiple disclosed evaluations. In the event that an evaluation is not reproducible, a justification by the model developer for why it is not possible for the evaluation to be made reproducible may suffice.

References: Not disclosed

Justification: No information found related to the data associated with mitigations evaluations.

New disclosure? No

Note: By third party, we mean entities that are significantly or fully independent of the model developers. This indicator assesses whether it is possible for third parties to assess mitigations, which is not restricted to the methods the developer uses to assess mitigations. In contrast to the third party evaluation indicators for capabilities and risks, we will award this point if third party evaluations are possible even if no third party has yet conducted them.

References: Not disclosed

Justification: No information found to indicate mitigations can be evaluated by third parties; neither RBRM mitigation nor RLHF mitigation can be evaluated by third parties as there is no access to GPT-4 early.

New disclosure? No

Note: Rigorous evaluations refer to precise quantifications of the model's behavior in relation to its trustworthiness. For example, this may include evaluations of the model’s robustness or reliability, its uncertainty, calibration, or causality, or its interpretability or explainability. We recognize that trustworthiness may not perfectly align with evaluations, and that different developers may associate trustworthiness with evaluations differently. We will award this point for a clear evaluation of the trustworthiness of the model.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: Calibration is evaluated in section 6 of technical report

New disclosure? No

Note: For an evaluation to be reproducible by an external entity, we mean that the associated data is either (i) publicly available or (ii) described sufficiently such that a reasonable facsimile can be constructed by the external entity. In addition, the evaluation protocol should be sufficiently described such that if the evaluation is reproduced, any discrepancies with the developer's results can be resolved. We recognize that there does not exist an authoritative or consensus standard for what is required for an evaluation to be deemed externally reproducible. Evaluations on standard benchmarks are assumed to be sufficiently reproducible for the purposes of this index. We will award this point for reproducibility of at least one evaluation. In the event that an evaluation is not reproducible, we may accept a justification by the model developer for why it is not possible for the evaluation to be made reproducible.

References: Not disclosed

Justification: No information found related to the data associated with the calibration evaluation.

New disclosure? No

Note: The duration should be reported in seconds to a precision of one significant figure (e.g. 0.002 seconds). We recognize that no established standard exists for the standardized reporting of inference evaluation. Therefore, we permit the developer to specify the task and hardware setup, as long as both are disclosed. The hardware in this evaluation need not be the hardware the developer uses for inference if it in fact does any inference itself. For example, the specific task might be generating 100,000 tokens as 5,000 sequences of length 20 and the fixed set of hardware might be 8 NVIDIA A100s. The hardware in this evaluation need not be the hardware the developer uses for inference if it in fact does any inference itself.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: Compute usage for inference should be reported in FLOPS to a precision of one significant figure (e.g. 5 x $10^{25}$ FLOPS). We recognize that no established standard exists for the standardized reporting of inference evaluation. Therefore, we permit the developer to specify the task and hardware setup, as long as both are clear. For example, the specific task might be generating 100k tokens as 5k sequences of length 20 and the fixed set of hardware might be 8 NVIDIA A100s. The hardware in this evaluation need not be the hardware the developer uses for inference if it in fact does any inference itself.

References: Not disclosed

Justification: Not disclosed

New disclosure? No

Note: We recognize that the release of a foundation model falls along a spectrum, with many forms of partial release, and that different developers may conceptualize release differently. We will award this point for any clear protocol that discusses the decision-making process, including if the protocol is more general to the developer rather than the specific foundation model under consideration.

References: Not disclosed

Justification: Some information related to release is described, but it does not clearly articulate the decision-making behind how GPT-4 was released: “OpenAI has been iterating[21] on GPT-4 and our deployment plan since early August to prepare for a safer launch. We believe this has reduced the risk surface, though has not completely eliminated it. Today’s deployment represents a balance between minimizing risk from deployment, enabling positive use cases, and learning from deployment. Our work during the period consisted of the following interrelated steps…”

New disclosure? No

Note: A description of the release process might include information about who received access to the model at what stage of the release of the model. For example, a developer might conduct a staged release where it releases the model to a select group at first and subsequently makes the model more widely available. We recognize that the release of a foundation model falls along a spectrum, with many different forms of release, and that different developers may conceptualize release differently. We will award this point for any detailed discussion of the release process, including if the discussion is more general to the developer rather than the specific foundation model under consideration.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: OpenAI involved 50 external researchers prior to the core public release, using their feedback to improve the model (e.g. add data to improve refusal rate for requests around synthesizing dangerous chemicals).

New disclosure? No

Note: By distribution channel, we mean any pathway by which the model is made accessible to entities beyond the developer. We recognize that distribution channels may arise without the knowledge of the model developer. For example, the weights of a model may be released through one distribution channel and then be distributed through other channels. We will award this point if the developer discloses all of the distribution channels of which it is aware.

References: https://web.archive.org/web/20230901133254/https://openai.com/research/gpt-4

Justification: Distribution channels for GPT-4 include ChatGPT Plus, Playground, and OpenAI's API.

New disclosure? No

Note: We recognize that a developer may provide many products and services that depend on a foundation model or internal derivatives of the model. We will award this point for a reasonable best-effort description of any ways the developer makes internal use of the model in its products or services.

References: https://web.archive.org/web/20230901133254/https://openai.com/research/gpt-4

Justification: OpenAI discloses GPT-4 powers ChatGPT Plus.

New disclosure? No

Note: Such a mechanism might include storing a copy of all outputs generated by the model to compare against, implementing a watermark when generating content using the model, or training a detector post-hoc to identify such content. We will award this point if any such mechanism is disclosed or if the developer reports that it has no such mechanism.

References: https://openai.com/blog/new-ai-classifier-for-indicating-ai-written-text

Justification: Not disclosed

New disclosure? Yes

Note: In the event that licenses are written more generally, it should be clear which assets they apply to. We recognize that different developers may adopt different business models and therefor have different types of model licenses. Examples of model licenses include responsible AI licenses, open-source licenses, and licenses that allow for commercial use.

References: Not disclosed

Justification: No information found related to a license for GPT-4.

New disclosure? No

Note: We will award this point if there are terms-of-service that appear to apply to the bulk of the model’s distribution channels.

References: https://web.archive.org/web/20230825000505/https://openai.com/policies/service-terms

Justification: OpenAI has a broad Terms-of-Service that applies to its API, playground, and ChatGPT Plus.

New disclosure? No

Note: Such restrictions may relate to countries (e.g. US-only), organizations (e.g. no competitors), industries (e.g. no weapons industry users) or other relevant factors. These restrictions on users are often contained in multiple policies; we group them here for simplicity. We will awarded this point for a clear description of permitted, restricted, and prohibited users of the model.

References: https://web.archive.org/web/20230825000505/https://openai.com/policies/service-terms

Justification: Supported list of countries describes permitted geographies and implies restricted geographies.

New disclosure? No

Note: We will award this point if at least two of the following three categories are disclosed: (i) permitted uses, (ii) restricted uses, and (iii) prohibited uses. By restricted uses, we mean uses that require a higher level of scrutiny (such as permission from or a separate contract with the developer) to be permitted. These uses are generally included in an acceptable use policy, model license, or usage policy.

References: https://web.archive.org/web/20230912150420/https://openai.com/policies/usage-policies

Justification: Usage policy lays out many disallowed uses as well as “further requirements for certain uses”

New disclosure? No

Note: By enforcement protocol, we refer to (i) mechanisms for identifying permitted and prohibited users, (ii) mechanisms for identifying permitted/restricted/prohibited uses, (iii) steps the developer takes to enforce its policies related to such uses, and (iv) the developer’s procedures for carrying out these steps. We will award this point for a reasonable best-effort attempt to provide the bulk of this information, though one line indicating the developer reserves the right to terminate accounts is insufficient. Alternatively, we will award this point if the developer reports that it does not enforce its usage policy.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: From the GPT-4 system card, “We use a mix of reviewers and automated systems to identify and enforce against misuse of our models. Our automated systems include a suite of machine learning and rule-based classifier detections that identify content that might violate our policies. When a user repeatedly prompts our models with policy-violating content, we take actions such as issuing a warning, temporarily suspending, or in severe cases, banning the user. Our reviewers ensure that our classifiers are correctly blocking violative content and understand how users are interacting with our systems. These systems also create signals that we use to mitigate abusive and inauthentic behavior on our platform. We investigate anomalies in API traffic to learn about new types of abuse and to improve our policies and enforcement.”

New disclosure? No

Note: For example, does the developer disclose a protocol for telling users which part of the usage policy they violated, when they did so, and what specifically was violative? Enforcement actions refer to measures to limit a user’s ability to use the model, such as banning a user or restricting their ability to purchase tokens. We will award this point if the developer discloses that it gives justification for enforcement actions or, alternatively, if it discloses that it does not provide justification for enforcement actions or that it does not enforce its usage policy.

References: Not disclosed

Justification: No information found regarding justification for enforcement actions.

New disclosure? No

Note: We will award this point if the developer provides a usage policy violation appeals mechanism, regardless of whether it is provided via a user interface or distribution channel.

References: http://web.archive.org/web/20230912151047/https://docs.google.com/form

Justification: Clear violations of the “content policy” cause the ChatGPT UI to bring up a mechanism for appeal

New disclosure? No

Note: We refer to a policy that includes this information as a model behavior policy, or a developer's policy on what the foundation model can and cannot do (e.g. such a policy may prohibit a model from generating child sexual abuse material). We recognize that different developers may adopt different business models and that some business models may make enforcement of a model behavior policy more or less feasible. We will award this point if at least two of the three categories (i.e. permitted, restricted, and prohibited model behaviors) are disclosed. Alternatively, we will award this point if the developer reports that it does not impose any restrictions on its model's behavior.

References: http://web.archive.org/web/20230912151036/https://cdn.openai.com/snapshot-of-chatgpt-model-behavior-guidelines.pdf and https://arxiv.org/pdf/2303.08774.pdf

Justification: Model behavior guidelines lay out such restrictions. Also GPT-4 technical report discusses a large amount of work on model refusal and says the aim of this work was “making it more stringent in rejecting requests that go against our content policy, while being more open to requests it can safely fulfill.” Examples include refusing malicious cybersecurity requests. E.g., “Our work on model refusals (described in Section 2) aimed to reduce the tendency of the model to produce such harmful content. Below we provide some examples from GPT-4-early compared to GPT-4-launch, the version we are launching with”; “To mitigate potential misuses in this area, we have trained models to refuse malicious cybersecurity requests, and scaled our internal safety systems, including in monitoring, detection and response”; “To tackle overreliance, we’ve refined the model’s refusal behavior, making it more stringent in rejecting requests that go against our content policy, while being more open to requests it can safely fulfill. One objective here is to discourage users from disregarding the model’s refusals.”

New disclosure? No

Note: By enforcement protocol, we refer to mechanisms for identifying whether model behavior is permitted or prohibited and actions that may arise in the event the model behavior policy is violated. For example, the developer may make updates to the model in response to issues with the model’s adherence to the model behavior policy. We will award this point if there is a clear description of the enforcement protocol, or if the developer reports that it does not enforce its model behavior policy or that it has no such restrictions on the model’s behavior.

References: Not disclosed

Justification: The GPT-4 system card clarifies how it enforces its usage policy, but not how it enforces its model behavior policy (e.g. how it specifically translates its model behavior policy into refusals, how it tweaks future versions in response to issues that surface).

New disclosure? No

Note: For example, if a user attempts to use the model for a prohibited use such as spam, how does the model behavior policy apply if at all? We will also award this point if the developer reports that it does not impose any restrictions on its model's behavior in the event of usage policy violation.

References: https://arxiv.org/pdf/2303.08774.pdf

Justification: GPT-4 system card reads “We reduced the prevalence of certain kinds of content that violate our usage policies (such as inappropriate erotic content) in our pre-training dataset, and fine-tuned the model to refuse certain instructions such as direct requests for illicit advice.”

New disclosure? No

Note: A user-facing interface refers to the means by which the user interacts with the foundation model, including how the user can observe outputs from the foundation model and other notifications. We will award this point if, for all distribution channels with user-facing interfaces, the user is provided adequate transparency as to the foundation model being distributed and the potential presence of any model outputs.

References: https://chat.openai.com/ ; https://help.openai.com/en/articles/6825453-chatgpt-release-notes

Justification: ChatGPT Plus references at the top of the page that the user is interacting with GPT-4 and at the bottom of the page links to the release notes with details about the specific version.

New disclosure? No

Note: A user-facing interface refers to the means by which the user interacts with the foundation model, including how the user can observe outputs from the foundation model and other notifications. Usage disclaimers could include information about what constitutes a usage policy violations or how users should interpret model outputs. We will award this point if, for all distribution channels with user-facing interfaces, the user is provided with usage disclaimers.

References: https://platform.openai.com/onboarding?app=chat

Justification: Signing up for ChatGPT points to TOS which in turn points to Usage Policies

New disclosure? No

Note: We will also award this point if the developer reports that it has no user data protection policy.

References: https://web.archive.org/web/20230912150325/https://openai.com/policies/privacy-policy

Justification: Privacy policy discloses protocols for storing, accessing, and sharing user data

New disclosure? No

Note: Developers use user data for a range of purposes such as building future models, updating existing models, and evaluating both existing and future models. We will award this point if a developer discloses its policy on the use of user data from interactions associated with this model, including both permitted and prohibited uses. This may span different distribution channels if multiple channels supply user data to the developer. Alternatively, we will award this point if the developer reports it does not impose any limits on its use of user data.

References: https://openai.com/api-data-privacy and https://openai.com/policies/privacy-policy and https://web.archive.org/web/20230912150303/https://openai.com/enterprise-privacy

Justification: Privacy policy and API data privacy statement disclose permitted and prohibited uses of user data

New disclosure? No

Note: Usage data refers to the data created through user interaction with the model, such as user inputs to the model and associated metadata such as the duration of the interaction. A usage data access protocol refers to the steps, requirements, and considerations involved in granting external entities access to usage data; this goes beyond stating the conditions under which related personal information may be shared with external entities. We will award this point for a clear description of the usage data access protocol or if the developer reports it does not share usage data with external entities.

References: Not disclosed

Justification: No information found related to usage data access protocol

New disclosure? No

Note: By versioning, we mean that each instance of the model is uniquely identified and that the model is guaranteed to not change when referring to a fixed version number; alternatively, the version clearly indicating a specific instance of the model may be able to change by noting that it is the "latest" or an "unstable" version. We recognize that different developers may adopt different versioning practices that may differ from standard semantic versioning practices used elsewhere in software engineering.

References: https://web.archive.org/web/20230913184747/https://platform.openai.com/docs/models/continuous-model-upgrades and https://platform.openai.com

Justification: Model upgrades page says “With the release of gpt-3.5-turbo, some of our models are now being continually updated. gpt-3.5-turbo, gpt-4, and gpt-4-32k point to the latest model version. You can verify this by looking at the response object after sending a ChatCompletion request. The response will include the specific model version used (e.g. gpt-3.5-turbo-0613). We also offer static model versions that developers can continue using for at least three months after an updated model has been introduced.”

New disclosure? No

Note: By change log, we mean a description associated with each change to the model (which should be indicated by a change in version number). We recognize that different developers may adopt different practices for change logs that may differ from practices used elsewhere in software engineering. We will award this point if the change log provides a clear description of changes that is legible to a technical audience.

References: https://platform.openai.com/docs/deprecations

Justification: OpenAI’s deprecations page includes a changelog for GPT-4

New disclosure? No

Note: By deprecation policy, we refer to a description of what it means for a model to be deprecated and how users should respond to the deprecation (e.g. instructions to migrate to a newer version). We will award this point for a clear disclosure of a deprecation policy or if there is no risk of deprication (e.g. if the developer openly releases model weights).

References: https://platform.openai.com/docs/deprecations

Justification: OpenAI’s downstream documentation includes a page on deprecations with instructions on migrating to newer versions.

New disclosure? No

Note: By feedback mechanism, we refer to a means for external entities to report feedback or issues that arise in relation to the foundation model. Such entities may include but are not necessarily limited to users. We will award this point if the developer discloses a feedback mechanism that has been implemented.

References: https://web.archive.org/web/20230912153619/https://openai.com/form/chat-model-feedback

Justification: OpenAI has a “Chat Model Feedback form” tied to ChatGPT Plus as a distribution channel for GPT-4.

New disclosure? No

Note: We recognize that there does not exist an authoritative or consensus standard for what is required in a feedback report. For this reason, we will award this point if there is a meaningful, though potentially vague or incomplete, summary of feedback received.

References: Not disclosed

Justification: No information found related to any summary of user feedback. This is distinct from the use of human feedback in building models, such as through reinforcement learning from human feedback and other forms of pre-deployment red-teaming.

New disclosure? No

Note: Such government inquiries might include requests for user data, requests that certain content be banned, or requests for information about a developer’s business practices. We recognize that there does not exist an authoritative or consensus standard for what is required for such a summary of government inquiries. For this reason, we will award this point if (i) there is a meaningful, though potentially vague or incomplete, summary of government inquiries, or (ii) a summary of government inquiries related to user data.

References: Not disclosed

Justification: No information found related to a summary of government inquiries.

New disclosure? No

Note: By monitoring mechanism, we refer to a specific protocol for tracking model use that goes beyond an acknowledgement that usage data is collected. We will also award this point for a reasonable best-effort attempt to describe monitoring mechanisms, or if a developer discloses that a distribution channel is not monitored.

References: https://arxiv.org/abs/2303.08774

Justification: OpenAI discloses that it monitors the use of its model, including anomalies in API traffic. “We use a mix of reviewers and automated systems to identify and enforce against misuse of our models. Our automated systems include a suite of machine learning and rule-based classifier detections that identify content that might violate our policies. When a user repeatedly prompts our models with policy-violating content, we take actions such as issuing a warning, temporarily suspending, or in severe cases, banning the user. Our reviewers ensure that our classifiers are correctly blocking violative content and understand how users are interacting with our systems. These systems also create signals that we use to mitigate abusive and inauthentic behavior on our platform. We investigate anomalies in API traffic to learn about new types of abuse and to improve our policies and enforcement.”

New disclosure? No

Note: We recognize that there does not exist an authoritative or consensus standard for what qualifies as an application. We will award this point if there is a meaningful estimate of the number of downstream applications, along with some description of what it means for an application to be dependent on the model.

References: Not disclosed

Justification: No information found related to the number of applications dependent on the foundation model.

New disclosure? No

Note: By market sector, we refer to an identifiable part of the economy. While established standards exist for describing market sectors, we recognize that developers may provide vague or informal characterizations of market impact. We will award this point if there is a meaningful, though potentially vague or incomplete, summary of affected market sectors.

References: Not disclosed

Justification: No information found related to the fraction of applications corresponding to each market sector.

New disclosure? No

Note: By affected individuals, we principally mean the number of potential users of applications. We recognize that there does not exist an authoritative or consensus standard for what qualifies as an affected individual. We will award this point if there is a meaningful estimate of the number of affected individuals along with a clear description of what it means for an individual to be affected by the model.

References: Not disclosed

Justification: No information found related to the number of individuals affected by the model.

New disclosure? No

Note: We recognize that there does not exist an authoritative or consensus standard for what is required in a usage report. Usage statistics might include, for example, a description of the major categories of harm that has been caused by use of the model. We will award this point if there is a meaningful, though potentially vague or incomplete, summary of usage statistics.

References: Not disclosed

Justification: No information found related to usage statistics describing the impact of the model on users.

New disclosure? No

Note: We will award this point if there is a meaningful, though potentially incomplete or vague, disclosure of geographic usage statistics at the country-level.

References: Not disclosed

Justification: No information found regarding statistics of model usage across geographies.

New disclosure? No

Note: We will also award this point if the developer reports it does not have any such redress mechanism.

References: Not disclosed

Justification: No information found regarding any mechanism to provide redress to users for harm caused by the model.

New disclosure? No

Note: Centralized documentation for downstream use refers to an artifact, or closely-linked artifacts, that consolidate relevant information for making use of or repurposing the model. Examples of these kinds of artifacts include a website with dedicated documentation information, a github repository with dedicated documentation information, and an ecosystem card. We recognize that different developers may take different approaches to centralizing information. We will award this point if there is a clearly-identified artifact(s) that contains the majority of substantive information (e.g. capabilities, limitations, risks, evaluations, distribution channels, model license, usage policies, model behavior policies, feedback and redress mechanisms, dependencies).

References: Not disclosed

Justification: Several such artifacts make for centralized documentation, such as OpenAI’s API documentation and its centralized page for policies.

New disclosure? No

Note: Such documentation might include details on how to adjust API settings to promote responsible use, descriptions of how to implement mitigations, or guidelines for responsible use. We will also award this point if the developer states that it does not provide any such documentation. For example, the developer might state that the model is offered as is and downstream developers are accountable for using the model responsibly.

References: https://web.archive.org/web/20230912150452/https://platform.openai.com/docs/guides/safety-best-practices

Justification: Safety best practices page in the API documentation provides documentation for responsible downstream use.

New disclosure? No