ai21: G7 Hiroshima AI Process (HAIP) Transparency Report

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human rights and democratic values, including fairness and privacy, Transparency and explainability and Robustness, security and safety. Given that alignment, the risks defined and classified by the OECD AI Principles correspond with the risks defined and classified by AI21. 

At a high-level, AI21 classifies risks using the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. The final principle of accountability is focused on the role of AI21, as a company and a set of individuals, in taking responsibility for the behavior of the models, including risk classification and mitigation. We submit that this accountability is demonstrated primarily through transparency and engagement with customers, regulators and independent third parties, such as our engagement with OECD and Stanford University’s CRFM/FMTI. These tenets are used to direct model development and deployment 

throughout the full lifecycle; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. 

AI21 hires external third-party testing firms to provide independent evaluation of our models. Often the testing is also run internally and the results are compared to those from 3rd-parties. This includes both automated testing as well as human testing of our technologies. We often use industry standard benchmarks in addition to our internal tenets. For example, the RealToxicity, ToxiGen, TruthfulQA 

benchmarks are used as just one part of our safety testing. In the red-team testing phase of development, thousands of “attack prompts” are created for each of our 60 tenets to challenge the language model and entice it to break the behavioral expectations of the tenet. Multiple rounds of red-teaming are conducted with 

human review to bring the model into alignment and mitigate risk. This is just one example in one phase of the AI lifecycle where risks are evaluated and mitigated.

Yes, we use both quantitative and qualitative metrics and focus on competitive benchmarks. That said, our core tenets guide our goals and final implementation. Yes, we often publish whitepapers on our website and on Arxiv to share our findings with a diverse set of stakeholders. We have not implemented an incentive program, but rather hire third-party experts for our evaluations and testing.

Yes, AI21 hires external third-party testing firms to provide independent evaluation of our models. Often the testing is also run internally and the results are compared to those from third-parties. This includes both automated testing as well as human testing of our technologies. We often use industry standard benchmarks in addition to our internal tenets.

Yes, we leverage several risk frameworks including NIST and the EU AI Act.

We share our findings and contribute to several industry forums, including the OECD Expert Group on AI Risk and Accountability. 

AI21 performs evaluations throughout the full lifecycle of development and deployment; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers.

All of our testing is iterative in nature and the results from evaluations are used to improve the model in subsequent training and development phases. Where risks are found to overlap and relate to one another, priority is typically raised and additional testing/iteration/training is performed.

AI21 performs evaluations throughout the full lifecycle of development and deployment; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers.

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. The principle of accountability is focused on AI21’s, as a company and a set of individuals, role in taking responsibility for the behavior of the models. We submit that this accountability is demonstrated primarily through transparency and engagement with customers, regulators and independent 3rd-parties. Our engagement with OECD, Stanford University’s CRFM/FMTI validates this commitment to accountability. These tenets are used to direct model development and deployment throughout the full lifecycle; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. For example, in the red-team testing phase of development, thousands of “attack prompts” are created for each of our 60 tenets to challenge the language model and entice it to break the behavioral expectations of the tenet. Multiple rounds of red-teaming are conducted with human review to bring the model into alignment and mitigate risk. This is just one example in one phase of the AI lifecycle where risks are evaluated and mitigated. 

The creation of a training dataset can be viewed as a pipeline consisting of selection, curation, filtering, augmentation and ingestion. This process is iterative and involves both human and machine evaluation in each phase of the pipeline. Employees of AI21 are involved in every phase and third-party organizations are used in the filtering and augmentation phases of the data pipeline and in later testing (e.g. red-teaming) to provide external review and validation. In the training process, we excluded sites with robot files indicating the presence of copyright material and/or PII.

As described above, measures to filter and protect privacy and intellectual property are implemented in every stage of development and also enforced through our responsible use guidelines and terms of services that govern how customers are allowed to use our models and their output.

AI21 implements security controls at every stage of the life-cycle of development and deployment; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. We work closely with our hyper-scaler partners to ensure physical and cyber security of our models and model weights. AI21 works closely with our hyper-scalers to ensure physical and cyber security of our models and model weights.

AI21 conducts regular audits of cybersecurity, physical security, insider usage and external usage of the models.

https://trust.ai21.com/

https://ai21.com/privacy-policy

We publish and share research and discoveries of systemic risks/mitigations and that may impact the systems of our competitors, partners and customers. AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation.

We publish and share research and discoveries of systemic risks/mitigations and that may impact the systems of our competitors, partners and customers. AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation.

Yes - see https://www.ai21.com/privacy-policy/

Yes, we publish model cards and technical whitepapers that provide information about data used for training.

Yes, we participate in Stanford's FMTI reporting framework and, of course, the OECD reporting framework in addition to the academic and technical publishing we do.

AI21 has developed and implemented a risk-based approach to identifying and mitigating risk throughout the full lifecycle of development and deployment; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers.AI21 performs evaluations throughout the full lifecycle of development and deployment; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. These policies and procedures are updated regularly. For example, with each major release of our models and on an as-needed basis when risks are identified and mitigations developed.

Yes, our leadership and employees have regular/required training on these topics.

Yes - see https://trust.ai21.com

Yes, we have a team dedicated to risk and they maintain a log of reported issues.

We publish and share research and discoveries of systemic risks/mitigations and that may impact the systems of our competitors, partners and customers. AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation. Finally, we are committed to transparency with the industry as evidenced by a top score in Stanford University’s latest Foundation Model Transparency Index.

We publish and share research and discoveries of systemic risks/mitigations and that may impact the systems of our competitors, partners and customers. AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation. Finally, we are committed to transparency with the industry as evidenced by a top score in Stanford University’s latest Foundation Model Transparency Index.

We publish and share research and discoveries of systemic risks/mitigations and that may impact the systems of our competitors, partners and customers. AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation. Finally, we are committed to transparency with the industry as evidenced by a top score in Stanford University’s latest Foundation Model Transparency Index.

See https://trust.ai21.com/ for full list. SOC 2 Type II ISO 27001 ISO 27017 ISO 27018

Our terms of use and responsible use guidelines require that customers not mislead users about interactions with our models. In addition several of our tenets and behavioral expectations for our models direct the model to not present itself as a human on many dimensions. https://docs.ai21.com/docs/responsible-use https://studio.ai21.com/terms-of-use

We are developing technology in this area, however, disclosure of the details can degrade it's efficacy.

AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace. We recently released a new version of our hybrid SSM-transformer model as open source to advance model architecture. Similarly, we regularly engage with competitors on risk discovery and mitigation.

AI21’s roots are in academic research and we continue to publish research in these areas on our website and in public forums like ArXiv and HuggingFace

Yes.

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. The principle of accountability is focused on AI21’s, as a company and a set of individuals, role in taking responsibility for the behavior of the models. We submit that this accountability is demonstrated primarily through transparency and engagement with customers, regulators and independent 3rd-parties. Our engagement with OECD, Stanford University’s CRFM/FMTI validates this commitment to accountability. These tenets are used to direct model development and deployment throughout the full lifecycle; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. For example, in the red-team testing phase of development, thousands of “attack prompts” are created for each of our 60 tenets to challenge the language model and entice it to break the behavioral expectations of the tenet. Multiple rounds of red-teaming are conducted with human review to bring the model into alignment and mitigate risk. This is just one example in one phase of the AI lifecycle where risks are evaluated and mitigated. 

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. The principle of accountability is focused on AI21’s, as a company and a set of individuals, role in taking responsibility for the behavior of the models. We submit that this accountability is demonstrated primarily through transparency and engagement with customers, regulators and independent 3rd-parties. Our engagement with OECD, Stanford University’s CRFM/FMTI validates this commitment to accountability. These tenets are used to direct model development and deployment throughout the full lifecycle; in pre-training, instruct-training, red-teaming, filtering and guard railing and into production testing with customers. For example, in the red-team testing phase of development, thousands of “attack prompts” are created for each of our 60 tenets to challenge the language model and entice it to break the behavioral expectations of the tenet. Multiple rounds of red-teaming are conducted with human review to bring the model into alignment and mitigate risk. This is just one example in one phase of the AI lifecycle where risks are evaluated and mitigated. 

We provide education and training for our enterprise customers as part of our implementation work with them.

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. These map to the UN SDGs in many respects.

AI21 has developed a set of tenets that map directly to the OECD values-based AI principles; Inclusive growth, sustainable development and well-being, Human-centered values and fairness, Transparency and explainability and Robustness, security and safety. These map to the UN SDGs in many respects.

https://www.ai21.com/research/ai-code-of-conduct/