Magazine #2 | Summer 2023
Step by Step Towards Sustainable AI
The sustainability of an AI system depends on many decisions taken during its lifecycle. At the SustAIn project, we have developed criteria that help us illustrate which steps should be heeded.
Phase 1: Planning and Design
To a significant degree, the sustainability of an AI system is determined during the design and planning stage. As such, careful consideration should be given from the start as to whether the development and deployment of a complex and resource-intensive AI system is even necessary for a particular task. If it is, then additional fundamental questions must be asked: What data must the system process and how? What are the risks? And what security measures will be implemented?
General Requirements Checklist
Code of Conduct: A Code of Conduct sets out values and standards, such as transparency and equity, that provide guidance to all participants during planning and development.
Stakeholder Participation: All stakeholders related to the AI system must be identified and consulted during planning.
Documentation: Important decisions about the system must be documented in conjunction with its functions.
Risk Management: Once potential risks have been identified, appropriate safety measures must be taken.
Responsibilities: Where responsibility lies for the results produced by the AI system must be clearly defined.
Social Requirements Checklist
Transparency and Responsibility: A determination must be made about what information is disclosed and how responsibility for the system is distributed.
Non-Discrimination and Fairness: When humans are affected by an AI system’s decisions, the fairness of its application must be determined. In addition, measures to eliminate bias and discriminatory consequences must be defined.
Technical Reliability and Human Supervision: Possible technical risks must be identified. A determination must be made about how harmful system operations can be remedied.
Self-Determination and Data Protection: Appropriate data protection measures must be implemented to ensure that data subjects know how their personal data is used. They must be given the opportunity to withhold personal data.
Inclusive and Participatory Design: The design must be barrier-free and inclusive according to applicable standards.
Cultural Sensitivities: Local knowledge assets (from stakeholders, for example) must be integrated into the development process. Team diversity is essential.
Environmental Requirements Checklist
Energy Consumption: Requirements for the necessary performance of the model and the resource budget must be defined, metrics for capturing energy efficiency established and test procedures for early detection of failing experiments developed.
Greenhouse Gas Emissions: Measures for offsetting CO2 emissions must be defined and tools for recording emissions must be identified.
Sustainability in Use: Positive and negative potentials must be identified, analyzed and, if possible, quantified. Key performance indicators must be defined to record whether the potentials are being fully utilized.
Indirect Resource Consumption: Certifications and efficiency metrics for data centers must be the basis for ensuring that the hardware and the data centers used meet sustainability requirements.
Economic Requirements Checklist
Working Conditions and Jobs: When deployed in the workplace, impacts of the planned AI system on the workforce, working conditions and potential job losses must be analyzed and, if necessary, minimized.
Phase 2: Data
Data is at the core of all machine learning-based AI systems. System sustainability depends on the quality and amount of data to be processed, its interpretation and the way in which the data is obtained.
General Requirements Checklist
Documentation: Data sheets must be implemented to document information about the data collected and used and make it publicly available.
Stakeholder Participation: Individuals whose data will be used during the training or use of the AI systems must be consulted to ensure data is interpreted in an appropriate manner.
Social Requirements Checklist
Non-Discrimination and Fairness: Datasets must be examined for possible bias.
Technical Reliability: It is essential to ensure that current, complete, representative and reliable data is used.
Self-Determination and Data Protection: To protect personal data, principles such as data minimization, encryption, aggregation and anonymization must be adopted.
Environmental Requirements Checklist
Energy Consumption: A determination must be made on how much data is needed to train and operate the systems. If necessary, methods must be developed to minimize the amount of data needed.
Economic Requirements Checklist
Working Conditions and Jobs: When digital crowdworkers label datasets, they must be paid fair wages and be provided with good working conditions.
Phase 3: Development
How is an AI model developed? Which model should be selected? How is the model trained? These questions have far-reaching implications, particularly for the social and environmental sustainability of AI systems. Energy consumption must be minimized during the development phase. Risks must also be contained.
General Requirements Checklist
Responsibility: A body set up for this purpose should ensure compliance with the Code of Conduct during the development and evaluation of the models.
Documentation: Model cards must be created to document the experiments performed during development and the corresponding key performance indicators.
Stakeholder Participation: Individuals who may be affected by the AI systems or who are expected to use the AI systems must be involved in development to ensure the prevention of negative outcomes.
Social Requirements Checklist
Non-Discrimination and Fairness: Tools and evaluation criteria for measuring fairness must be implemented.
Technical Reliability and Human Supervision: Model weaknesses must be identified and corrected, and control measures defined.
Inclusive and Participatory Design: Inclusive design principles must be observed, including in the design of user interfaces.
Cultural Sensitivities: Systems must be developed to be applicable in different local contexts – through training with local datasets, for example.
Environmental Requirements Checklist
Energy Consumption: The energy efficiency of systems must be recorded. Preference should be given to pre-trained models and models with lower complexity. Methods should be deployed for optimizing energy efficiency.
Greenhouse Gas Emissions: CO2 emissions generated in the development process must be rigorously recorded. Selection of appropriate training locations and timing will increase CO2 efficiency.
Sustainability in Use: A system’s resource consumption must be recorded and compared to its sustainability potential.
Indirect Resource Consumption: The resource efficiency of the hardware must be recorded and optimized.
Phase 4: Implementation
During the deployment and operation of AI systems, as well as during additional training runs, risks to data subjects must be monitored and steps taken to ensure data protection. During the application phase, the consumption of energy and resources must be kept low.
General Requirements Checklist
Stakeholder Participation: Stakeholders must be consulted on new releases.
Risk Management: The potential risks identified must be monitored, and any new risks that may arise must be identified.
Social Requirements Checklist
Transparency and Responsibility: Stakeholders and end users must be informed about the use and operation of the AI system.
Non-Discrimination and Fairness: The AI system’s decisions must be evaluated to determine whether they meet previously established fairness standards.
Self-Determination and Data Protection: Persons whose data is used must receive information about that use. Simple consent or revocation options must be guaranteed.
Inclusive and Participatory Design: Barrier-free accessibility as well as access for disadvantaged groups must be ensured.
Environmental Requirements Checklist
Energy Consumption: Energy consumption during deployment must be documented and optimized.
Greenhouse Gas Emissions: CO2 emissions must be recorded and optimized during use.
Sustainability in Use: It must be determined whether the use of the AI system can be made more sustainable by, for example, conserving resources.
Indirect Resource Consumption: Resource efficiency of the hardware must be recorded and optimized.
Economic Requirements Checklist
Working Conditions and Jobs: When deployed in the workplace, impacts of the system on the workforce, working conditions and potential job losses must be determined and, if need be, minimized.
