In this blog

Machine Learning offers improvements to each of the three Environmental, Social and Governance components of ESG. As the focus on ESG gains momentum, harnessing the power of advanced technologies becomes essential to navigate the complexities of sustainable practices. With advances in Machine Learning and Artificial Intelligence, streamlining and enhancing ESG initiatives has never been easier. 


Machine Learning can be leveraged to reduce emissions, but is also responsible for emissions

ML models can be used to identify patterns in data center usage driving electricity consumption, facility cooling, or high loads taxing performance. Once these models identify the major drivers of increased energy consumptionIT teams can reduce these drivers and optimize hardware electricity use.

However, Machine Learning can also be a source of emissions. Before starting to decrease the environmental impacts of Machine Learning, it's important to know the existing footprint. There are a few tools for doing this, like the Machine Learning Emissions Calculator.

When and where ML operations are performed can significantly impact a model's emissions profile

The time-of-day electricity is consumed has a major impact on the emissions generated. Electric grids continuously generate electricity, however the demand for this electricity varies throughout the day, and the weekScheduling models to update in off-peak hours can lead to substantial carbon emission reductions.

The physical location of ML operations is another driver of emissions. Leveraging the cloud can significantly reduce the emissions generated by ML, since large-scale data centers operated for multiple companies are generally more efficient than on-prem operations. If the cloud is not an appropriate fit, optimizing the location of hardware performing ML operations can be another powerful way to reduce emissions. Choosing to locate data centers in energy grids with lower carbon intensities[1] can reduce emissions without changing user behavior. Many resources identify the emissions of different electric grids.

Map showing the relative carbon intensities of different electric grids in the contiguous United States on 4/7/2023, greener means
Map showing the relative carbon intensities of different electric grids in the contiguous United States on 4/7/2023, greener means "cleaner"


Best practices lower the carbon footprint of models

Training choices can dramatically affect emissions from Machine Learning. Selecting hyperparameter values to use in training is a major driver of carbon emissions from ML. If there is no literature available to help establish good values, randomly looking for relevant ones is actually better than using grid search. Overall, optimizing model training provides significant ways to reduce a major source of emissions, like using pre-trained models to eliminate all emissions from training. If pre-trained models are not an option, shortening necessary training procedures will also reduce emissions. Longer training time means more power drawn, and thus increased emissions, in addition to longer hardware use, forcing an enterprise to purchase additional hardware to meet its needs. Carefully choosing efficient makes and models of hardware to train on also has major impacts. For example, CPUs can be 10% as efficient as GPUs while TPU 3 can be 4-8x more efficient than GPUs.

There are many ways to improve code to reduce its carbon impact. For instance, limiting SQL queries run inside a loop and loops, and using static and local variables are some easy ways to make models greener.

Finally, developers should compare models to each other to select the most efficient.. When creating a model that is planned to retrain downstream (e.g. on a new domain, fine-tuning on a new task, etc.), developers should report the model's training time and required computational resources, as well as model sensitivity to hyperparameters. Reporting allows users to choose the cleaner option when deciding between alternative models.


Addressing social issues with AI + ML solutions

Machine Learning solutions can address social issues in a range of ways – deep learning models predict lead-contaminated water pipes in Flint, Michigan; speech AI platforms raise accessibility for audio recognition for those with auditory impairments.  AI algorithms are crucial in the healthcare industry as well, elevating our social wellbeing through predictive models identifying early-stage cancer, diagnose liver diseases, and heart attacks. During the wake of COVID-19, ML Models were particularly helpful in predicting how the pandemic will evolve and affect different regions to direct medical efforts and inform travel guidelines

Democratizing AI across the board

From predictive models that address social challenges to technologies that transform our day-to-day, ML solutions can be applied to social challenges and enhance daily routines.  Democratization of AI brings opportunities for smaller companies and individuals to harness the power of AI, leveling the playing field and fostering innovation as well as productivity. One example of the proliferation of machine learning models is Chat-GPT, a Large Language Model (LLM) developed by OpenAI that can be used for anything from language translation, text summarization, and content generation. Read the many use cases for Chat-GPT related to ESG goals here.

ML frameworks, in particular ML for Healthcare, play a crucial role in revolutionizing healthcare. By leveraging vast amounts of patient data to create more accurate diagnostic tools, personalized treatment plans, and predictive models for disease outbreaks, the way healthcare can be proactive and predictive is significant. For example, a large-scale study demonstrated the impact of an early warning system to enable early detection of sepsis, one of the leading causes of in-hospital deaths. When deployed successfully, MLH can improve patient outcomes by helping clinicians make decisions faster and more accurately. 

Data availability, accuracy and security is an ongoing challenge

One main consideration when solving complex social challenges is data availability, accuracy, and security. Large datasets that go into creating these solutions are complicated and require protection. Potential problems associated with gathering data are twofold – (1) lack of data availability; (2) skewed datasets that contribute to biased results. 

Private and public organizations are often unwilling or unable to share data due to regulations on data use and privacy considerations. Datasets with personal information some models require also come at a high cost that most non-profit organizations may not be able to pay. Lack of data availability can lead to skewed datasets composed of problematic historical data and unrepresentative datasets. Using these datasets in ML models can cause biases deepening existing social inequalities, reinforcing cultural prejudices, and unfairly targeting specific populations.

A wide sample set is also crucial in making solutions more accessible. Speech recognition tools can enhance how we interact with audio and speech – live captioning, meeting transcriptions and even virtual assistants not only make content accessible for people with varying levels of hearing and vision but can also transform ways users interact with different products. Speech recognition AI platforms should be able to accommodate regional dialects and accents to make tools more available, not determine one type of "speech" as a norm over another. 

Effective AI solutions leveraging ML need to enforce rigorous validation best practices at all stages of the process and ensure proper data security and governance practices are maintained, all while making datasets available to NGOs serving communities.

Solutions need to center transparency and accountability to community 

Furthermore, it is crucial to center transparency and accountability to communities these tools are developed to serve. For instance, in 2018, more than 100 civil rights, digital justice, and community based organizations – who represented millions of people impacted by mass incarceration – released and signed a shared statement highlighting concerns with the adoption of algorithmic based decision making tools as a substitution for ending money bail.

At every step of the model building and training process, researchers should identify and address biases, and seek to ensure that privacy and data security is enforced at various stages of the ETL pipeline. Outputs of solutions should also not consider a panacea for intersectional and complex social problems.

It is crucial to implement robust governance frameworks, ensuring fairness, interpretability, and accountability in models. Additionally, the deployment of AI in critical areas like healthcare and finance necessitates careful consideration of privacy, security, and regulatory compliance. As AI continues to advance, it is essential to navigate these social implications responsibly to foster a future where AI benefits all of society. 


"Hasta la vista, baby" to privacy abuse

From Hollywood portrayals of AI taking over the world in The Terminator and Ex Machina to AI model demographic biases, ethical concerns and artificial intelligence have long been intertwined. Improved automation and generative AI tools like ChatGPT have led to real AI ethical concerns. Data privacy and consumer consent are gaining traction in the individual's favor, as years of data collection on the individual level, with minimal laws governing online privacy, have led to multiple large-scale invasions of privacy

MLOps emphasizes tracking and documenting changes in the machine learning pipeline. Maintaining auditability is essential for governance, so organizations can understand who makes decisions and that processes fall in line with compliance and ethical standards.

Companies directly benefit from governance regulations

MLOps best practices can help companies align their machine learning initiatives with their organization's core values and bottom line. Creating this alignment reinforces governance structure and bolsters the organization's commitment to ESG standards. Ranking high in ESG score can directly improve revenue, both through tax benefits and cost cuts, as well as increased consumer appreciation. 

Consumer interest has grown enough that multiple organizations have come up with scoring systems to rank companies based on ESG practice effectiveness. For example, the MSCI ESG Ratings measure a company's resilience to long-term ESG risks, the Dow Jones Sustainability Index allows investors to invest in an index based on sustainability and environmental practices, and the Corporate Knights Global 100 ranks the world's most sustainable corporations each year. These scoring systems provide crucial tools for investors to have more informed insight when assessing potential investments and which companies are utilizing sustainable and responsible practices. Sound MLOps governance practices allow organizations to address biases, ensure fairness, align with ESG principles, and adhere to ethical guidelines. Extreme examples of companies trending away from sound ethical operations, such as the 2015 Volkswagen emissions scandalFIFA misleading fans and more, all stem from organizations that strayed too far from proper governance and compliance processes. 

ML Governance encourages fostering education and model transparency

With machine learning models permeating people's lives more and more in this digital world—from uncovering treatments to COVID-19 to the Google search engine—increased awareness around model training and biases is invaluable. 

Strong ML governance encourages training and ongoing education for individuals and teams on machine learning projects. Teaching team members about the latest ethical considerations and best practices improves knowledge around data privacy, as well as increasing model transparency.

Machine Learning can be a valuable tool for organizations with ESG goals 

Machine Learning offers improvements in each of the three realms of ESG. Optimizing hardware electric use by modifying ML code and performing operations best practices can reduce emission output. Standardizing rules and regulations around ethical, effective practices for ML and AI models smooth their introduction while addressing current concerns around maintaining an ethical AI and data privacy culture. Protecting our people and our planet through sustainable and responsible business practices is becoming a necessary part of business. With the significance of ESG becoming increasingly visible, business leaders need to use Machine Learning solutions thoughtfully to meet the evolving expectations of stakeholders.

Read more to learn how WWT's Machine Learning and ESG capabilities are ready to help you drive more sustainable business outcomes.