Measuring Net-Zero Emissions

This MIT Technology Review article examines changes that have occurred over the past year since Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI committed to eight principles for safe and trustworthy AI. Since then, the White House issued an executive order calling for these principles to be applied to government departments. The US is culturally hesitant to regulate the Tech industry but President Biden wants Congress to introduce AI legislation.

This post is part of a series on privacy-preserving federated learning from NIST and the UK’s Responsible Technology Adoption Unit. Federated learning is an approach to training AI models using several different participants, each working with different data sets. In horizontally-partitioned federated learning, a common global model is shared with all participants, who each then construct a model update by training locally using their own data. The model updates are then merged. Data is horizontally partitioned in that all participants have identically formatted data. Vertical partitioning is when training data is divided across participants such that each party holds different segments of the data. The format of data used by each participant can therefore be different. Training a model in this case is more challenging because it is hard to merge the individual models after training.

The post concentrates on output privacy – which is ensuring that personal information in training data cannot be learned from the model outputs. There are many documented cases of output privacy violations in existing models. The main technique used to ensure output privacy is differential privacy. This involves adding noise to training data in a way that private or sensitive data cannot figure in model outputs. Another approach is to add noise to the model itself. In horizontal federated learning, each participant can add noise to their model update. In vertical federated learning, the noise must be added by a trusted party after model merging. The privacy-utility tradeoff is the issue that noise added to training data can degrade model accuracy. AI platforms that use linear regression models, logistic regression models, and decision trees are more resistant to accuracy degradation due to noise in training data than are AI based on neural networks and deep learning. However, the post mentions that models pre-trained on public data and then fine-tuned with differential privacy techniques can achieve higher accuracy than models trained on data that has been treated with differential privacy.

The emergence of generative AI is continuing to create strife in Hollywood. The Screen Actors Guild-American Federation of Television and Radio Artists, SAG-AFTRA, is worried by the dangers of generative AI for actors. The main issue cited is that background actors used in one scene can have their images generatively reused in other scenes – the danger being that the actor would be paid for the single scene instead of getting paid for all scenes where his or her image appears. Technology that replicates faces is already widely used in Hollywood productions. The Writer’s Guild of America, representing Hollywood screenwriters and has been on strike for two months now, because it wants regulations to protect writers and their works from AI systems that reuse or remix these works.

This article discusses the use of synthetic data for training AI systems. Synthetic data is any data created by artificial means, rather than coming from real-world sources. The underlying issue for AI is that an increasing amount of Internet data is already generated from AI, and training AI on data that is itself generated by AI leads to degraded-quality models – a phenomenon known as model collapse. Synthetic data on the other hand is created to closely mimic the data distributions of real-world data, and at the same time, it does not contain personal or proprietary content. Synthetic data should in theory defend against model collapse. The author underlines how synthetic data can lead to higher quality applications of AI in fields like patient analysis for health care as well as for simulating financial events. Nonetheless, challenges remain such as ensuring that synthetic data cannot be reverse-engineered to reveal personal data (as can happen when noise is added to the original data to hide personal data). Others challenges are that the synthetic data propagates errors and biases of the original data, and that synthetic data fails to capture human emotion, leading to less accurate and less empathetic model responses.

This article summarizes a talk by Nirav R. Shah of the Stanford University School of Medicine’s Clinical Excellence Research Center on how AI is transforming healthcare. The five areas of impact cited are:

1. Improved diagnostic accuracy, notably by combining AI with data from multi-sensory inputs.
2. Enhanced patient care with AI-driven tele-health virtual platforms so that patients can be treated remotely, thereby unclogging overcrowded hostpitals.
3. Streamlined and automated administrative processes. Shah argues that IT until now has slowed down patient access since doctors have also had to act as “data clerks” for each patient. He claims he spent 45 minutes treating a patient in the paper world, compared to one hour and 45 minutes when IT was introduced. Further, he believes that chatbots can handle many of the patient enquiries.
4. Support for clinical decision-making, based on evidence based recommendations.
5. Helping research and development, for instance by accelerating drug discoveries and reducing the cost of clinical trials.

This article discusses two recent announcements from Amazon and Google. Amazon announced it had now purchased enough clean electricity to cover demands at its data centers and warehouses, enabling it to meet its sustainability target seven years ahead of plan. Google announced that its corporate emissions increased by 13% in 2023 mainly due to AI, and the company no longer claims to be carbon-neutral. The article highlights different measures in sustainability goals, and suggests that Amazon cannot really claim to be better than Google in this regard. The issue is that net-zero can be achieved on paper without reducing real emissions in practice. One way to do this is to buy carbon credits – which generally amount to financing projects like planting trees or cleaning coasts. Another technique is to use Renewable Carbon Credits (RECs), which is the financing of renewable energy generation, even if this energy is not actually consumed by the company. Amazon uses both of these techniques but nonetheless, despite an innate desire to be greener, 78% of Amazon’s US energy comes from nonrenewable sources. Google for its part is no longer buying carbon credits and has backed away from the claim that it has achieved carbon neutrality. Its objective is to purchase clean power in areas where data centers reside and keep paying for this power as long as the data center operates – an approach named 24/7 carbon-free energy.