The European Union is regulating artificial intelligence (AI) systems under incorrect assumptions that could result in undesirable outcomes. The EU is assuming more data is always useful to AI and incentivising “big data” approach.
AI is not restricted to data-hungry techniques. There are several alternative techniques that rely on less data, require less computation and less memory. They are more in line with the principle of data minimisation in the GDPR. The Commission should take a smarter approach and incentivise AI techniques that require less data, that benefit society and that assists with climate change mitigation.
EU Commission believes “AI can only thrive when there is smooth access to data [emphasis added].”1 It believes that it is in “the nature of AI” to rely on “large and varied datasets [emphasis added].”2 It believes more data automatically produces more insight and understanding. The AI Act, the Data Governance Act and the Data Act assume that more data is needed for AI.3 These beliefs are wrong.
What matters is not how large the datasets are, but the statistical properties of datasets such as how varied the data are. The context in which data is collected is also important. However, the Commission seems to believe that all data is interchangeable. Instead of focusing on “re-use, sharing and pooling of data,”4 the Commission should recognise that more data is not always useful to AI.
“The more the data, the surer we fool ourselves.”5 Quantity of data is not a metric of AI innovation. Many AI techniques do not require large amounts of data. In fact, innovation in mathematical techniques enable powerful AI models based on small datasets.6
It is true that some AI techniques are data hungry. These techniques, such as deep neural networks, demand vast computation and memory, which are expensive. They need an infrastructure to collect, store and process lots of data. Despite all these resources, these data-hungry AI techniques produce diminishing returns with more data.7 Exorbitant amounts of data are needed to make expensive marginal improvements.
The process of discovering a large model that performs well involves a range of steps that include searching and finding the optimal parameters, testing them, and training multiple prototypes. The cost of merely training a large model with a set of parameters that has already been optimized by another team could cost more than $2 million. A recently published research on large models trained more than 1000 variations to find one optimised model.8
Furthermore, large models require large memory. Limited memory limits the model size. Specialized hardware with large memory can accelerate the training process. However, specialized hardware required to train the models could cost in the range hundreds of thousands of dollars. Even worse, specialized hardware does not generalize to different computation tasks, thus requiring further investment for newer models.
This means that the actual cost of discovering just one large model that performs well could be in millions or even billions of dollars. Currently, such costs are only affordable to a small set of companies such as Facebook, Google and OpenAI. Such costs may create barriers for small and medium sized enterprises (SMEs).
We should also consider the carbon footprint of training a large model, which is estimated to be about five times the carbon footprint of a car over its lifetime. This, along with large amounts of data being used to surveil humans, is an immense cost for our society to pay for marginal gains in performance of data-hungry AI models. Thus, we should incentivise the use of AI techniques that require less data.