Potential Impact of Data-Centric AI on Society

Model-centric AI aims to produce an optimal model for a given data set, with the objective of maximizing its performance on a specific task [4]. This is achieved through various model modifications, such as adjusting the loss function, fine-tuning the hyper-parameters, and employing optimization techniques, all aimed at enhancing the model’s overall performance, accuracy, precision, and generalization abilities. The accessibility of deep learning models through open-source communities and the availability of paid application programming interfaces (APIs) have made it more convenient for researchers and developers to utilize optimized models across various domains. Despite these advancements, ensuring data quality control remains a persistent challenge. Consequently, placing emphasis on tackling data-related issues becomes crucial to prevent inaccurate outcomes and enhance the ability of AI models to handle real-world complexities effectively.

In contrast, DCAI focuses on improving the training data set itself to enhance the performance of any given model on a specific AI task. This is accomplished through systematic or algorithmic changes to the data set, ensuring it contains high-quality, relevant, and diverse data during model training [5], [6]. By optimizing the input data, DCAI endeavors to elevate the model’s accuracy, robustness, and overall effectiveness in addressing the targeted AI task. DCAI encompasses several techniques, including outlier detection, error correction, data augmentation, feature engineering, establishing consensus labels, active learning, and curriculum learning. These methods aim to enhance data quality, representation, and informativeness to improve AI models’ performance and accuracy. The adoption of a DCAI approach extends beyond Big Tech companies, as it proves relevant and beneficial in industries with smaller or more challenging data sets due to regulatory or practical constraints (e.g., manufacturing and healthcare) [7]. For example, OpenAI has openly acknowledged that addressing errors in the training data and labels is a significant challenge with DALL-E and generative pretrained transformer (GPT)-3, underscoring the importance of data quality over the model itself. 1 The improvement of ChatGPT involved fine-tuning the model to prioritize data quality and minimizing harmful, untruthful, or biased output. Moreover, human rankings were incorporated to down-weight “bad data,” enhancing the reliability and accuracy of the AI system. Emphasizing data quality plays a pivotal role in refining AI models and mitigating potential issues related to biased or inaccurate outputs.

The importance of DCAI for the benefit of society is presented in Figure 1. Maintaining data quality and adapting data to meet evolving needs is essential for maximizing the benefits of DCAI [8]. The industry that adopts DCAI can benefit in the following ways:

By prioritizing high-quality data, DCAI systems can avoid “garbage-in, garbage-out” and yield more results that are dependable.

Recently, several research articles have proposed DCAI methods in different domains. These include data-centric defense (DCD) for mitigating model inversion attacks [10], crop disease identification in agriculture [11], unsupervised anomaly detection in industrial production [12], transformer-based language models (bidirectional encoder representations from transformers (BERT) and GPT-3 for automated occupation coding [13], enhancing deep neural network (DNN) model robustness [14], a comprehensive community library for biomedical natural language processing (NLP) data sets [15], automatic surgical phase estimation [16], rapid nanoparticle energy prediction [17], identifying incongruous regions in data [18], a multidomain benchmark for image classification [19], data-centric approaches to improve GAN training with less data [20], and a data-centric super-resolution (dcSR) approach for video quality enhancement [21]. A brief summary of recent DCAI works is presented in Table 1.

DCAI has the potential to revolutionize many industries and fields by enabling more efficient and effective decision-making based on insights extracted from data. There are countless potential new applications for DCAI, as the field continues to evolve rapidly [22]. A few potential applications for the well-being of society as shown in Figure 3 are briefly discussed below.

DCAI has the potential to revolutionize many industries and fields by enabling more efficient and effective decision-making based on insights extracted from data.

New ventures and innovations in DC technology have been rapidly growing in recent years. They are leveraging data to develop innovative products and services that are transforming industries and disrupting traditional business models [26], [27]. Some of these are focused on developing advanced analytics solutions to help businesses analyze large data sets and extract insights to make data-driven decisions. AI-based young enterprises are leveraging ML and other advanced techniques to develop intelligent applications that can automate tasks, make predictions, and provide insights. Internet-of-Things (IoT)-oriented startups are developing connected devices that collect and transmit data, allowing businesses to monitor and optimize their operations. Blockchain-focused startups are developing decentralized applications that use distributed ledgers to store and manage data securely. These applications have the potential to disrupt traditional industries such as finance, supply chain management, and healthcare. Cloud computing is another area where startups are developing infrastructure and platforms that enable businesses to store, process, and analyze data in the cloud [9], [28]. DC technology startups are transforming industries by providing new insights, increasing efficiency, and enabling new business models. As data continue to play an increasingly important role in business and society, these young enterprises are well-positioned to continue driving innovation and growth.

DCAI education and academic research involve the study of techniques, methods, and algorithms for analyzing, processing, and making predictions based on large data sets. It is a multidisciplinary field that draws upon knowledge from computer science, statistics, mathematics, and engineering [29], [30], [31]. In terms of education, DCAI programs typically cover topics such as data mining, ML, deep learning, NLP, and big data analytics. Students also learn about the various programming languages and tools commonly used in the field, such as Python, R, TensorFlow, PyTorch, and open source tools in DCAI [32], [33], [34] like Albumentations —Image augmentation, Amazon SageMaker Debugger, AutoAugment learning augmentation policies from data, Bootleg —self-supervision for named entity disambiguation at the tail, CleanLab —clean data set labels, DataPrep —prepare your data with a few lines of code, Deequ —unit tests for data, HoloClean: an ML system for data enrichment, Knodle —knowledge-supervised deep learning framework, Meerkat —handling data for ML models, mltrace —a python package to make ML models observable, nlpaug —augment NLP data sets, Picket: guarding against corrupted data in tabular data,Raha & Baran —configuration free error detection tools, skweak —weak supervision for NLP, etc.

Many universities around the world offer degree programs in DCAI, such as bachelor’s and master’s degrees in data science or AI, as well as PhD programs in related fields. There are also many online courses like MIT Introduction to DCAI and massive open online courses (MOOCs) available for students who want to learn about DCAI on their own or supplement their formal education. In terms of academic research, DCAI involves developing new algorithms, models, and techniques for analyzing and making predictions based on large data sets [35]. This research can be applied in a variety of fields, such as healthcare, finance, marketing, and social media. Some of the current research topics in DCAI include explainable AI, which seeks to make AI models more transparent and interpretable, and federated learning, which aims to train models in a decentralized manner while maintaining privacy and security. Other research areas include reinforcement learning, transfer learning, etc. [36].

Data are often referred to as the new oil, but refining them into meaningful insights can be a challenging process. Many companies are still facing difficulties when it comes to effectively managing and utilizing their data. Despite investing significant time and resources, their efforts often yield little results. The primary challenges are data volume, consistency, quality, and bias that industries face when adopting a DC approach to AI. These challenges can significantly impact the performance and reliability of an AI model, so it is important to address them properly [37], [38]. Organizations must ensure that they have a sufficient volume of high-quality data that is consistent and representative of the data that the model will process after deployment. Blindly collecting as much data as possible can be inefficient and costly. Therefore, it is important to establish the kind of data that is needed before acquiring more data. In addition, organizations must prioritize an effective system of data annotation to ensure consistency in the labeling of the data. An AI model trained on inconsistent data annotation can quickly become unreliable.

Biases can be introduced into AI systems when the training data used to develop the models are themselves biased. If the training data is not diverse or representative enough, or if it reflects historical biases and stereotypes, then the resulting AI models may perpetuate or even amplify these biases. In addition, biases can also be introduced during the selection of features or the design of the algorithm used to develop the AI model. Therefore, it is important to carefully select and prepare training data and algorithmic models to minimize the potential for biases in AI systems [39], [40]. For example, in 2018, Amazon had to abandon its AI-based hiring tool, which was designed to help automate and streamline the recruitment process. The tool was trained on resumes submitted to the company over a 10-year period, which were predominantly from men as the tech industry is male-dominated. As a result, the tool learned to prefer male candidates over female candidates, reflecting the biases present in the training data. Amazon’s tool was ultimately not used because it perpetuated the existing gender bias in the tech industry, rather than mitigating it. This case highlights the importance of ensuring that training data is diverse and representative and that AI models are tested for fairness to avoid perpetuating or amplifying biases [41], [42].

The solution to DCAI challenges requires a multifaceted approach. Data sources should be diversified to ensure the algorithm is exposed to a more representative sample of the population. Algorithms should be audited regularly to identify biases and provide opportunities for correction. Ethical considerations should be included in the development process [3]. Human oversight can help identify potential biases and ensure ethical decision-making. Monitoring and evaluating the algorithm’s performance can help identify and mitigate biases over time, ensuring reliable and ethical results.

The future of DCAI technology is exciting and holds enormous potential. With advances in AI algorithms, ML models, and data analytics tools, companies can gain deeper insights into their operations, customers, and markets. This allows them to make more informed decisions, optimize processes, and create more personalized and engaging customer experiences. One of the most significant trends in DCAI technology is the increased use of NLP and conversational AI. NLP enables machines to understand human language and respond accordingly, opening up new possibilities for customer service, virtual assistants, and chatbots. Conversational AI takes this a step further by allowing machines to carry on a conversation with a person in a more natural and human-like way [30], [33]. Another trend in DCAI technology is the integration of AI and IoT devices. This allows companies to collect and analyze data from a wide range of sources, including sensors, machines, and other devices. These data can then be used to optimize operations, improve product design, and create more personalized experiences for customers [34], [37]. Finally, there is the potential for DCAI technology to make significant strides in healthcare, financial services, and other industries. AI can help healthcare providers diagnose diseases more accurately, develop new treatments, and improve patient outcomes. In the financial sector, AI can help identify fraud, streamline lending processes, and improve risk management.

As technology and DCAI continue to evolve, it holds the promise of transforming our lives, work experiences, and interactions in unprecedented ways. Its current applications in healthcare, education, and finance demonstrate its ability to improve outcomes and enhance human experiences. However, despite the immense potential, there are significant concerns regarding the impact of DCAI, such as privacy infringement, job displacement, and biased decision-making. Therefore, it is crucial to develop and deploy DCAI based on ethical principles to ensure the equitable distribution of benefits across society.