Generative AI: What’s All the Noise?

Insight provided by Tintra

Recently, a media storm has erupted with concerns about the potential harmful consequences of introducing generative AI to the public. This commotion has highlighted the need for awareness of not only the potential for innovation in the development of new technologies, but also for critical insight into the multitude of risks associated with their introduction.

In recent years, the development of generative AI has begun to redefine the possibilities of artificial intelligence, namely within human creativity itself. What was once confined to the research laboratories of academia and industry is now available to the public and spreading on a global scale at an unprecedented rate. Able to create new data or content not explicitly defined by a human expert, generative AI uses the patterns and relationships it has learned from the training data to generate new data similar to this source material. Since it is not limited to a fixed set of predefined rules, generative AI can also be used to generate data that is more diverse and creative than the original training data. The emergence of generative chatbots such as Chat-GPT is a good example of this rapid transformation.

With over 100 million users in less than two months, Chat-GPT has surpassed the growth rates of other popular platforms such as TikTok, which took nine months to reach the same milestone, and Facebook, which took a staggering 4.5 years. This impressive growth highlights the enormous potential of generative AI and its ability to rapidly gain widespread use. Although generative AI has the potential to revolutionize many fields, including art, music, and literature, its recent widespread use has begun to highlight a number of important issues that need to be addressed to ensure ethical, responsible, and effective adoption.

Let’s explore:

Algorithmic bias

One of the most pressing concerns is the introduction of harmful bias into an AI system. Algorithmic bias refers to the tendency of machine learning algorithms to produce results that are systematically biased or discriminatory against certain groups of people. For example, an algorithm used to screen job applicants may unfairly discriminate against people with certain demographic characteristics, such as race or gender. This can lead to unfair treatment and maintain existing social inequalities. From algorithm design to the composition of the data set used for training, the presence of cultural biases can systematically encode cultural values ​​and privilege certain groups over others. This has led to the introduction of prejudiced discrimination in AI systems, creating prejudice against groups on a large scale. For example, the Gender Shades project by Buolamwini & Gebru, 2018 highlighted how facial recognition systems systematically performed worse on darker skin tones and least well on darker female faces. When certain groups are less well represented in the training data used for system development, there are a number of effects for downstream applications.

Quality and diversity of data

As the dominance of data-driven systems has increased, it is important to remember that data is not neutral. Rather, its composition, how it is collected, elected, and whose votes it includes, affects both system performance and how equity is modeled in those systems. These models underpin the many technologies that permeate our everyday lives and can not only reproduce but also reinforce bias and discrimination. The use of large data sets is often necessary to train models, which means that data sets are often obtained to prioritize sample size. There is a need for greater emphasis on the quality of training data and the values ​​that are the basis for data collection and curation strategies. Similar problems have been observed in connection with health services. For example, many health datasets do not adequately represent different demographic groups.

In publicly available health datasets, incomplete demographic reporting was observed and they were disproportionately collected from a small number of high-income countries. For skin cancer datasets, only 2% of datasets reported clinically relevant key demographic information, such as ethnicity and skin tone. A starting point for solving this problem is to specifically curate diversity in the training data. This can be achieved by gathering data from a range of different perspectives and viewpoints to reduce the risk of generating output that is overly influenced by a particular bias.

Plagiarism and identity fraud

Another major concern with generative AI is its potential use for plagiarism, and more worryingly, identity fraud. The use of generative AI in this way has become known as a deep fake. Deep spoofing can be used to manipulate video footage to create a false identity that is indistinguishable from the real person. This can lead to serious consequences in a number of contexts, such as political propaganda, online harassment or even financial fraud. In some cases, deep forgeries can be used to create false evidence in a court case or to damage someone’s reputation by falsely attributing statements or actions to them. Several possible methods have begun to be adopted to combat identity fraud through the use of deep fakes. One method currently being adopted is the investment in advanced deep fake detection tools. These tools can analyze ID documents and videos submitted by customers and determine whether they are authentic or not. However, approaching this problem through deep fake detection alone is problematic in that it creates an arms race between the fraudulent individuals who develop deep fake technologies and the institutions that try to combat them.

Conclusion

As generative AI continues to evolve, it is critical that we address potential concerns to ensure ethical, responsible and effective use. This new technology offers enormous potential for many industries, but it is only through careful ethical and cultural considerations that we can fully realize its benefits and avoid the potential for unintended harm.

To ensure the development and use of generative AI – which is fair for all – ethical frameworks must be followed throughout the entire research and development process, from designing the technology with the involvement of relevant stakeholders, to the selection, labeling and structuring of datasets. In addition, constant monitoring is required to identify and address any inherent biases during the design process, as well as the output of the system. By adopting these measures, we can help ensure that the promises of generative artificial intelligence are realized in a way that promotes, rather than risks harming, human well-being and social justice.