As Bias Plagues AI Models, SAP Says Will Ensure GenAI Is Free From Stereotypes

To address the problems of biases in AI, the German software giant has noted that it is working on the stereotypes.

ChatGPT, the viral AI chatbot, changed the world a year ago, and now generative artificial intelligence (GenAI) is becoming a transformative technology that has the potential to revolutionise enterprises as well as industries. India has said it will make responble artificial intelligence (AI) and regulate it so that it is used only for "contructive purposes" and used safely in the country. However, a big challenge for the industry world over is the inherent bias that AI tools have embedded in their algorithms and software, which continues to plague the technology.

German software giant SAP has said it is working to ensure generative AI is free from any kind of stereotypes. It claimed to have AI committees of experts long before the hype surrounding GenAI surfaced.

SAP is in favour of a risk-based approach to AI regulation. It noted that AI is a game changer for efficiency and climate action and India can lead in this area.

SAP Working With Govt On New Approach To AI 

"It is good to see that the Indian government is working to future-proof AI policies. In terms of AI regulation overall, we continue to believe governments around the world need to strike the right balance between innovation and regulation. We look forward to reviewing the recommendations and continuing to work with the government and industry bodies as we move forward with the new approach to AI," a SAP representative told ABP Live.

The company further stressed on the need to have a balanced approach that encourages innovation while regulating high-risk AI apps that have the potential to harm health, safety and fundamental rights.

It may be recalled that in June this year, NASSCOM, the association representing the IT and business process management sector, issued a comprehensive set of guidelines for responsible AI, specifically GenAI. These guidelines seek to define normative responsibilities for researchers, developers as well as users engaged with GenAI models and apps, fostering a responsible approach to the adoption of this technology. To promote further transparency and accountability, public disclosures of data and algorithm sources used for modelling will be mandatory. 

Synthetic Data Better Than Real Data To Prevent Biases

Betting big on GenAI, SAP introduced a range of generative AI solutions at its recently concluded TechED flagship event. SAP introduced a number of GenAI offerings, including vector engine on SAP HANA Cloud, SAP Build Code solutions and a generative AI hub in SAP AI Core. To address the problems of biases in AI, the German software giant has noted that it is working on the stereotypes and making sure it is free from biases.

"From a research perspective at SAP, we are experimenting on the use of synthetic data to develop training datasets that are not biased so we can tune in the datasets in a way that overcomes inherent data bias," Max Wessel, Chief Learning Officer, SAP, told ABP Live, during an interaction.

The requirement to expose an AI model to extensive data, during training, may include situations where the necessary data may be non-existent or lacks comprehensiveness. This is where synthetic data becomes a crucial component in the training process of AI models. Moreover, to train these GenAI technologies, datasets from the Internet undergo processing, with human intervention employed to mitigate harmful biases.

'Will Overcome The Inherent Biases Of AI Models'

According to a recent study published in the journal Organizational Dynamics, researchers stated that AI-generated content has the potential to uphold and sustain detrimental gender biases and machine bias refers to the bias ingrained in algorithms and software. However, Wessel is confident that AI models will get rid of the inherent biases in future.

"With innovation there and with some of the improvements, that we'll see in the underlying technology, I think it will overcome the inherent biases of these AI models, and the training datasets over time. But it also requires you there to be conscious of where the data is not a match to reality," Wessel noted.

In the world of AI, synthetic data emerges as a new area, which is said to alleviate the challenges associated with manual data acquisition, annotation and cleaning. This novel approach addresses the issue of obtaining data that might be unattainable through conventional means. Synthetic data generation is capable of delivering equivalent results to real-world data but in a significantly reduced timeframe and without compromising privacy.

Sixty per cent of all data used in the development of AI will be synthetic rather than real by 2024, says a study by global market intelligence firm Gartner.

Breaking down sythetic data for the layman, SAP's Wessel explained: "Synthetic data is the generated data that is realistic to the properties of original data but right when it was created, it was created in a way that matches the types of relationships that exist."

In fact, synthetic data is a remarkably straightforward concept -- an idea that appears almost too good to be true. Synthetic data essentially allows practitioners to effortlessly generate the required data digitally, on demand, in any desired volume, and customised to their exact specifications.