The Biden-Harris Administration recently enacted a sweeping Executive Order to forge America’s path in responsible AI development, encouraging both innovation and risk mitigation. The Executive Order spells out a multi-faceted plan touching upon AI safety and security, privacy, equity, civil rights, and more, with profound implications for organizations that are already embedded in the AI landscape or are taking active steps into this burgeoning field. The order essentially provides a blueprint for AI’s future, requiring immediate action to conform to new standards and practices.
Two risks highlighted in the Order stand out to us, as mitigating them falls squarely into Private AI’s area of expertise: the requirement to preserve privacy and to reduce bias and algorithmic discrimination. These risks take center-stage in the Order. In fact, threats to privacy and discrimination are pervasive to almost all areas where the US administration sees the use of AI posing risks to individuals, from access to housing and jobs to criminal prosecution. The reason for the pervasiveness is that all AI systems have this in common: they are trained on vast amounts of data, and if the data set contains personal and biased information, the output will reflect this. Concerning generative AI, the same holds true with regard to the input users provide to the LLM.
This article addresses the new Executive Order and details how Private AI can lighten the task of these two important compliance aspects, freeing up resources to address the many other requirements the US administration has set out.
Safeguarding Privacy through Redaction of PII
The Executive Order highlights the need to protect Americans’ privacy by accelerating the development and use of privacy-enhancing technologies (PETs). The importance of privacy protection is emphasized with reference to the chilling effect that invasion of privacy has on exercising First Amendment rights. The thought here is presumably that when individuals know their data could be improperly collected or misused, they may refrain from exercising their rights to free speech or assembly, for fear of surveillance or retribution. The Order further mentions fraud in connection with privacy risks. Without spelling this out further, it can be inferred that one risk the US administration is attuned to is that LLMs are known to reproduce personal information contained in training data in production. Malicious actors may tease out this information and use it for identity theft and other fraudulent purposes.
Specializing in privacy-enhancing AI-driven redaction software, Private AI is an invaluable partner in ensuring compliance with this part of the Executive Order. Historical approaches to solving these privacy problems are no longer working in the age of Big Data. Private AI’s leading edge technology moves beyond simple pattern matching, using machine learning to allow organizations to identify and redact Personally Identifiable Information (PII), Protected Health Information (PHI), and Payment Card Information (PCI) in structured and unstructured data forming the basis of AI systems.
Context-aware machine learning models provide superior accuracy, allowing organizations to process data in a way that respects privacy: Removing personal information before it gets ingested by AI is the safest way to reduce risk as it ensures that no personal information gets revealed in production. It, of course, has the further benefit that compliance with applicable privacy laws is greatly facilitated since no personal data is disclosed in the process.
Advancing Equity by Mitigating Bias
The Executive Order also demands action to ensure that AI advances equity and civil rights, keeping algorithms from exacerbating discrimination. While an evaluation of a mortgage, job, or rental application can of course also be met with bias when a human is in charge, an AI system taking over can assess infinitely more applications in hardly any time, thereby significantly broadening the impact of existing bias incorporated in its training data.
Private AI’s capabilities offer a solution. By redacting identifiers that reveal gender, ethnic origin, sexual orientation, etc., the data set is rendered largely “neutral.” Redaction thus doesn’t just safeguard privacy; it can be a tool for social justice by reducing bias in data-driven systems. Private AI’s technology does not only work to sanitize training data. It can also filter out personal identifiers from user input before it is transferred to the LLM, ensuring that any sensitive data, including those often linked to biases, are not disclosed, and can’t form the basis for a biased output generated by the model.
Towards a More Trustworthy AI Ecosystem
Beyond mere compliance, Private AI contributes to the larger goal of creating an AI ecosystem that is safe, secure, and trustworthy. By providing a privacy layer that can work in sync with AI systems like ChatGPT, Private AI enables the seamless yet secure interaction that is indispensable for building public trust. Its capability of replacing personal data with synthetic data also ensures data usability for many purposes because the semantic integrity of the data is retained.
The Executive Order by the Biden-Harris Administration is a clarion call for responsible AI innovation, one that blends safety, privacy, and equity. Private AI serves as an excellent partner for organizations aiming to navigate this complex new landscape. We agree with Biden that the future of AI isn’t just about who builds the most advanced algorithms; it’s about doing so responsibly. With tools like Private AI, we don’t have to choose between innovation and respect for individual rights and the values of our society. We can, in fact, have our (secure, private, and equitable) cake and eat it too.