The growing attention for generative AI tools such as ChatGPT has captured our collective imagination. The versatility of this technology has enabled professionals in every sector to envision how AI can alter human interactions with the digital world.
From content creation to more complex tasks like fraud prevention, the capabilities for generative AI seem endless. However, it won’t have a revolutionary impact – for example to the same effect that AI has had on personalised video streaming over traditional TV – unless it is populated with the data needed to truly bring it to life.
Without up-to-date high-quality data, generative AI will not sustain its status as a game changer that can reinvent the digital experience for businesses and consumers.
Generative AI’s success requires fixing our data ecology
Providing AI with the data it needs will require us to rethink how our data is shared to meet the compliance and copyrighting needs of data owners. In general, data ecology has historically been light in touch.
However, the arrival of technology like GPT-4 is motivating data owners to incorporate better data sourcing and governance procedures, such as data tagging. This is no bad thing and has led to the emergence of innovative data governance models, such as data cooperatives, which promote the idea that organisations and individuals can collectively control and benefit from their data.
In their simplest form these are just repositories of training data where ownership rights and attributes such as collection date and quality assurance are clearly marked. An example of this is the MIT Media Lab’s Data Provenance Initiative.
More proactive are data exchanges like Gaia-X, which aim to provide data owners (licensors) the ability to provide revokable and specific rights to organisations to use their data for the training of models.
Generative AI must deliver value without loss of IP or competitive advantage
Feeding generative AI models the high-quality data they need comes with a risk to customer privacy though, and AI’s future success hinges on trust and our collective ability to create systems that inspire faith from consumers.
Avoiding these privacy issues will require trusted networks (e.g., legal and data architectures like the Mastercard network), federated AI, and TEEs (Trusted Execution Environments). With the goal of enabling data to be used securely and privately, so that AI models can offer contextually relevant answers, but without revealing private aspects of the data owning community or its members. Privacy-preserving methods like these allow community members to see real value without data leakage risk.
This approach provides a protected roadmap to update data within AI models, while effectively safeguarding sensitive communities and personal information, and providing contextual, flexible and safe access through web searches. However, maintaining this will need unique privacy, security, and custodianship solutions that safely bring together data from multiple parties and stakeholders.
How to ensure the safety and performance of AI
Both EU and US government guidelines stress the need to protect consumers without hampering AI innovation. However, elements of the proposed EU regulation on AI maybe unachievable, for example the requirements on high-risk AI systems: “training, validation and testing data sets shall be relevant, representative, free of errors and complete”.
Virtually no dataset is likely to be complete and some degree of error is always present. Imposing these static rules irrespective of context misunderstands how AI is developed and not only fails to protect consumers but also deters desirable innovation.
Instead, we would advocate for a more flexible legal frameworks that introduce incentives for AI creators to release safe products. Today we already have well-established strict product liability principles applied to most technologies.
Why not make manufacturers liable for defective or falsely advertised AI systems? Imagine if AI manufacturers were held accountable for the harm caused to an injured party based on evidence. Both citizens and business would naturally be incentived to adhear to the framework, without needing to over complicate the processes.
To ensure fairness, the process would need to outline clearly when a defective AI system caused harm by creating a baseline (a human or another AI) to compare with. This is where auditing AI behavior becomes important.
Auditing for certification is already part of the proposed EU AI regulations, and in Singapore they have released AI Verify, an AI auditing and testing framework for companies based on internationally accepted AI governance principles. AI Verify does not define ethical standards, rather enables developers and owners to demonstrate their claims about the performance of their AI systems.
While still in a pilot phase, Singapore’s efforts represent further development of the debate on AI governance and regulation, and respond to the growing demand for trustworthy AI systems and interoperability in global AI regulatory frameworks.
As AI becomes ever more ingrained in our lives and regulators seeking to protect consumers, while promoting innovation, we must understand that there is no one-size-fits-all global solution. As seen with electricity, cars, bridges, highways…. rigid regulation of new technologies has always been a poor general approach to making technology safe.
Void of any flexibility, fixed regulation fails to anticipate new applications and capabilities, restricts only local developers who then can’t compete in a global market. Instead implementing a more flexible auditing regulatory framework will both narrow the scope of records for error assessments, making it easier to appeal an automated decision.
Finally, this is a learning process, we shouldn’t expect immediate globally applicable solutions but instead a focus on where the benefit of AI can be clearly demonstrated, and maintain a level of transparency that opens the door to for people to engage positively.