Artificial intelligence (AI) is now pretty much ubiquitous. With ChatGPT having beaten the Turing Test, AI is being increasingly leveraged for both business and consumer applications across finance, healthcare, education, entertainment, e-commerce, logistics and government. It’s already used in self-driving cars and could one day find itself assisting pilots on aircraft flight decks, or in autonomous aircraft. Moreover, AI built into combat drones is already a reality in Ukraine, and its adoption for many other military uses is surely not far off. How long, one wonders, before it’s deployed for nuclear missile defense systems, as depicted in the 1983 movie Wargames?
Given its widespread use, and potential catastrophic scenarios for its misuse, much attention is being directed to governing and managing AI, to ensure that it operates in an ethical, responsible, and trustworthy manner.
Ethical and Responsible AI … Built on Trust
Ethical AI and Responsible AI are strategic imperatives that refer to the principles and procedures that set a vision and guide the development and use of AI to ensure it is beneficial to society and that it respects human values, such as fairness, accountability and being mindful of privacy and security.Trustworthy AI has a more focused and tactical mission and generally has more immediate and direct impact on companies implementing AI and their customers, who increasingly want to know more about their service providers, and especially how they are making use of their personal information. Trustworthy AI is generally considered to be a prerequisite for and a building block of ethical and responsible AI.
Drilling down, for AI models to be trusted, they need to be accurate, reliable, and transparent in their decision making. They not only need to make the right decisions based on data inputs, but they need to demonstrate that the decisions are indeed correct. Many facets of AI model design and training need to be considered to create trustworthy algorithms, including legal, organizational, procedural and technology aspects.
By their very nature. Decentralized AI (DeAI) architectures are generally positioned as open source and transparent – and so more trusted. Centralized models from IT heavyweights are often presented as “black box” offerings, which are marketed on their accuracy, flexibility, and ease of use credentials.
Such services generally draw on investments and exclusive licensing arrangements – such as Microsoft’s relationship with OpenAI, and Amazon’s deal with Anthropic – and benefit from cloud services that feature significant GPU server power. But the downside of these AI black boxes is that users cannot inspect models or track data flows, and so cannot tell whether bias or tampering of results has taken place. For some users, the perception that models are vulnerable might cause them to look at alternatives.
It's All About Data
But given that the effectiveness of AI models is dependent on the data that is used to teach them, and the data presented to them at inference time, then the data management and integrity aspects of AI is a hot topic. Another hot topic is privacy, which might seem impossible to achieve given AI’s need for accurate data from sources that don’t want to provide it.
Decentralized AI architectures, and decentralization technologies (DecentraTech) – which are built on blockchain platforms, cryptographic primitives, and token incentive models – can be leveraged to address several trust, data, and privacy related aspects of AI. Considerations for creating trustworthy AI include:
Data Availability and Privacy – adopting a decentralized and federated approach to AI datasets allows model code to work only on subsets of data that might be stored behind a corporate firewall. Just the outputs of the AI model are exposed to the outside world, and not the raw input data. These outputs can be aggregated so that an aggregated output can be produced that encompasses all the datasets as inputs, including those that remain private.
Beyond federation, a set of cryptographic techniques, known as Privacy Enhancing Technologies (PETs) can allow data elements to be included in AI model processing without exposing the values of the data. PETs include zero-knowledge proofs, multi-party computation and homomorphic encryption. While PETs are considered cutting edge and are not yet widely adopted in production environments, one can expect to see their increasing real-life rollout in the next year.
Data Monetization – it is common for incentive models – often leveraging tokenization approaches – to be leveraged to attach a monetary value to private datasets, and so make it more likely that their owners will want to share them in a privacy preserving way. Data marketplaces are emerging that make it easier for model providers to discover and integrate diverse datasets to power their models.
Traceability and Provenance – by cryptographically signing individual data elements and models and including them in an audit trail allows a robust record of provenance to be created. The audit trail – using (virtually) tamper-proof blockchain technology – would include not both data inputs but also outputs, and which models processed them.
Subsequent analysis of this provenance record stream allows both data and models to be traced from their source until the AI outputs are presented to users.
Improving Trust for Model Providers and Users
Leveraging a combination of decentralization technologies and techniques as outlined above, providers and users of AI models can benefit and take comfort from increased trust profiles. To summarize:
Because AI models can learn from datasets that are otherwise private, the accuracy of models tends to be improved, which in turn feeds into better trust outcomes.
Provenance of data and models also underpins transparency of AI models, including determining how data inputs have been processed by models. Knowledge of provenance processes underpins increased trust.
Determination of whether data inputs might be subject to bias and how this has been addressed/neutralized also is a key input to determination of trust in models.
Building trustworthy AI models requires providers to understand all the data sourcing and processing concepts and issues outlined in this blog. Leveraging available decentralization and DeAI platforms and tools can also accelerate their creation – and some of those will be covered in Part 2.
Dive deeper into Decentralization Technologies and Trustworthy AI in Person
I will be discussing the role of decentralization technologies to underpin trustworthy AI at the AI STAC event in New York City on December 10, 2024, as part of a panel titled: "Who says?: Tracing AI outputs to their source.”
For more information on this event and the panel, visit https://stacresearch.com/aistacNYC.
תגובות