Trustworthiness in Artificial Intelligence – Exploring 5 Interesting Ideas

Reading Time: 8 Minutes

I came across this topic of Trustworthiness in Artificial Intelligence during my studies and found it was very interesting. In this era of modern computing and tech, everyone has access to powerful AI systems even through a basic device like mobile phone, instead of need for huge computers with large processors. But this brings us to ask an important question – can we trust the responses of AI?

In a high level, it is generally known that these AI systems comes with disclaimers or terms which emphasize that AI responses may be inaccurate/incorrect – and we need to verify certain outputs. But when we go deeper, we can see AI hallucinate and output wrong responses with high confidence and even base its reasonings based on incorrect judgements or primary knowledge.

1. The Idea of Trust

In the grand scheme of things, from social structures such as hospitals to corporate entities, they operate primarily on one basic aspect that is trust, and that is why trust is one of the most expensive things over any form of materialism. Inspite of AI booming from the past 10 years, algorithms played an important role in mankind for a longtime, algorithms made our lives easier and more efficient thereby driving innovation. But an important distinction between algorithms and modern AI is the “reasoning” or “how the system came to a conclusion” part.

Previously, the algorithms were purely mathematically predictable or condition based. Contrast to current age algorithms mostly seen in AI, are mostly Blackbox and trained on data that have nuances in terms of viewpoints, have biases and reasoning over distinguishable statistical distribution. This makes modern AI difficult to be trusted blindly without analyzing its sources and reasoning steps.

Broadly, trust can be established towards an entity when right reasoning is exhibited, or right principles are followed, and of course elements of ethical values and morality are also involved.

When it comes to an algorithm, it is reasonable to trust it when the reasoning or steps performed to reach a solution is mathematically explainable or the conditions in the algorithm can be explained.
Complexity comes into picture when non-linearity is leveraged by AI algorithms to solve complex problems, though this is more powerful than tradition algorithms and led to larger innovations, the trustworthiness of it comes into question as AI is susceptible to making mistakes.

To simplify further, trust can be based on the capability, transparency and reliability of a system, all of which an algorithm or an AI can satisfy.

Capability is more of a technical examination as it relates to how well a model can perform to unseen situations and solve existing problems with optimality. But transparency and reliability in AI are the grey areas that can’t be quantitatively pinpointed, of course statistical analysis can be done on the historical reliability, and its transparency is terms of sharing all facts and reasoning. But in modern AI systems, this cannot be completely established due to evolving nature of AI with growing data and model parameters.

2. Data Protection and Security

When it comes to trust in AI, the topic of data protection and data security cannot be left unnoticed. Since AI exhibits Blackbox behavior with the increasing complexity in model parameters and modern architectures, securing/processing data before training the model or giving it context becoming of at most importance. Processing data – such as masking sensitive information or replacing those data with synthetic data is crucial, as certain guardrails or restrictions can be bypassed by hackers who exploit AI. Examples of such attacks include membership inference and model inversion attacks which can leak sensitive data which was originally intended to only help the model learn patterns.

Therefore, to enable data protection, data governance or administration within the organization should be established when dealing with sensitive/private data. In addition, guidelines and guardrails should be set for training AI systems. When it comes to synthetic data, there is always a question of tradeoff between maximizing utility by training with more sensitive data or minimizing the risk of sensitive data usage by reducing utility.Various solutions exist for handling this situation, some of the frameworks that help validate synthetic data accuracy are Differential Privacy and Fidelity and Privacy Assessment through mathematical computations.

3. The Technical Aspect of Establishing Trust

From a technical standpoint, transparency of AI systems can be ensured by registering the system under a globally registered entity that verifies it based on a set of rules/questions and classify it based on risk – which can help AI producing organizations to redefine their models in terms of risk factors.

Another idea would be to make the AI system go through a certification process that can report aspects like AI’s technical features, degree of autonomy and self-learning capabilities. This certification can help consumers choose AI effectively that is verified. I also believe we should see AI more like an organization rather than an individual with its own thoughts when it comes to analyzing its responses for biases, this is because it is being trained from data all over the world which has different viewpoints.

An AI system can be elected based on multiple features or frameworks with which it generates outputs, like how ML models are trained on multiple columns to predict a target, like a decision tree. This approach towards AI can help improve its trustworthiness – as it becomes easy to pinpoint its reasoning and decisions – thereby solving the issue of transparency and reliability.

4. Autonomy in AI

Autonomy is the right to self-governance. AI or modern algorithms are not anymore simple conditional statements/automations rather entities that has feeds of multiple world views and constant updates, so in order base its transparency and reliability, I think we need to establish governing layers that helps the model behave or output widely accepted norms and also considers the nuances and other viewpoints thereby becoming diplomatic.

On the same light, autonomy could be the one of the major impediments to current and near-future AI’s being trustworthy. Because, apart from governing the layers of an autonomous AI, it is important to ensure that it is liable to its actions that impacts the lives of human beings.

Here autonomy could add a layer of complexity. One of the important discussions that comes into picture is the legal personhood concepts where elements of compensation to victims, efficient distribution of social costs and risks among all AI’s stakeholders and legal simplification and transparency through mandatory registration and certification are emphasized.

These aspects are theoretically feasible, but it can be a challenge to be implemented practically. The reason being, establishment of a global governing body that is voted based on all the users of AI or internet is a huge task. And even if it is established, it is quite a task setting compensation rules based on severity of AI’s output, as it can subjectively differ, and distribution of social costs and risk amongst stakeholders might further be a complex step to achieve.

5. Human Interaction

In summary, given the weight of all the nuances for trustworthiness in AI discussed above, a human level interference is important given the evolving nature in AI, as these advancements also comes with a cost of uncertainty in its certain actions. The human in loop aspect need not be necessarily in checking whether the AI responses are not exposing any sensitive data or whether it is factually correct – as these tasks can be automated with algorithms or another AI, rather the human element should be enforced to validate the moral and ethical aspects of the model.

To be more precise, humans can be involved in evaluating the diplomacy of the model and how reliable a model is, by passing it through various subjective based tests which humans can effectively design.

Another place where humans can play a role in ensuring trustworthiness is by governing responses when sensitive tags are applied to a response before giving it an end user – this solution may not look practical, but a fundamental concept in AI – reinforcement learning, is also not possible in real time. So, this approach would be a continual learning process for an AI where humans employed for tuning certain kinds of response can help improve the trustworthiness of an AI model over the time.

For next generation or near future AI’s, the trust factor becomes more complicated, as innovations are made that has more model parameters and complex architectures. In addition, since new models can be adapted by fine-tuning with minimal compute, more AI models are available.

Regulating everything through policies can be a real challenge as it can sometimes hinder growth and innovation, but at the same time, frameworks and governing bodies must regulate certain aspects discussed above to ensure trustworthiness of AI.

References –

That’s a wrap, thank you for taking the effort to read this post ! Feel free to drop your thoughts in the comment section below. Subscribe to sapiencespace and enable notifications to get regular insights.

_{Cover picture and title image credits – unsplash content creators}

Click here to explore through similar insights.

What’s your Reaction?