Trustworthy AI and Crayon’s blueprint for ethical and explainable Large Language Models

Trustworthy AI is a holistic concept that incorporates aspects of fair AI, responsible AI, and safe AI, among other elements.

The rapid advancement and availability of AI technologies have meant companies have been keen to jump on the bandwagon to adopt and capitalize on them. However, it has been demonstrated that Generative AI can throw up ‘untrustworthy’ results which can be plain wrong and biased. Such outcomes can harm reputations, generate bad press and break the trust of investors and consumers.

Governments around the world are keen to set up guidelines for responsible AI. In the UK, for example, it has created an Office for AI which is responsible for encouraging safe and innovative Use of AI. Elsewhere there is a joint US-EU initiative to draft a set of voluntary rules for businesses, called the “AI Code of Conduct”, in line with their Joint Roadmap for Trustworthy AI and Risk Management.

In APAC, Singapore has been instrumental in spearheading responsible AI adoption. Also, Indonesia and Thailand have made strides in government AI readiness, echoing their recent implementation of national AI strategies. Japan’s new AI guidelines are a big step towards responsible AI use, stressing the importance of avoiding biased data and promoting transparency. Nations such as New Zealand and Australia, South Korea, India, and others are drawing up governance and ethics guidelines, some voluntary, and some are taking steps towards legislation. In December 2023, the EU published its AI Act: deal on comprehensive rules for trustworthy AI¹.  More recently in Australia, the Federal Government issued an open edict that it will set mandatory guard-rails in place for the use of AI in business and industry, which could include outright prohibition of ‘high-risk’ AIs.

Across the international community, governments and businesses are striving towards trustworthy AI at both macro and micro levels. Developers, technology service providers and systems integrators in the APAC region should anticipate accelerated regulatory review of development practices, adoption, implementation and use of AI in business and industry.

For many of our partners, the initial foray into AI may start with Generative AI and Large Language Models.  In our previous post, we explored the concepts of Fair AI and responsible use of AI.  In this update, we dive into Trustworthy AI, what it comprises and Crayon’s blueprint for ethical and explainable Large Language Models (LLMs).

Defining Trustworthy AI

Let’s look at how trustworthy AI relates to, and differs from Fair AI, Responsible AI and Safe AI.

Fair AI: This is a subset of trustworthy AI, specifically addressing the need for AI to be unbiased and to treat all users equitably. While fairness is a critical aspect, trustworthy AI also requires the system to be reliable, lawful, ethical, and robust.

Responsible AI: Responsible AI is a broader category that includes fairness but also extends to ethical obligations, societal impacts, and ensures that AI acts in people’s best interests. Trustworthy AI encompasses all these aspects of responsible AI but also implies a level of trust from users that the AI system will behave as expected in a wide range of scenarios and over time.

Safe AI: Safety is another essential facet of trustworthy AI, emphasizing the need for AI systems to operate without causing unintended harm. Trustworthy AI takes this further by not only ensuring that the AI doesn’t cause harm but is dependable, maintaining data privacy, and is secure against attacks or failures.

While fair, responsible, and safe AI are components that contribute to the trustworthiness of an AI system, trustworthy AI is the overarching goal, ensuring that AI systems are designed and deployed in a manner that earns the confidence of users and the general public and is worthy of that trust.

Trustworthy AI can be considered an umbrella term that implies an AI system is:

• Ethical: It respects human rights and democratic values.
• Robust: It is robust from a technical perspective (minimizing errors and unexpected behaviors), and a social perspective (it can be used under various conditions and by different user groups without causing societal harm).
• Lawful: It complies with all applicable laws and regulations.
• Socially beneficial: It contributes positively to society, enhancing individual and collective well-being.
• Transparent: Its decision-making processes can be understood by humans, and those impacted by AI decisions should be able to seek explanation and redress.
• Accountable: There should be mechanisms in place to hold the system and its operators accountable for its outcomes.

We want to ensure that AI systems are worthy of our trust and here at Crayon we underscore the importance of trustworthy AI by integrating ethical considerations into the development and deployment of LLM-based solutions in all AI projects we deliver.

From ‘black box’ to explainable AI

The concerns surrounding large language models (LLMs) as “black boxes” with opaque decision-making processes are well-founded, considering their intricate architectures. Nonetheless, advancements in the field lead to increased explainability and control over these models.

A case in point is Crayon’s approach. Crayon conducts research in AI explainability. By adopting emerging methods and tools, insights into LLM behaviour are enhanced. For instance, techniques analogous to those used in medical diagnosis tools for identifying influential image parts could be adapted to pinpoint text segments in LLMs affecting certain responses.

Crayon’s contribution is based on learnings from more than 300 successfully delivered AI projects.

The ability of LLMs like ChatGPT to self-explain can be exploited to comprehend the reasoning behind their responses. For instance, in sentiment analysis, LLMs might identify and enumerate the sentiment-charged words shaping their conclusions, providing a transparent layer, and facilitating comparison with traditional methods like LIME saliency maps.

Rigorous assessment of these self-explanations against established methods (e.g. SHAP, LIME) validates their efficacy. Metrics are employed to gauge the faithfulness and intelligibility of these explanations. And, acknowledging the dynamic nature of AI systems, including LLMs, Crayon commits to perpetual monitoring and enhancement of explainability features, revisiting current interpretability practices in response to LLM advancements.

We embrace user-centric design and believe it is imperative to ensure that explanations are accessible and comprehensible, catering to various expertise levels, from laypeople to AI specialists.

Further, Crayon integrates ethical considerations into the development and deployment of LLM-based solutions, ensuring transparency about AI explanations’ limitations and actively addressing biases and potential harms.

We have adopted an “Explainable by Design” strategy, employing specific paradigms according to the problem context. These include:

• Chain of thought: Encourages LLMs to decompose problems into sequential steps, enhancing transparency and interpretability.
• Tree of thought: Similar to the chain of thought but branches into multiple paths, enabling exploration of different decision-making avenues.
• Program of thought: Structures LLM responses in a programmatic format, using external functions for logical or mathematical tasks.
• Context Injection: Controlling the context provided to LLMs influences their outputs towards more explainable and reliable responses.
• Fine-tuning: Tailoring LLMs to specific datasets or tasks improves predictability and interpretability.
• Soft/ hard prompt engineering: The design of prompts significantly impacts LLM outputs, with soft prompts being more adaptable and hard prompts more structured.

Crayon’s strategic use of these paradigms maximizes LLMs’ inherent controllability and explainability, ensuring effectiveness, transparency, and trustworthiness. This proactive stance integrates explainability into LLM deployment’s core. Implementing these strategies not only enhances LLMs’ explainability and control but also yields considerable business benefits:

• Mitigating risks: Incorporating explainability in AI significantly lowers the risks of unforeseen outcomes and biases, crucial for avoiding reputational and legal issues.
• Future-proofing AI solutions: Adopting responsible and explainable AI practices prepares solutions for future technological and regulatory changes. This approach ensures their longevity and relevance, enabling continuous adaptation to emerging trends and societal expectations, and safeguarding long-term investments in AI technology.
• Building user trust: Comprehensible AI solutions foster user trust, essential for widespread adoption and effective utilization of AI across various domains.
• Ensuring appropriate deployment: A deeper understanding of AI solutions supports better decision-making, ensuring the right AI solution for the right problem.
• Competitive advantage and market differentiation: Adopting responsible and explainable AI practices distinguishes a company in the market, a factor increasingly important to customers and partners.
• Facilitating compliance and governance: Explainable AI systems simplify demonstrating compliance with evolving AI regulations, averting future legal complexities and penalties.

Integrating explainability and control in LLMs aligns with ethical AI practices and offers substantial business advantages, including risk mitigation, user trust enhancement, appropriate solution deployment, competitive edge, and regulatory compliance, are all crucial for the sustainable and responsible growth of AI-driven enterprises.

Incorporating explainability and control in large language models (LLMs) aligns with the global shift towards trustworthy AI, emphasizing transparency and fairness. This approach not only builds trust among stakeholders in sectors where clear, justifiable decision-making is crucial, like healthcare and finance, but also addresses the increasing social awareness of technology’s impact.

Explainable AI plays a key role in identifying and correcting biases, ensuring fair and equitable operations as a significant step towards trustworthy AI. This commitment reflects a deep understanding of AI’s societal context to address society’s concerns, aims to comply with what organizations and governments are legislating for, and demonstrates a dedication to responsible innovation and sustainable business practices, essential for maintaining operational legitimacy and social license.

Your choice of distributor makes a difference.

The appetite for AI is high and partners are on the frontline of educating customers about the right ways to think about the value of AI in their businesses.

Having a trusted AI sherpa to guide your own learning journey is essential and that’s where Crayon offers a unique difference to our partners.

Far beyond the cloud solutions distribution and licensing expertise we offer, Crayon partners benefit from nearly a decade worth of frontline, hands-on AI solution development. 

Crayon is an Azure OpenAI Global Strategic Partner with two Data and AI Centers of Excellence (COEs) including one based in Singapore.  We are the only services company with cloud distribution capabilities across the APAC channel to have over 300 applied AI projects in market, which includes more than 2,500 models running on a proprietary accelerated MLOps framework. 

Our work in this field is governed by the Crayon Responsible AI Guidelines (CRAIG). CRAIG grounds our services around sustainability, ethics, trust, robust engineering, and security. It establishes the in-depth policy mechanisms of our genuinely responsible AI organization. Our sales and delivery processes have tight integration of this policy. So too in our governance and support structures, and knowledge dissemination.

This is material for any partner that is using, integrating or implementing AI in their own business or for their customers, and wants absolute confidence that their distributor of choice understands the nuanced perspectives involved. 

If you are exploring the future of AI as a practice in your business and want to know more about the Crayon Responsible AI Guidelines (CRAIG), reach out to our Technology Advisory Group.

This post was originally published on Crayon.