Responsible AI Through Crowdsourced Software Testing

Part 1 of this two-part blog serves as the foundation, highlighting the critical role that quality engineering (QE) must play in addressing the business risks associated with the accelerating adoption of artificial intelligence or AI-powered software.

Emeka Obianwu , VP, Alliances and Acquisitions

October 29th, 2024

Explore How Testlio Leverages AI

Additionally, this blog will introduce responsible AI quality gates as a mechanism for quality engineering teams to implement controls.

Part 2 will dive deeper into responsible AI quality gates and describe how the unique attributes of crowdsourced testing empower product teams to implement responsible AI quality gates.

The AI Golden Age

The golden age of AI is upon us, fueled by three areas of innovation and investment:

AI Technology. Advances in machine learning, natural language processing, and computer vision have enabled AI to tackle complex tasks with unprecedented accuracy and efficiency.
Democratization. Platform-as-a-Service (PaaS) providers like AWS Bedrock, and Microsoft Azure now offer software developers API access to LLMs and associated services.
Compute Infrastructure. Led by Nvidia, semiconductor industry innovations and investments are in place to power the computing infrastructure needed to support expansive usage of LLMs.

Emergent User Experience Business Risks

As the golden age of AI empowers machines to assume decision-making that historically required human involvement, businesses are being subject to a new set of risks concerning end-user experiences.

AI-generated content or interfaces can lead to inconsistent quality, causing user frustration and dissatisfaction. If the AI produces biased or culturally insensitive outputs, it can alienate segments of the user base and harm the company’s reputation.

Additionally, generative AI hallucinations, errors, or inability to apply contextual awareness can lead to flawed or dead-end user flows that negatively impact user interactions. These issues not only damage the brand’s credibility but also lead to potential legal liabilities and financial losses.

Enter Responsible AI

To combat these threats, technology providers and enterprises from across many industries are embracing a new movement: Responsible AI. Responsible AI seeks to define and advocate for a set of software quality, testing and governance practices that together mitigate these emergent risks associated with the modern era of AI.

“Responsible Artificial Intelligence (Responsible AI) is an approach to developing, assessing, and deploying AI systems in a safe, trustworthy, and ethical way.”

Quality Engineering & Responsible AI

Responsible AI calls upon product teams to expand their definition of software quality to ensure software releases meet functional requirements while adhering to ethical standards across different end-user scenarios. The following incremental responsible AI testing goes beyond functional accuracy by specifically tackling ethical, legal, and traceability considerations.

Bias and Fairness: Specific testing to identify and mitigate biases in AI outputs, ensuring that the system treats all users fairly and does not propagate inequalities.
Regulatory Compliance: Specific testing to ensure compliance with cross-border and ethical standards, with continuous updates to maintain ongoing regulatory alignment.
NSFW (Not Safe for Work): Specific testing to uncover risks of generative AI prompt interactions surfacing inappropriate, insensitive, or confidential content.

These three additional elements of software testing are now critical additions to quality engineering programs. They ensure the development of AI systems that not only fulfill their intended purposes but also adhere to responsible AI principles, fostering trust and positive societal impact.

Introducing the Four Software Testing Quality Gates of Responsible AI

The above offers additional dimensions for product teams to adopt. They are the “what”. Quality gates represent the “when”, as in when these additional areas of testing can and should be invoked across the software delivery lifecycle.

Shift left responsible AI quality gates:

Human-in-the-Loop Engineering. Leverage humans across the selection of foundation LLM models, custom model development, and prompt engineering to evaluate and ensure the accuracy, transparency, and interpretability of outputs.

Implement Prompt Guardrails. Embedded prompt guardrails should be built in, including automated checks that scan AI outputs for inappropriate or sensitive content and context-based notifications to help end users understand why AI generated a given response.

Shift right responsible AI quality gates:

Release Checkpoint. Expand the scope of pre-release functional testing to include the combination of bias, compliance, and NSFW testing. Empower engineering and QA teams to treat these categories of testing as part of the decision-making criteria before release.
In-Market Validation. Leverage humans to continuously conduct “in-the-wild” testing of production AI applications across a diversity of locations, languages, and device types. Speed time-to-resolution and unlock valuable end-user insights on AI feature adoption.

The AI golden age is placing new demands on product teams to assure a safe, ethical, and high-quality digital end-user experience. As organizations increasingly rely on AI to drive innovation and improve user experiences, quality engineering programs should incorporate responsible AI quality gates into their software testing and continuous delivery process.

Stay tuned for Part 2 of this series to learn how the unique attributes of crowdsourced testing give product teams a cost effective way to implement responsible AI quality gates.