Building Inclusive AI: Addressing Hidden Biases

This article discusses two less visible biases. We’ll explore how to avoid these biases. The solution involves embracing inclusive datasets and hyper-localized solutions.

Uncover the Overlooked Biases

Selection Bias

This bias occurs when diversity lacks representation for all users. It stems from overlooking outliers and underrepresented groups. Mainstream large language models (LLMs) today train predominantly on English text. This limits AI advancements to English-speaking environments.

Even in English-speaking regions, problems persist. An AI system trained on US English data may fail to recognize slang words in the UK, Australia, or Singapore.

Popularity Bias

This bias occurs when AI algorithms build on popular choices only. Popular choices from the majority become the only visible choices. This approach excludes people who make less mainstream choices.

For instance, an e-commerce platform’s AI algorithms might recommend bestsellers only. This neglects niche interests and specialized needs.

Crafting Responsible and Inclusive AI

We need broader sourcing and sampling to capture true diversity. We must embrace localized and personalized solutions that avoid majority defaults. At WIZ.AI, we commit to making AI inclusive for all. This forms an indispensable part of practicing responsible AI.

Hyper-localized Solutions for Diverse Consumers

Enterprises and their consumers need easy access to our generative AI-powered customer engagement solutions. This requires multilingual capabilities. We embed hyper-localization in our products’ DNA, including our flagship Talkbot.

Smart bots understand multiple languages and dialects. These include Bahasa Indonesia, Thai, Tagalog, and Singlish. This ensures our Talkbots have barrier-free conversations with diverse consumers in the region.

LLMs-powered Hyper-localization

Our hyper-localization effort includes training localized LLMs. We launched an LLM for Bahasa Indonesia. Our LLM for Thai undergoes training currently.

During LLM training, we feed the model diverse real-world customer conversation calls. These reflect local culture, dialect, and language contexts in actual applications. This minimizes selection bias.

Our R&D team revealed an interesting finding. LLMs’ learning ability makes training easier and faster than traditional AI models. Being inclusive remains very difficult. There can always be underrepresented groups and missing cases.

One enterprise client offers hundreds of cigarette brands. Local Indonesians have multiple ways to refer to a single brand. Traditional AI training required feeding all naming varieties during training. The system needed this to understand sayings in actual conversation.

With LLMs, we can inform the system that “new nickname = X brand” in a prompt. The LLM picks up this new knowledge instantly. No need to retrain from scratch.

Capturing Cultural Sensitivity

We consider local cultural norms beyond languages. When we first launched our Talkbot in Indonesia for debt collection, we used a straightforward tone. Our bot would tell Indonesian consumers about consequences directly. People in Indonesia don’t consider this approach rude.

We launched the product in Thailand later. We strategically revised the conversation tone to be more polite and friendly. When Thai users might be late for repayment, our bot reminds them gently about potential negative consequences. Thai people value courtesy highly and avoid confrontation in daily interactions.

Never Let Any User Feel Excluded

When products feel “off” to us, this could indicate overlooked perspectives in design. At WIZ.AI, we try our best to avoid these moments. We design our products to cater for diverse users.

Inclusion of Vulnerable Groups

Our Talkbot requires no human training on the back-end. Customers aren’t forced to adopt digital self-service solutions on the front-end. Talkbots work well on various telecommunication mediums. These include traditional landline telephones, analogue phones, and smartphones. This ensures effective coverage for all customer segments.

One enterprise client in the Philippines has hundreds of mom-and-pop stores as distributors. These store owners, mostly older people, usually struggle with online booking systems on smartphones. Some don’t even own smartphones.

Previously, the enterprise client sent human agents to each store weekly. Agents collected orders in person. Now with smart Talkbots, the client makes automated calls to shop owners over telephones. This saves vast human effort weekly.

Inclusive Knowledge Base

WIZ.AI’s experts used to spend days studying real-life customer service recordings. They would brainstorm possible conversation scenarios for particular use cases. Consider users returning products as an example. How many different expression ways might they have?

LLMs empower us to generate hundreds of possible expressions and conversation scenarios in seconds. This enables a truly inclusive knowledge base!

The WIZ.AI team conducts extensive user testing before deployment. We gather feedback to create solutions that work effectively across customer segments. The goal is providing memorable and positive experiences for our enterprise clients.

Building an Inclusive Culture

We believe products reflect who builds them. Broad representation enables holistic perspectives. WIZ.AI hires local talents across the ASEAN region. Our technology, product management, and customer experience teams include people from Singapore, Indonesia, Philippines, Thailand, and Malaysia.

Towards a Future with Democratized AI Access

At WIZ.AI, we innovate continuously to democratize AI access. We make AI solutions inclusive for all. As we enter the AGI era, we remain committed to developing AGI solutions for enterprise users worldwide. We expand AI access for our customers. Only through embracing diversity can AI enable a fair and equitable future.

Our LLM for Bahasa Indonesia will open for tests soon. We consider making it open source in the future. Further developments will focus on feeding more diverse data types. These include Indonesia’s many local dialects and everyday slang.