Annotation and opportunity: a blueprint for inclusive AI

Sama annotates data for AI algorithms. Huawei's Executive Editor-in-Chief Gavin Allen spoke with the company’s CEO about bringing the power of mobile AI to underdeveloped markets.

Gavin Allen: Your company does data annotation. What does that mean?

Wendy Gonzalez: Data annotation is the process of labeling or tagging unstructured data (e.g., emails, photos, voice recordings) so that machine learning or Gen AI models can learn from it.

For example, before an autonomous car can drive itself, it needs to detect lanes and traffic signs. Annotation precisely identifies those lanes and signs. AI models need something reliable to learn from; otherwise, it’s “garbage in, garbage out.” Identifying the datasets that should be used to train the AI is a critical part of developing a high-performing model.

Gavin: What factors do you have to address when annotating data for mobile AI systems?

Wendy: We tend to see a common set of applications and a lot of multi-modal images, video, and text – everything from product recommendations, to virtual try-ons for glasses or clothing, to image classification. You need the right datasets, trained on the required parameters, and a quality evaluation framework to measure that this is an accurately structured piece of data. Those principles apply whether you're building a large enterprise model or a mobile application.

Gavin: Which industries are being transformed most quickly by mobile AI?

Wendy: Previously, the work mostly involved object and text recognition: let me take a picture of this wine bottle and find out its rating, for instance. Now, we’re seeing much more happen in AR/VR and wearables, with the technology getting very sophisticated. When you're wearing these intelligent sets of glasses, not only do you need pixel-point accuracy to detect objects, but the technology also has to track your eye movements. As with watches and wearables for fitness and tracking, we're seeing a lot of really interesting developments.

Gavin: And Sama assists with the data annotation for the models?

Wendy: Yes, we validate and correct those outputs, and we might support new edge cases for models already in production. There are all sorts of fascinating use cases where mobile AI development is creating benefits. Farmers typically have to bring all of their grain and products to market, which is a huge, laborious effort: weighing their wheat, understanding sales value, and feeding the commodities pricing.

But now, farmers can take their mobile phones, photograph their stacks of wheat, and have computer vision automatically calculate it all, without having to haul it all to market.

Similarly, during Covid, people in underserved parts of Kenya were using their mobile phones to automatically generate text and upload images to the local market of the snacks and food they wanted to sell. You could safely have goods delivered to somebody's home, retaining the ability to earn an income – which was life-changing in a lockdown. And it was powered by global AI apps.

Gavin: How do you ensure low-income and low-connectivity communities benefit from mobile AI?

Wendy: Sama’s mission is to bring people into the digital economy. Our annotation teams come from underserved communities in East Africa, and we provide living wages to foster financial independence and skill development.

We also support connectivity—our scale helps influence infrastructure, but government involvement is key. Kenya’s ICT program, for example, expands digital hubs into rural areas. Mobile apps are transformative—India’s mobile infrastructure shows how powerful, edge-based applications can drive progress.

Gavin: So, you’re creating jobs, connectivity, and community mobility?

Wendy: Exactly—we’re building an ecosystem. Talent is distributed equally, but opportunity is not. It's just about opening the door. Sama – which means equal in Sanskrit – opens that door. These talented individuals walk through it, accessing positions that would normally be very difficult to get. Unemployment rates can be upwards of 70% in the areas we serve. These opportunities produce income that goes into the local economy, and drives connectivity and infrastructure.

Many of our team members advance into management, ICT roles, or start businesses. One even built a successful roadside assistance app, now a thriving company. It’s proof of how one opportunity can ripple through an entire community.

Gavin: How do you prevent mobile AI training from reinforcing bias and ensure diverse data representation?

Wendy: It’s a major challenge. A strong upfront data strategy and model evaluation framework are critical. Think of global e-commerce for clothing: you need inclusive sizing and styles for all body types and ethnicities. The same applies to AI.

The issue worsens with existing LLMs, especially open-weight ones, which don’t make all of the source code public and may not share all details of the training data. You can’t just have models train models, because that perpetuates bias. Instead, you need a clear plan: defining the model’s purpose, sourcing diverse data, augmenting gaps, and continuously testing. Data and contexts evolve, so ongoing evaluation is key.

The framework must define what “good” looks like—whether avoiding gender bias or ensuring self-driving cars recognize bicycles. Without it, bias is inevitable. It all comes down to having the right data, rigorously checked.

Gavin: To what extent is unreliable connectivity a barrier to mobile AI adoption? And how should companies and governments work together to bridge that digital divide?

Wendy: You can't go anywhere without the internet or wireless infrastructure, so it’s a strong barrier. During the Covid lockdown, we had to move everybody to “work-from-home” and there were entire areas where our workforce lived that weren’t covered. So, we called the local providers and got them to put connectivity into these areas.

You have to have connectivity. We’re working with an organization trying to ensure sustainability in fishing. If you overfish, you not only can cause ecological damage, but you potentially could put everybody out of work. So, as part of an education program, we got them to use their phones to log their information, taking a picture of where they were, sharing data and accessing guides on when to move downriver to other areas. It was very powerful. They’re out in very remote areas and it's literally their day-to-day livelihood. You couldn’t do something like that without wireless, without connectivity.

Gavin: Are you optimistic that mobile AI – the alliance of technology and connectivity – can do more to bridge inequality gaps?

Wendy: Access to mobile applications will be a game-changer. In emerging countries, such as Kenya and Uganda, where we work, even in universities you don't necessarily have access to a computer. A common discussion when we meet these universities is “Can we use your GPUs?” Well, imagine a situation where we get large amounts of mobile devices – the computing power is significant. Access to that scarce resource won’t just help with education: now you’re able to leapfrog. And it's very common for anybody who leaves our company to want to purchase their laptop. Because that's how they can make a living.

Gavin: When you were appointed CEO, you were hailed as “part of a small group of female leaders within the male-dominated AI industry.” Is that gender imbalance improving; and, if so, what benefits will that produce?

Wendy: I’m definitely a fan of diverse experiences and viewpoints – you end up with better products. Gen AI isn’t exclusively built for one demographic, or one type of person, whether that’s gender, profession, or any other form of diversity.

AI should work for, and be trusted by, everybody, so there has to be diversity as part of the development process. There’s still not a huge amount of it, and we need to get more people of different backgrounds into positions where they can make a difference – such as more women engineers, or data scientists. It needs to start there, with the ones who are building it. And progress will come, a step at a time.