Meta AI dials users in India as battles with OpenAI, Google and Apple commence

It took a while. Meta AI, which is the company’s generative artificial intelligence (AI) chatbot, is now rolling out for all users in India. This means that all of Meta’s apps, such as Facebook, WhatsApp, Instagram and Messenger, will now have the colourful circle within which shall reside the ability to search the web, text to image tools, get contextual inputs such as organising projects or trip itineraries in group chats, create stickers, learn more about a post on the feed and can be called upon for suggestions or answers to queries, at any point. This means Meta’s AI battles with OpenAI’s ChatGPT (and as an extension, Microsoft Copilot), Google Gemini as well as the upcoming Apple Intelligence, begin in earnest.

In case you are curious about Meta’s timing for tapping the India market now, it has to do with gaining a foothold before Google’s widened availability of AI integration within Android phones (such as Gemini replacing Assistant) makes its mark, as well as Apple’s upcoming Apple Intelligence suite which includes a contextually smarter Siri assistant and an integration of the GPT-4o models within iOS 18, iPadOS 18 and macOS Sequoia. However, Ryan Cairns, who leads engineering for AI at Meta, insists that isn’t the sole reason behind choosing this month to make Meta AI available in India.

“The benefit of the way that we’ve rolled out Meta AI is that it’s available in all of our apps, regardless of the platform. You can access it in WhatsApp, whether it’s on Android or iOS. You get it on the web, and we really want to be everywhere that you are. Right now, our apps provide both the biggest distribution and most availability to the end user,” Cairns tells HT. Numbers lend credence to Cairns’ argument.

India is, by far, Meta’s biggest market. By enabling Meta AI for every user in the country, and therefore a presence on every device they use, it gives Meta a significant numbers advantage. WhatsApp, according to the latest numbers by research firm Statista, has 535.8 million users in India – its biggest market, with Brazil in a distant second with 118.5 million active users. Meta’s social media platform Facebook also sees India’s contribution with 378.05 million active users at the end of 2024, significantly ahead of the US (193.8 million) and Indonesia (119.05 million). For Instagram, India’s contribution of 362.9 million active users clocked at the turn of the year, is far ahead of the US (169.65 million) and Brazil (134.6 million).

“I think that’s something which doesn’t really exist when you look at it from the operating system perspective,” adds Cairns. It is true that a lot of Android’s AI functionality as it rolls out to phones in the coming months, won’t be available on most apps outside of that device – a Chromebook, for example. Apple Intelligence, which will be an extensive layer of smartness around the OS’ for iPhone, iPad and Mac, will figure much less prominently if you’re using the same apps on an incompatible (read, older generation) Mac or iPhone, or even an Apple Watch.

In fact, if I am not wrong, Meta AI is presently the only AI chatbot that accesses Google Search as well as Microsoft Bing, for web searches.

India gets a smartened version of Meta AI, compared to how it was when first released in September last year. At its foundation now is the Llama 3 foundational open-source model, which was released in April this year. In fact, working behind the scenes for Meta AI will be Llama 3 in two sizes – one with 8 billion parameters and a larger 70 billion parameter model. Compared with Llama 2, core improvements are around lowering false refusal, improve response quality, diversity and improve reasoning.

“It’s our most advanced LLM to date,” says Cairns. He details that in the testing phase in India, it was observed that a lot of use cases revolved around information gathering. “The biggest uses that we’ve seen from our test are things like asking factual questions, getting instructional and how to advise and tutoring and learning support,” he says. Meta details that these can be done using one to one threads with the assistant, or even in group conversations.

HT asked Cairns if any part of Meta AI works on-device, and for now, he clarifies that all computation and processing is done online, that is, on a server. “We are primarily server side. Today, we don’t do much to sync conversation history and things like that, but if you use the website, you can see a hint of the thread history that we’re looking at. You can imagine that that might be useful on different devices, either to go deeper in a topic or to resume a particular task,” says Cairns, before letting in on Meta’s thought behind the evolving Meta AI structure.

“Let’s say, if you were trying to generate code for a project. It’s multi-step, and you’re taking a lot of iterations. You might do that in different places. We think it’s important eventually to get to the point where those things sync, but we don’t have that today,” he adds.

There is often the concern with generative AI, about incorrect responses for genres or topics that may otherwise require some level of expertise before a specific recommendation can be given. HT asked Cairns if there are any genres or question types that Meta AI will stay away from answering. “Health and financial advice are two good areas where we can’t answer. We can’t give medical or financial advice. Areas of self-harm are another topic we avoid, he says.

Meta details their approach when a user asks a question on a sensitive topic. “One of the things that we had fixed were refusal rates, We wouldn’t necessarily give advice, but someone might ask a question. We could say, well, that’s a really interesting question. Here’s some balancing results so that people can do some more information digging on their own, but not necessarily refuse the prompt outright,” Meta details.

Meta has a strong generative artificial intelligence (AI) product, and the timing does give it a strong platforms as generative AI tools contend for user attention and retention. Meta’s advantage is its apps have a very active user base, which places Meta AI front and center for access. The company insists that no user data will be used to train the AI models. “We use publicly available information to train our models. It’s an industry wide practice, not unique to our services, and we’re going to continue to do that,” says Cairns.

For now, Meta AI will only support the English language in India. However, while they aren’t revealing plans for local language support, Cairns does refer to the fact that the Meta AI models are updated every couple of weeks or so. Which means adding more languages, alongside fixing feedback around hallucinations, information accuracy and safety, wouldn’t take much time.