Apple aims for certainty, not hallucinations, in AI
When it comes to improving Apple Intelligence and keeping the promise of contextual intelligence next year, Apple is testing its own internally developed artificial intelligence chatbot, code-named Veritas, the Latin word for “truth.”
Perhaps that code name is also an ambition. If so, it is a welcome target, given the studies that show the percentage of AI chatbot answers that are hallucinations (the current buzzword for “wrong”) is as high as 10% to 70%, depending on the model and benchmark used.
If Apple can meet or exceed that, then it has a marketing campaign right there.
Will truth set Siri free?
As a thought experiment, let’s imagine that Apple finally manages to create a system that delivers wrong answers at 10% or less. Imagine, in other words, that Siri is always able to provide accurate answers. Would that change your daily interactions with the tech? I guess it would give the company the chance to use marketing expressions like:
“Siri: The truth will set you free.”
“Siri: Truth conquers all.”
Or even “Siri prevails.”
That’s assuming, of course, that Apple’s AI can live up to its code name. Right now, reports tell us, the company doesn’t intend to put its Veritas out there, with the bot confined to internal testing — presumably until it lives up to that name.
What else do we know?
Bloomberg says Veritas is being used to experiment with new Siri features. It presents itself similarly to other existing tools, so that means typed queries, conversations, and the chance to revisit previous conversations.
The decision not to introduce it to the public may seem a little counterintuitive, given the race to AI that is taking place, but the company is making use of it to finesse upcoming Siri actions and upgrades. In other words, I think, it’s being used to help ensure those new actions are much more consistent when used.
So, when you do use Apple Intelligence, you should be able to rely on the results and not have to double-check for false information embedded in whatever it is you use it to create. An AI you can trust.
That’s a logical expression of Apple’s stated approach to AI. At WWDC, Apple SVP Software Engineering Craig Federighi said the company is focused on developing AI to do the kind of things we do every day, rather than creating what he described as a “bolt-on chatbot on the side.”
Getting it right first time
But, of course, that focus on tasks makes it even more important to ensure that the tasks Siri gets to do are done efficiently, accurately, and effectively. Anything less than that will not be good enough, particularly after such an extensive and well-publicized wait for an upgrade to the original chatbot.
There are other details emerging to characterize Apple’s approach.
For example, we already know that Apple Intelligence works with OpenAI’s ChatGPT, but in the future we can expect it to work with other AI engines, beginning with Google Gemini. It seems likely we’ll see it work with Anthropic’s Claude, too, given its inclusion in Xcode even as the latter company ramps up its global capacity, presumably to deal with a future spike in demand for its services.
It’s a dual-solution approach, with Apple building its own AI models for some tasks and working with external services for others. The idea is that Apple Intelligence will offer access to third-party services from within its interface, just as it does already. Apple is also developing its own search experience, called Answers, which we must also hope won’t be prone to hallucinations.
Where are we now?
Apple is under pressure, and the company knows it.
In its response to the AI arms race, it is working to control the battlefield. It doesn’t want to engage in a chatbot war, as doing so would cost many billions of dollars and in an already widely populated market would deliver little to differentiate its products. It is, after all, inevitable that not every major AI company will make it through the next decade. Consolidation is going to take place — if only because of the high energy and technology infrastructure costs of running operations like these.
Instead of choosing to place its flag on that hill, Apple is looking to identify the AI-augmented activities most likely to make a difference to its customers. It also seeks to differentiate its approach in order to bring a unique AI value proposition to its ecosystem.
That’s a nuanced approach, and it absolutely requires the front-end AI to be absolutely and unfailingly accurate in the responses it creates. If you want hallucinations, try another platform. We shall wait and see if Apple can get to that point, but right now, no one else has, either — so perhaps the company has time to get there.
Would you use a smarter Siri that didn’t hallucinate? How likely do you think it is that Apple will be able to bring it to reality? If Apple cannot, who can, and why? Please let me know your insights.
You can follow me on social media! Join me on BlueSky, LinkedIn, and Mastodon.Microsoft upgrades M365 Copilot with Agent Mode – ComputerworldRead More