Choice of your AI assistant? Nope, probably not
AI interoperability, and why it matters
On an iPhone, you’re stuck with Siri. Fine — maybe you don’t use it that much.
But what happens once Siri has become good enough (for instance, through a deal with OpenAI) that you depend on it?
Answer: you’re still stuck with Siri. And that’s a problem.
We’ll increasingly use AI-enabled operating systems (OSes). At their heart will be AI assistants like Siri or Google Assistant.
As AI improves, these will become seriously useful — maybe even central to how we use our devices. We’ll speak to them, task them, get their help with emails, managing our finances, posting to Instagram.
The usefulness of an AI assistant will depend on two things:
Intention-translation: Its ability to translate our Intentions into complex Inputs for the OS which faithfully match the original Intentions. The assistant hears ‘Set up a meeting with Santa Claus to discuss this year’s gift requests’, and knows what it needs to do: find a convenient time and send Santa Claus a calendar invite.
Permissions: The Permissions the assistant is given to turn those complex Inputs into Results via the OS. The assistant is able to access my calendar to find a good time for my catch up with Santa Claus, and send over a calendar invite to Santa Claus’s email, which it pulled from my contacts.
Both are important. But while anyone can freely innovate in Intention-translation, big tech firms like Apple and Google call the shots over who gets Permissions in the OS.
That’s bad — because without Permissions, Intention-translation is limited in its usefulness. We’re already seeing this happen. ChatGPT is smarter than Siri when it comes to Intention-translation. But the ChatGPT app is limited in its usefulness on our phones, because it has limited OS permissions. For instance, it can’t read and send messages across your various messaging apps, act across multiple apps, or change system settings. (Hence their deal with Apple — although it’s important to note that this would likely be to enhance Siri’s capabilities, not provide an alternative assistant.)
A smart assistant without permissions is like Jeeves being permanently stuck in your attic. You can still talk to him through the trapdoor, but he can’t do much for you.
Apps like ChatGPT, which aren’t given permissions at the OS level, can try to work around this by maxing out on the permissions that normal apps can request (like the ability to manage your calendar, or to look through your contacts), as well as getting some permissions through the internet — requesting that we sign into third party services like our bank, or our email. In other words, Jeeves can’t get out of the attic, but there are some strings he can pull through the trapdoor.
But:
This is laborious, compared with access at the OS level. For many services, you’ll need to log in on a case-by-case basis on Jeeves’s behalf before he can use them.
Not all services and data you use will have great APIs, or any API at all (e.g. most messaging apps).
Defaults matter. Jeeves is stuck in the attic — and the default assistant is downstairs, getting things done. Who do you turn to?
By contrast, if Jeeves is an OS-level assistant with extensive permissions, everything’s laid out on a plate for him. He roams freely around your house, doing your digital laundry, tidying your digital rooms. You don’t need to log into email for him, because he can already do and see everything on your device that you can. It’s seamless.
So what we’re up against is walled gardens, AI edition: a future where OS-makers like Apple and Google are able to limit meaningful choice over all-powerful AI assistants we use in our OSes, unless they’re forced to grant permissions to third party assistants — probably through regulation.
Assuming that doesn’t happen, the only assistants with the permissions to be seriously useful will be big tech’s own. And even then, only when run on the devices they make themselves.
So, for instance, Facebook probably won’t be releasing an AI assistant for iOS any time soon — let alone an organization like Mozilla which is developing AI with far purer motives than any big tech company.
(And yes, I’m ignoring the niche worlds of people doing things like jailbreaking or sideloading, as well as desktop computing where OSes are more open. For most people, mobile OSes are the primary way they’ll interact with assistants.)
Big tech companies will invoke concerns around privacy and abuse of AI to justify locking Jeeves in the attic. And there’s certainly a strong argument for vetting the AI assistants that are allowed on our devices — but big tech shouldn’t have a say in that vetting process.
We can use the word salad ‘AI interoperability for consumer-facing devices’ to describe this area, and I’ll be digging deeper in a series of articles — trying to address questions like:
Why is a lack of AI interoperability going to be a problem?
How will it undermine the struggle for trustworthy, responsible AI?
How will it make billions of people worse off, particularly in the Global Majority?
What’s wrong with being stuck with the assistant that ships on your phone?
And most of all:
What can we do about it?