A Furry Little Robotic Demon-Spawn Helped Me Understand AI Policy
Stop me if you've heard this before: a Calculator and a Furby walk into Congress to talk about Artificial Intelligence
What do we mean when we say "artificial intelligence?" A colleague and I recently wrangled with this question while we were trying to update a few legal documents at work. It's not that we don't know what artificial intelligence is - we do. But it's challenging to capture what we mean when say "artificial intelligence" in a few words. As my colleague and I worked through various definitions, I lamented that my dumb calculator would satisfy some of these definitions, and then that my Furby was, for all intents and purposes, a HAL-level AI based on these definitions.
Thus was born the Calculator Test and the Furby Test
Other contenders from my snarky commentary were, in no particular order:
The Tamagotchi Test
The Roomba Test
The "Took an Arrow to the Knee " Test
The Super Fancy Bidet Test
The robot-mouse cat toy that my cat murdered in 3 seconds flat Test
These tests are dual-purpose. First, they provide you with a framework for evaluating your definition of AI, or any other definition of AI based on real-world, well-known examples from recent-ish history. Second, more importantly, it challenges your framing of how you think about whether your software solution qualifies as AI under various laws, or doesn't. Drawing parallels between a calculator, a Furby, or any other similar style toy or device will allow you to evaluate your own machine-based systems more robustly by challenging your preferred framing.
Who knew that a furry little demonic toy from 1998 would teach me something about Artificial Intelligence policy today?
Why the Calculator and the Furby?
I'm going out on a limb and assuming that everyone knows what a calculator is, but not everyone might know what a Furby is. For anyone who doesn't know what a Furby is, it's a Robotic toy launched in 1998 that claimed to "learn" English from interacting with people, could interact and "learn" from other Furbys, and was actually banned from the NSA over concerns of it leaking Top Secret information. It's a cute, furry owl-hamster-demonic-entity thing that for some reason I absolutely love ... for whatever that says about me.
Honestly, the Calculator Test is pretty utilitarian and uninspired, but it gets the job done. A calculator is one of the most basic functions a computer machine can fulfill - basic math. So, if we can apply a given definition of AI to a calculator and have it reasonably (and this word is doing a lot of heavy lifting here) satisfy the requirements, then the definition is too broad. If it can applied to a calculator, it can be applied to every single piece of software ever conceived, which isn't great. If you satisfy the Calculator test, it's time to back to the drawing board.
The Furby Test is more stringent. The toy responds to inputs and gives the illusion of learning but doesn't actually utilize any modern machine learning techniques as we consider them today. (I'm referring to the 1998 version of the toy, the 2012/2023 versions are more sophisticated). But knowing that it responds to inputs and gives the illusion of making decisions and learning, if your definition of AI satisfies the Furby Test, meaning that your definition would include the Furby, but doesn't satisfy the Calculator test, you are on the right path and likely just need to tweak some things.
Why the Furby? Like I said, I kinda love the stupid little things. But they also represent a really important mix of responding to human and robot interactions and convincingly imitating AI-like functionality. They aren't something that we consider when we talk about artificial intelligence today, but they bear all of the hallmarks of artificial intelligence.
These two tests validate whether a given definition of AI is too broad, and by how much. It works like this: if I read a definition of AI and can reasonably make that definition apply to a calculator, the definition is too broad. If I can read a definition and make it apply to the 1998 Furby, but not the calculator, it might be too broad but it's definitely on the right track.
What is a legal definition?
Before we talk about these tests, let's start with why they can be helpful as we try to navigate legal definitions. A legal definition is how something is defined within a piece of legislation and it determines what's covered in the legislation and what's not. Most pieces of policy will have a definitions section, where they spell out and define all the key terms through the legislation. For example, gambling also has a legal definition, based on the three pillars of "chance, consideration, and prize," and that definition is how some sleazy video games get their commonly understood as gambling tactics to avoid actually being legally defined as gambling. It’s also important to remember that a legal definition will differ from a popular culture definition.
Why does this matter AI? We want to make sure that definition captures only what we intend it to, and not the myriad of AI-like things. If we are sloppy with our definition of AI, we could accidentally have our agreements and laws cover types of software we don't intend, or miss a piece of what we do intend to cover. Like gambling, we don't want folks to be able to skirt the law with AI through sloppy definitions.
A Common AI Definition
So let's try it out. Let's start by taking a common definition of Artificial Intelligence from Coursera, "Artificial intelligence (AI) refers to computer systems capable of performing complex tasks that historically only a human could do, such as reasoning, making decisions, or solving problems."
That's straightforward, right? Let's apply the Calculator Test.
Is a calculator a computer system? Yes, I think it is. The definition doesn't give us a lower bound on the capabilities of the system, so I can reasonably say it's a computer system as defined here.
Does it perform complex tasks that historically only a human could do? Maybe? We don't know what "complex tasks" actually refer to here, and it could easily be argued that even a dumb calculator is capable of finding the answers to complex mathematical problems that historically required a human to complete.
Does its task include reasoning, making decisions, or solving problems? Yes, it does solve math problems for us.
VERDICT: FAILS the Calculator Test because a calculator can reasonably be said to satisfy all 3 criteria of the definition.
Since we know the Calculator qualifies for the definition, the Furby definitely will so there's no need to run the Furby test until we can pass the Calculator test.
Take it Up a Notch with actual legislation
So let's take a look at how some laws have attempted to define artificial intelligence. The US National Artificial Intelligence Initiative Act (15 USC 9401(3)) defines AI as, "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. Artificial intelligence systems use machine and human-based inputs to perceive real and virtual environments, abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action." This definition has 3 criteria:
A machine-based system
that can, for a given set of human-defined objectives
make predictions, recommendations, or decisions
influencing real or virtual environments
use machine- and human-based inputs to perceive real and virtual environments,
abstract such perceptions into models through analysis in an automated
manner
use model inference to formulate options for information or action
The "model" here refers to the AI model, which is defined in the same Act as "a component of information system that implements AI Technology and uses computational, statistical, or machine-learning techniques to product outputs from a given set of inputs.
The Calculator Test
Machine-based System - no guardrails, so yes.
that can, for a given set of human-defined objectives - Debateable here, but I'll say yes under the logic that entering the equation into the calculator is a human-defined objective
make predictions, recommendations, or decisions - I'm struggling to say that the output for a mathematical equation is a prediction, recommendation, or decision. I'm going to say a calculator doesn't satisfy this requirement
influencing real or virtual environments - A calculator is a siloed system, so it doesn't have any influence on a real or a virtual environment.
use machine- and human-based inputs to perceive real and virtual environments - This one is difficult because some virtual environments can only be expressed by equations and mathematical formulae. I'm going to say it doesn't satisfy the criteria for this because although there are human-based inputs and virtual environments can be expressed through mathematical means, I don't think the calculator can express those.
Since the calculator doesn't satisfy this requirement, it cannot satisfy the following 2 requirements either.
VERDICT: PASS - A calculator does not fit this definition of artificial intelligence.
The Furby Test (woo I'm excited!)
Machine-based System - no guardrails, so yes again.
that can, for a given set of human-defined objectives - I think the Furby achieves this with flying colors because it is more sophisticated, and its objectives are codified into its literal code
make predictions, recommendations, or decisions - This one is iffy. I'm stuck on what we mean by "decisions" here - if it refers to a pre-programmed decision based on inputs, then Furby satisfies this requirement. However, since it doesn't call out inferred or pre-determined decisions, I'll say that Furby satisfies this criterion as well.
influencing real or virtual environments - The 1998 Furby had limited perceptive capabilities, as we see from the sensor array between its eyes. Because it could communicate with other Furbys and 'teach them things', I'm going to conclude that's satisfying a virtual virtual environment. Furby satisfies this criterion.
use machine- and human-based inputs to perceive real and virtual environments - Weirdly, I think Furby satisfies this criterion with flying colors. But I think it's because of a loophole - the definition doesn't require that machine- and human-based inputs be directly tied to specific perceptions. So while the Furby didn't respond to specific words people said, it did respond to the noises that people made to make it mimic learning.
abstract such perceptions into models through analysis in an automated manner - This one is tricksy because we don't really know what the 1998 Furby was doing with the sounds it was processing. It responded to those sounds to some degree in an automated manner. Interactions between two different Furbys required perceptions of the other Furby, and then some processing of the "conversation" that the Furbys were having without any human interaction. I say Furby satisfies this criterion as well.
use model inference to formulate options for information or action - This is where I think Furby fails us. Whatever Furby was doing, I doubt it was using machine-learning model interference to formulate options for information or action. I think it probably used more traditional software development techniques to formulate options for information or action. There's a chance here that I could be wrong, but this would be a weak justification. Furby fails to satisfy this criterion.
Verdict: PASS - If it weren't for the specific callouts for model interference to formulate options, the 1998 version would have satisfied this criterion, which would have meant that the definition was still probably too broad.
Make It Your Own
Use this approach when you are trying to come up with your own definitions of Artificial Intelligence if you aren't using the EU AI Act definition or the US National Artificial Intelligence Initiative definition. But more likely, you'll want to use this as a model to apply to your own solutions to determine whether they satisfy the definitions and what that means for your legal obligations.
Either way, I just got to write a whole freakin' article tying Furby to AI legislation. Pretty damn awesome, if I do say so myself.
What about you? What tools or techniques are you using to evaluate AI definitions and your own AI solutions for regulatory compliance?