When I was around 10 years old, I spent some time contemplating the type of scenario usually reserved for folks on acid or at the very least strong weed:
Imagine, I thought,
if you spoke a completely different language to me. Imagine further that all of the words in that language were exactly the same as the words in my language, but they meant completely
different things. So if I say “blue,” you hear, “blue,” but I’m thinking about a certain range on the electromagnetic spectrum, and you’re thinking about, I
don’t know, pasta.
Now imagine that it just so happens that your interpretations “work” with my interpretations. So I ask a question that means something completely
different to you, and the words you use to answer what that question means to you comprise a sensible response to the question I was trying to ask.
Confused yet?
This is what it
comes down to: I actually have no way of knowing how you interpret what you see and hear. I can only guess, based on cues like the way you respond, your tone of voice, your facial
expressions, your body language.
advertisement
advertisement
And I’m wrong -- as are you -- all the time. As humans, we are constantly making up stories about other people’s intentions, motivations
and actions. We say things like, “He just doesn’t care about his work,” or “She was really grumpy in that meeting,” but we don’t actually know.
Maybe he’s working as hard as he possibly can but doesn’t have the skills. Maybe she just got some bad news about a family member and is distracted. Or maybe we’re right: he
doesn’t care, and she was grumpy. But we don’t know.
Because we’re all wrong so often, we’ve built it into the way we interact with each other. We expect
misinterpretations and misunderstandings. We expect disagreements and distinct points of view, frictions and fragmentation.
What we don’t expect is some of the completely
left-field stuff we get from generative AI. We don’t expect, for example, that an adult will not be able to tell time on an analog clock, reporting every single watch face as showing 10:10
because that’s what the vast majority of images in the data corpus are set to.
We don’t expect a seeming hyper-genius not to be able to do math reliably. We don’t expect humans to believe that cities are in the ocean.
These are all glitches in the current crop of large language models, or LLMs, such as ChatGPT. And
maybe it’s not such a big deal. But the problem is not their inaccuracy. The problem is that they’re inaccurate in ways we don’t expect and can’t fathom -- and that makes the
inaccuracies hard to manage.
I would expect LLMs to have absorbed a lot of historic inequity. If our datasets are embedded with biases, so will our AI be. So when I’m asking for an image
of a CEO, it’s unsurprising to me that it’s more likely to show men with lighter skin.
What I did not
expect was that generative AI would insist on racism. If you ask for an image of a Black African doctor treating white children, it straight-up refuses, only generating images
of Black children, and not infrequently depicting the doctor as white.
As these models get embedded across more and more of our systems, as we rely on them more and more, the risks of
unpredictable, anomalous malfunctions grow. It’s not because AIs aren’t perfect -- nobody is. It’s that they’re imperfect in ways we haven’t yet learned to
accommodate.
AIs are speaking a language that seems the same as ours, but it’s not. Let’s not be fooled into thinking we understand each other.