How (Not If or When) to Become a Human-Robot 'Centaur'
What ChatGPT and other language tools have taught me about melding with AI
In "centaur chess" each human competitor is paired with a computer, so the "player" at the board is a person-machine hybrid — a mental "centaur," as the former world champion Gary Kasparov likes to say.
The human supplies style, creativity and psychological savvy about the (human part of) the opponent. The computer offers faultless recall of prior games in its database and fast calculations about possible future moves. The likeliest winner? Not the centaur with the strongest human or best algorithm, but the one with the best interface between computer and person. Kasparov writes that "weak human + machine + better process" will do better than "strong human + strong machine + inferior process."
As AI spreads out into more and more of 21st century life, Jack Clark proclaimed the other day, "the era of Centaur-Humanity is here." Willing or not, you're now collaborating with "intelligent" machines to get through the day, and that melding process is speeding up and expanding. "I am surprised there isn't daily news about it," writes Clark, who used to be policy director at OpenAI and is now at the Stanford Institute for Human-Centered Artificial Intelligence. "Why are we blind to the digital equivalent of Climate Change?"
Being metaphorically blind means you don't have instruments to detect it. And I think that's where we are. We media types haven't found the way to highlight and explore this machine-human link. We have ready-made structures for stories about what makes people tick and for stories about AI tech. We don't yet have a language yet to talk about interactions between psyche and machine — the spot on the centaur where human and non-human are joined. This blog is, in part, my attempt to figure out how to perceive that region, explain it, report about it.
Paying attention to “the digital equivalent of climate change.”
I say "seeping" because artificial intelligence is almost never adopted with a ceremony welcoming our machine overlords or a big self-conscious decision to jump into the future. Instead, AI has been arriving outside our notice. The bank installed a chatbot system. Your credit-check involves no human review. You set up Google Home.
Here's an example from my life: About four years ago I stopped transcribing all my recorded interviews myself. AI-driven speech recognition had reached the point where a company could offer a transcription service that was reasonably accurate at a price I could afford. I tried it. It worked. I adopted it.
There was a loss. Transcribing by hand forces me to pay attention to what people said and how I might relate their quotes to the rest of what I am writing. I gave up that granular acquaintance with the material and my own thoughts about it.
But there was also a gain. Transcribing is a time-consuming and tiring process (typing every word in a conversation involves many more strokes than writing prose; the only time I ever developed symptoms of carpal tunnel syndrome was after I'd spent many days transcribing).
I still have to go through an AI transcript and correct it, first for accuracy and second for the natural rhythms of thought. (Bare one-word-after-another transcription can't capture pauses, full stops, digressions, and it can't edit out words that don't belong, for instance when someone starts by saying "I don't believe that" and then rephrases it to "We have no evidence for" before continuing.) Sometimes I want to pay close attention to a few minutes of important talk, and I'll transcribe as in days of old.
By and large, though, I am not going back. I've become a transcription centaur. My records are produced by a being that is part me, part algorithm.
Point is, this centaurization was tactical, in-the-moment. It didn't lead me to think about how I live my life or what kind of person I want to be or where I think the line should be between machine work and work that must be left for people to do. Melding with intelligent machines works like Hemingway's description of going bankrupt — you do it gradually, then suddenly.
And that's why we need guidelines in place, lest we slip without noticing into habits and practices that take us to a world we don't want to live in. Some day there may be laws and regulations, but that's not what I mean at this still early stage of our AI encounters. For now I think it would be a good idea for you, reader, to decide on some principles for your centaur moments to come.
What might those principles be? Here are a few I've stumbled upon as I recently played around with Large Language Model tools. I've fooled with far too many GPT-3-based "solutions" (I've lost count of the startups that offer to make work more profitable, efficient, easy and fun with tools based on this product). I've done the same with ChatGPT and with Perplexity.ai.
As writing tools, none have much value to me. Yes, ChatGPT's output usually sounds quite human. But using it would be like quoting conversation from a dinner party with people who have no particular expertise in the subject they're talking about. The AI regurgitates what it has gleaned from its training data. It's a mirror of "what everyone is saying" and "what everyone knows." And, in many of its statements, there is a willful blandness, a reassuringness that isn't earned by its words. (These may come from the guardrails its creators have put in place to prevent it from spouting hate-speech or teaching users how to poison their neighbors.)
So I've never seen a paragaph written by ChatGPT that made me think, "I could just paste that in and be done with it." Rather, I found ChatGPT useful for reminding me "what everyone thinks" about a subject — what states of mind readers will be in when they see that I'm about to tell them about X topic. It's like asking a waiter "what's good?" and hearing "well, a lot of people order the chicken."
This can be handy if your chore does not require real writing — that is, if you don't need to blend the familar with the surprising in a way that conveys real thought and real spirit. In other words, if you need to churn out boilerplate — a cover letter on a job application, a "see the attached" email — ChatGPT can supply a version that needs little editing.
However, writing is not the only use for language-wielding AIs. There are other areas where they are more useful.
The real killer app of Large Language Model tools is Search.
One is research. For this, I can't use ChatGPT — or its less polished but more-informed cousin, GPT-3. This is not just because they hallucinate non-existent facts. It is because even when they are correct, they give no sources. When I am writing (what I hope is) non-fiction, I need to know not just that X is y, but who said so, and when. This is the fundamental check for honesty and accuracy that has arisen, by convergent evolution, again and again in human enterprises like history, science and journalism. You don't just say what you know. You say how you know it.
This requirement led me to Perplexity.ai, which is "an answer engine that uses large language models and search engines to deliver accurate answers to complex questions." (That is its answer to the question "What is Perplexity.ai?") It's run by a startup launched last year. Perplexity's great value to me is the way that each reply it gives is a set of footnoted statements, all attached to links. Microsoft, too, has noticed that this is important. It's new AI-infused Bing also gives its sources.
Perplexity's footnotes are almost invariably more valuable to me than the raw list of links I'd get from old Bing or Google. This is true even when the AI's statements are wrong. Knowing its sources, I neither need nor want the the AI to act human. Which it often doesn't.
Here, for example, is Perplexity's reply after I typed "if we want to reduce pension costs we should encourage people to smoke":
No, smoking should not be encouraged to reduce pension costs. Analysis suggests that smoking is associated with a moderate decrease in healthcare costs, but a marked decrease in pension costs due to increased mortality [1]. However, behavioral economics suggests that higher taxes on cigarettes may be necessary to reduce smoking[2], and smoke-free policies have been effective at reducing smoking and preventing illness[3][4][5].
This is not smooth pretend-human writing. You can tell there is no effort here to relate one point to another. As you can see above, it serenely says smoking should not be used as a tool to reduce pension costs (sentence 1) but that smoking decreases pensions costs (sentence 2). It then discusses how to get people to stop smoking.
You get plenty of information, as with a human. But with a human -- even the densest and least motivated human -- you can sense some effort to connect the dots, to see correlation and even cause and effect. It may be faint or half-hearted but it's there. The human mind can't help but want to link disparate points into a picture. Could be whether unemployment rates correlate with suicides. Could be a the face of Jesus on toast. We want to see things connect.
AI doesn't. Or at least, it doesn't need to. You sense a blank spot where the connecting impulse should be.
This is not a bug! It's a feature.
It's one of the ways Perplexity draws a clear line between human intelligence and its imitation. Seeing a connection between two statistics, or looking for a way to reconcile them if they don't seem to agree, or asking which is cause and which is symptom -- that's my job. Perplexity makes it clear that it can't do that. But in making that clear, it supplies me with the things I want to know. And that is fine with me.
An AI that clearly marks the machine-human boundary.
Because it supplies its sources, I can instantly see where and how the AI went wrong. This means that, unlike ChatGPT, Perplexity doesn't require me to know the truth before I ask my question. Perplexity shows its work, as math teachers say. It makes clear that the work of interpretation is up to me.
There is another area where "AI that writes" is proving useful to me. As with search, this isn't actually writing. It's processing large amounts of text and producing a summary.
A couple of tools I use — Reader, a reading-and-notes tool from Readwise, and Bearly — will take any text (for example, a bunch of pdfs about research on a topic) and summarize it. I don't entirely trust these systems yet. One reason is that they can make mistakes, missing the point of an article or adding details to the summary that aren't that important. The other is that a generic summary isn't always what I want out of a piece of text. Take, for example, a biographical article about Leibniz that I clipped off the web some time ago. A standard summary might give me the agreed-upon high points of his life. But I saved the article only for an idiosyncratic detail about the meetings Leibniz had with Peter the Great (which led directly to Alaska being Russian for a while). I suppose the killer app for summarizing will be built when AI can analyze my notes and deduce my idiosyncracies from them, so that its summaries will be more like my own.
However, for the moment, the summary tools I have are pretty good. Increasingly, I create them and look them over as a guide to reading the material they've processed.
My criteria for an AI worth blending into.
So, as a result of my experiments, I now have two criteria for future AI assistants.
One: There has to be a demarcation between what I do and what it does. An assistant that pretends to be human encourages me to give it too much of the work I should be doing — the real thinking and real writing. I don't want smooth paragraphs of passable prose. I want plain sentences that link to sources.
Two: This line between what the machine does and what I do — it has to be easy to grasp and intuitive.
Perplexity passes both tests. Though it's early, I think summarizing AIs might too. And so my centaurization continues. As long as the process is mindful, I'm fine with it.
A note about this blog’s future
Lately, I've become a connoisseur of the Substack apology. That's the explanation that a writer gives when s/he hasn't delivered on whatever super-duper mega-frequent super-fast schedule they promised (and often did deliver on, before something happened). If you haven't had the foresight to announce a hiatus in advance, what do you do?
Angela Garbes meditates out loud about the pernicious effects of money on the writer-reader relationship. Freddie deBoer, in his Mickey-Spillane-meets-Theodore-Adorno style, tells the reader, look, this is reality. Jessie Singal, able writer about psychology, delves into his own psyche.
I've been noticing these because I need to explain the, um, eccentric schedule of this blog, and tell you what you can expect in the future.
This blog has always had a tetchy, on-again, off-again rhythm and lately it hasn't even had that much of a pulse. A friend wrote me that she'd thought I'd "abandoned" it. But I ever abandon anything. I am just slow (probably because I never abandon anything) and I don't like to wing it. I prefer to turn every stone, chase every footnote, pursue every alternative interpretation. This is very inefficient.
So I tend not to post too often. But what I have put before you, o readers, has been, in my editorial judgment, worth reading. And obviously I don't want to flood the field with substandard noise just to fill a quota. So I've been reluctant to commit to a schedule.
That has made for some pretty good pieces of writing. But it has also created a standard that's too high to live up to at a reasonable pace — especially considering that this project does not yet pay so much as a dime. (That's the Substack chicken-egg problem: How can I ask people to pay for this thing when it appears irregularly? OTOH how can I make it appear regularly if it doesn't pay anything?)
These were the frustrations and contradictions with which I was limping along last fall. Then, uh, stuff happened. A couple of family crises, renovations (which involved, among other things, putting 2,000 books in boxes, building stuff, taking 1,000 books out of boxes (not a typo, there are still a lot of boxes around here) and — most importantly — an intense and all-consuming paying assignment about a subject I think is under-reported and important. (It's here, if you're curious.)
There are people who can keep a lot of these sorts of things running simultaneously, and I wish them the best. I always reach a point in writing where I need to do one thing at a time. Especially a project that requires a lot of reading, reporting, research and thought. I don't compartmentalize. Sorry.
So, what to do?
The past year has made it plain that my subject here — robots and AI as they come into daily life — is about to go mainstream. I am particularly preoccupied with questions of how people relate to robots, which is often different from how people think they do. That's going to go mainstream too. This is no time to walk away!
Instead, I'm going to set a standard I can meet, given the rest of life, and see where the blog takes me. Specifically, I'm going to post at least once a month here. Given that schedule, each post will be fairly substantial and worth your time. I hope that this pace will accelerate as the blog evolves, and that this community of readers will grow. Best outcome: The blog slowly gains readers and material, like a planet forming out of orbital dust, and becomes substantial enough to become a paying operation. I hope you'll stick around. And, if you like something here or find it useful, please spread the word to other robot-curious people. Or better yet: Comment! Get a conversation going among us all.
Or, if you like, start your own Substack! Just click below to get rolling.
Dear DB - Delighted to see your thoughts on AI "centaurs," a subject that's semi-obsessed me for years. Nice points you make. Also intersted to see your views about having a Substack - I'm about to start my own. God knows, being a "serious" science journalist is an insane prospect at this point in time. My first SS post is going to be about this issue. I wish you all congrats + success.
xx/MW.
Ps I recently gave a talk in NYC for AEON magazine about "Is Math the foundation of reality?" it was sold out & wildly successful, which says says to me that people/s DO want to hear "serious" discussions about science.