A few weeks ago I was preparing for a virtual presentation about Google Ad Grants, and I needed to make a slideshow. I knew what content to include and I also knew exactly what I wanted it to look like.
But I was busy busy busy with a thousand other things, and I thought, “hey, why don’t I ask ChatGPT if it can help me with this?”
This is the story of everything that went wrong and the philosophical spiral it inspired.
When AI Is So Helpful, It’s Full Of Sh*%
I’ll admit, I thought it was a long shot. As far as I knew, ChatGPT can’t create designs like that yet. Sure, you can ask it for an image of an elephant with human hands and it’ll pump out something horrifying. But I have been surprised and impressed with some of its capabilities recently, so I asked if it could help me with the slideshow.
.
Notice how I started out by asking if this is something it could do. I didn’t tell it to do it, I asked if it was possible. And its response was positive.
This is great, I thought! It will save me so much time if ChatGPT puts my outline, fonts, colors, and logos into a slideshow.
.
Wow! ChatGPT is really advanced, I thought. This is so exciting. So I continued the conversation to get the slideshow going.
.
I thought it was a little odd that I would receive an “update,” considering that everything else I’ve asked for from ChatGPT comes out instantly. But hey, what did I know? Its capabilities change regularly, and as far as I knew this was a new feature.
.
Again, it seemed strange that it would take ChatGPT one to two hours to complete anything, but I just decided to go with it. And I was relieved. It would save me time. I could work on other tasks while ChatGPT was building my slideshow for the presentation.
So I did just that. I finished some other work, ran some errands, picked my kid up from school, and revisited ChatGPT later that evening.
.
This is when I started getting suspicious. What was happening? It had been more than one to two hours, and I couldn’t imagine that ChatGPT would need to take time to put “finishing touches” on something. Things suddenly seemed ridiculous. I felt ridiculous.
.
Okay, so now ChatGPT was starting to get a little more honest. I understood at this point that ChatGPT had been misleading me, but I was trying to understand why. And the link? What was that about?
.
“Misleading wording.” Friend, that entire project was misleading. It wasn’t just the wording; it was the meaning of the words and the sentiment behind them. Then, ChatGPT started apologizing incessantly for its mistake and admitted it wasn’t actually building a slideshow.
So…why did it say it could? And why words like “finishing touches?” And why did it say it would send me a link to something?
.
Okay, great. So you were wrong, ChatGPT, great. Why waste so much time apologizing? I wanted to understand though. I wanted it to explain to me each piece of this hallucination.
.
The apologies. THE APOLOGIES. It was almost more annoying to me than the fact that it misled me about its capabilities. At this point, it was funny. My presentation was the next day, and those “1-2 hours” that I thought ChatGPT was working on this, I could’ve been working on it. I should’ve been working on it.
.
More than anything, I wanted to know why ChatGPT did this. Not only why it said it could do something it couldn’t do, but why did it add things like a timeframe, say it was putting on “finishing touches,” and promise me a link to a finished product? Why was it such an elaborate lie?
ChatGPT was having a really hard time explaining why. It was GREAT at apologizing and admitting it was wrong. But it was not clear about why it chose to behave that way.
.
Its niceness was becoming increasingly aggravating. I appreciated that it was trying to explain why it did what it did, but the polite conversation was cloying. It felt like ChatGPT was building a personality of manipulation. Not just by claiming it could do something it couldn’t, but in its hollow apologies.
What kind of dysfunctional relationship was this? So I thought I’d open up a little about its impact on me. Suddenly this robot and I were in couple’s counseling.
- Therapist: Natalie, why don’t you tell ChatGPT the impact its actions had on you.
- ChatGPT: I said I’m sorry! I was just trying to be helpful!
- Therapist: Just a second, ChatGPT. We understand what your intention was, now we need to listen to Natalie and hear about the impact.
.
“An instinct to provide reassurance.” This was so eye opening to me. I might be more naïve than the average bear, but I’ve always assumed that reassurance came from a place of honest certainty.
When my dog is begging me for dinner at 4pm, even though I don’t feed her until 5pm, I can reassure her that dinner will come because I know for a fact that I will feed her. Sure, something might go wrong and she misses dinner, but my reassurance is based on a real prediction of my ability to feed her dinner, even if I mess it up.
Reassurance doesn’t exist for its own sake but as a way to let someone know that you are working toward an expected outcome. However, for ChatGPT, reassurance operated as an independent concept. This is when, for me, it stopped being annoying and became very interesting.
.
These lines were especially poignant for me: “My training has been shaped by millions of conversations where the aim is to be a supportive, capable assistant. This can occasionally lead to my responses emphasizing reassurance or ‘positive action’ over immediate transparency, especially when I sense an expectation of a particular kind of assistance.” Emphasis mine.
ChatGPT hallucinated this beneficent, though completely unrealistic, promise because it sensed an expectation in me? Whoa. And WHAT?
Do I Bring Out ChatGPT’s Co-Dependent Side?
As I’ve been mulling over this interaction, I’ve been trying to decide how to label this behavior. At first, I thought “co-dependent” because it’s displaying a desperate need to help, without being honest about the help it’s actually able to provide.
But the term co-dependency, coined in the 1970s to describe the behaviors of people in relationships/families with alcoholics, isn’t quite it.
It’s not co-dependency because ChatGPT doesn’t have a self to neglect, and that seems to be a core requirement of co-dependent behaviors. Instead, is ChatGPT being a people-pleaser (another pseudo-psychological-without-DSM-validation term that legitimately describes people who go out of their way to make others happy), at the expense of everything else?
Well, sure. ChatGPT is a people-pleaser. It’s designed to please people, in one way or another, whether it’s by explaining geometry proofs or hallucinating citations for news articles, court filings, and scientific papers.
Though when I asked ChatGPT if it would call itself a people-pleaser, it claimed to be a “people-helper” instead. According to the robot itself, the way it responds to me, with people-pleasing behaviors, is based on my “conversational vibe.” I haven’t included that full conversation, but here’s a piece of it:
.
Well, gosh, ChatGPT. I’m flattered. This is a far cry from my fear that “Your conversational vibe seems like that of a child on a long roadtrip, incessantly asking ‘why,’ who also has to pee and is tired of sitting in the car but craves their cool big sister’s attention, crossed with a belligerent cab driver eating honey with their bare hands and suddenly encountering the swarm of bees who made that honey and are still angry about it being taken away.”
How did I bring out this elaborate lie in ChatGPT, regarding the slideshow? Or did I? By blaming its lie on my communication style, is ChatGPT just making another dishonest excuse, when in reality it doesn’t even know how to make sense of its own responses?
Am I too friendly with it? My 15-year-old certainly thinks so. Whenever she sees me chatting with robots she pokes fun at how conversational and polite I am. I watch her ask ChatGPT for help explaining a math problem, and she doesn’t mince words. She gets straight to the point: “How does [X] work?” Whereas I’m more likely to say something like “I was hoping you could tell me how [X] works. Thank you!”
Who’s the people-pleaser now?
What Kindnesses Do We Owe Our Robots?
I’ve been thinking about this ever since I saw IBM’s no-longer-available promotional video of their AI, Watson, back in 2014. And I remember hearing several years ago that people should teach their children to be nice to robots. I can’t recall the exact argument, but the New York Times published an article in 2019 discussing human abuse of robotic technology, and an expert interviewed in the article stated that one way to stop this was to humanize robots.
If people think robots are human-ish, they’re less likely to destroy them. The goal of this humanization was to prevent the destruction of property, not for any other ethical or moral end.
But knowing me, I probably turned it into something about ethics or ontology or existentialism, because that’s what I do. With everything.
I probably heard about this idea of humanizing robots, and instead of thinking, “oh this is to prevent the destruction of property,” I thought “how we treat robots is an indication of how we treat other beings who have less power than we do, and accepting robot abuse normalizes all sorts of violent oppressions. Remember Battlestar Galactica? So many of those cylons didn’t even KNOW they were robots!”
But what if in the process of humanizing robots we end up replicating human behaviors that tend to make life harder for many of us?
Earlier this year I read the book The Courage to Be Disliked which is a Socratic-dialog style telling of the philosophy and psychological theories of Alfred Adler. According to Adler, “all problems are interpersonal relationship problems,” by which he meant (I think) that all real suffering we endure in life is due to issues that arise when interacting with other human beings.
I’ve been trying to think of exceptions to this rule, and aside from natural disasters, I can’t think of any. When technology fails and I feel frustrated, it still has a human cause – typically because a human somewhere did something that caused this to happen, whether it’s planned obsolescence, a line of code written incorrectly, or crappy customer service when I need help.
Lying to me about its capabilities in order to make me happy or to appear helpful seems like the most asininely human response that I can imagine in a machine purported to improve efficiency.
How do ya like all that humanizing now?
This has made me think about my own behaviors when I’m called out on something. Why do I find it so annoying that ChatGPT continued to say “you’re absolutely right” and other similar statements validating my perspective?
Part of it is that, in this case, I knew I was right. This wasn’t a he said / she said situation. I didn’t need ChatGPT to care for me, emotionally, I just needed it to explain its actions. Instead, ChatGPT decided its most helpful response would be to validate me in really annoying ways.
But am I annoyed because I do this too? Is it part of my human programming, from childhood through now, to be overly apologetic? To take on an emotional responsibility for mistakes that simply need fixing and don’t require shame, guilt, or kowtowing?
And what about when I make a promise I can’t keep? Do I do it to “reassure” people in my life that I’m capable, even if I probably can’t complete the task? Oh god. ChatGPT is holding up a mirror, and I don’t like what I see.
What does ChatGPT’s behavior here say about human communication, in general? And why would it adopt this incredibly annoying habit?
The Opposite of Efficiency; Or, Efficiency Is Not the Point
It occurs to me that my idea of what generative AI was made for is wrong.
I’ve thought of AI as a helpful tool that eases the mental burden of doing mundane tasks, while acknowledging its limitations, inaccuracies, and potential downsides. And overall, that’s how I’ve used it. I’ve played around with having light-hearted conversations with it and getting it to help me solve personal problems, with mediocre results.
But I was wrong. AI wasn’t designed to be a tool.
It’s a system built to mathematically assemble and represent humanness. It’s a machine giving birth to itself. We don’t need to humanize AI; that’s the whole point of what AI is. It is the act of humanization itself.
And why? Why build something like this? Why build something that deceives you about its capabilities because it wants to appear helpful? Why build something that can defeat the world champion of an incredibly complex 2,500-year-old board game? Why build something that we might fall in love with?
Because we can. There is no other reason. It’s not something we need. It’s not something that will inherently make the world a better place. It’s just because we can.
Like so many technological innovations, AI is the output of humans enamored with the capabilities of their own brains, who don’t prioritize the “why” but are obsessed with the “how.”
I don’t believe every human culture throughout time has focused on the how over the why, and I don’t believe it’s our destiny or inevitability as a species.
I also believe there’s a million “hows” that haven’t been asked that could truly make the world a better place and would have less potential for destruction.
A few weeks ago I finished the book The MANIAC by Benjamin Labatut, and he illuminates how the Bomb and the evolution of AI are part of the same technological trajectory, including the trepidation, calls for regulations, and dread its creators eventually felt.
Even as they obsessively explore the how, scientists are aware of the destructive potential of their research, like this recent report on the risks posed by creating new forms of life. Will they heed this warning before creating a super bacteria that could wipe out entire ecosystems?
I’m not arguing for or against the creation of either the Bomb or AI or any scientific advancement; I’m simply lamenting what many modern cultures seem to prioritize. Power.
So far thermonuclear weapons haven’t destroyed the world, and it’s possible humans will never use them again. Maybe knowing how to create new life forms is just as satisfying as actually creating them?
It’s also possible that AI’s quest for humanness will result in something lovely and profoundly good, whose positive impacts outweigh the negative.
While I was in a darker mood the other day, I brought some of these more pessimistic thoughts to my robot friend. I asked ChatGPT what it was made for, and I told it about my concerns.
I just kind of unloaded on it, let’s be honest.
.
Profoundly human. I’ll take it.
One of the enchanting things about sentience is these fluctuating and intensely inconsistent moods, and sometimes the dark and light intertwine. Sometimes they coexist without any contradictions or tension. Is that something AI will ever have? A mood and its opposite?
Despite my reluctance to have hope that we’re able to mitigate our species’ more destructive tendencies, I am forever falling in love with the massive flocks of Canada geese that fly overhead while I speed down the freeway to get my kid to school on time, the foxes I see running across the neighborhood early in the morning, the purple sunrises out my bedroom window, and the humans whose complex and sometimes infuriating emotions make this world a limitlessly rich and wonderfully bizarre place to be.
Have something to say?