On the Dangers of Overused AI Metaphors
Calling language models "bullshit generators" and people who want AI regulation "modern Luddites" is bad for everyone
In our attempts to understand the new from the old and the unknown from the known, we risk either stripping away too much truth or adding too much falsehood so that our inquiries inevitably become futile.
Metaphors and historical comparisons are always imperfect. They often entail a trade-off between fidelity and simplicity with a touch of motivation to drive an argument forward. Overusing them inadvertently damages the quality of the conversation in a way that affects our understanding and that of others.
This discoursive illness pains AI today.
Successful metaphors as signals of partisanship
AI metaphors fall victim to a singular phenomenon: death by success.
When I first heard GPT-3 referred to as a “stochastic parrot,” term coined by linguist Emily M. Bender, something clicked for me. It nicely captured one of the most problematic—and idiosyncratic—features of language models (i.e., that they, in contrast to humans, output intention-less pseudorandom utterances). The idea went viral and resonated with a lot of people: anyone could point out the limitations of language models with just a pair of words. A succinct, winning argument on AI debates.
But it’s been two years of seeing it everywhere. The term has been tampered with to the point of emptying it of meaning: the metaphor has eaten the substance within. When I read it now, I realize it doesn’t play the role it was conceived for anymore because its ideological charge impedes it; it’s no longer a pointer to some deep truth about the unreliable nature of language models but a loaded expression that signals the author’s partisanship. Their stance. It’s a symbol—like a flag—not an argument.
The same happens on “the other side” of the AI debate (not everyone is equally prone to exhaust metaphors but I’ve seen it from both parties). For instance, the idea that ChatGPT isn't that different from a human brain because they too are just “prediction machines” or “glorified surprise minimizer[s],” as neuroscientist Erik Hoel puts it, is overused. Like saying language models are “stochastic parrots,” it might have some truth to it but it’s now lost in the narrative war.
This isn’t me cherry-picking. We’re not short of examples: from “ChatGPT is a blurry JPEG of the web” to Stable Diffusion is “automated plagiarism.” And from “ChatGPT is like an e-bike for the mind,” or “like a super human,” to it is “AGI”… I could go on and on. The crux of the matter isn’t the degree to which these analogies diverge from the truth (which can be, indeed, a problem) but the fact that they’re more often used as dialectical weapons to dismiss the opposite party’s ideas—or attract the attention of the ingroups—than to highlight their (limited) validity as arguments.
I’ve seen people throw metaphors at each other in debates on social media as if they were the ultimate attack (not that Twitter is the summit of intellectualism, but still). It may not be problematic given that online conversations tend to be low-quality but it’s very detrimental when a high-profile AI public figure like Sam Altman, OpenAI CEO (1.4M+ Twitter followers), mockingly—and unscientifically—twists the “stochastic parrot” analogy into a debate-ending claim:
Altman’s unfalsifiable statement gives the impression that there’s nothing left to add to the discussion—and oblivious witnesses would certainly believe him as an authority on the matter. Emily M. Bender, unsurprisingly, disagrees with him: “You are not a parrot and a chatbot is not a human.”
Tweets like Altman’s don’t help anyone. They only blur the conversation by tapping onto people's emotions at a rhetorical level. Instead of engaging with the arguments behind the metaphor he dismisses it with a half-joke. And, of course, the “stochastic parrot” expression (or Emily M. Bender) isn’t at fault; this is an inevitable consequence of its success and our reprehensible tendency to reduce the AI discourse to a dogmatic dispute through the abuse of these often well-intended but only-partially-correct analogies.
Historical events as blurry mirrors of the present
As it turns out, AI people love history as much as they love metaphors.
In an exclusive interview for Forbes, Alex Konrad and Kenrick Cai asked Altman this question, “do you see any parallels between where the AI market is today and, say, the emergence of cloud computing, search engines or other technologies?” He replied:
“Look, I think there are always parallels. And then there are always things that are a little bit idiosyncratic. And the mistake that most people make is to talk way too much about the similarities, and not about the very subtle nuances that make them different.”
Nailed it. I disagree with his remark on the “parrotism” of humans, but I agree with him here. Drawing parallelism between AI and previous emerging technologies to support claims in favor—or against—is hardly ever to put it in an objective light but to forward a narrative. For instance, disruptive tech eventually creates more jobs than it destroys (in historical timeframes) so we can expect AI to have the same effect (that was roon and Noah Smith’s main defense of generative AI in an essay eloquently entitled “autocomplete for everything”).
And it’s not just other technologies. Historical events that share some commonalities with the present state of affairs can be easily extended, with some rhetorical sleight of hand, to reinforce any given argument—and it works whether you think AI is a net positive or negative for society.
Some comparisons I've seen alluded to repeatedly: Luddites with people who feel text-to-image models like Stable Diffusion threaten their livelihoods; AI with fire or electricity due to their huge potential to transversely redefine society; the generative AI hype with web3/crypto hype; or AI art’s repercussions on traditional artists with the appearance of photography in the 19th century. And I'm not free of guilt either: just last week I compared the value of learning prompt engineering today with learning English as a non-native speaker when I was younger.
The most recent one I’ve found was from a writer who minimized the relevance of Clarkesworld's temporary closure (due to an unprecedented amount of AI-generated submissions) by drawing an analogy with the emergence of online magazines 30 years ago, “is this really a crisis of creativity? Or an opportunity?” she asked.
Is the emergence of AI writing tools like ChatGPT really comparable to the appearance of digital newspapers? Are people asking for better regulation of generative AI companies behaving like Luddites? A more fine-grained analysis, stripped of eloquent metaphors, would definitely help.
As Altman says, we risk focusing too much on the similarities to the point of unfaithfully projecting historical events into a present with radically distinct circumstances—this rounds the edges of reality to fit our biased perception, biasing in turn the perception of those who listen to or read us.
Don’t fall for the easy argument
I’ve illustrated this issue with analogies, metaphors, and historical comparisons from both sides of the AI debate (i.e. those who think modern AI is nothing more than a powerful statistical tool vs those who think AGI is five years away).
I wanted to emphasize that, even though I think some metaphors are more problematic than others (e.g. comparing ChatGPT with the human brain is pretty far-fetched), I’m not advocating any particular position here (not even my own, which aligns more with the pro-“stochastic parrots” group), but emphasizing a generalized issue that affects us all.
This new trend—which has existed forever but ChatGPT has amplified it by orders of magnitude—might render the whole discourse on AI useless and meaningless: The more we talk about AI through meaning-diluted concepts and dubiously comparable events, the more we alienate those who aren’t familiar with the original intention—that is, most people.
Almost everyone engages in this practice. It goes hand in hand with the popularization of AI and the necessity to communicate with the general public in a language that’s more easily understood, but we can do better. We shouldn’t be talking in terms of “dunking” on each other.
My solution—which I’m not naive enough to think will replace this ubiquitous tendency, but could help to some degree—is to not abuse metaphors and comparisons as standalone arguments. It’s important to contextualize them to make them useful. We should accept the boundaries of applicability before we decide to fully substitute de object of debate with an imperfect replacement.
It’s not practical to do it always, but at least occasionally so that we refresh the meaning those appealing analogies successfully captured that now time and use have erased.
This essay is a good take. I think one reason these AIs are so disconcerting is that the future evolution of AI poses both potentially unbounded downside risk *and* potentially unbounded upside. There's no consensus on which outcome is more likely, and there won't be consensus for a while, which is also unsettling and makes people feel uncomfortable. It's easier to put things--and people--in a box, so the immediate reaction may be to try to do that.
There aren't many risks that fall into this uncertain/unbounded bucket (most risks are clearly asymmetric in one direction or another--with either the downside or upside outcome having higher likelihood--and many are also bounded in magnitude on the upside or downside or both).
"Don’t fall for the easy argument" is good advice. Here is a less simple argument. The system has a limited amount of knowledge that can be accommodated. The number of questions it can generate answers to is infinite. Therefore, most of the possible answers must certainly turn out to be based not on the knowledge gained in the process of training.