The Danger of Anthropomorphizing

Not everything should be described as if it had human imperatives

Nov 11, 2024

Is this car smiling at me, or am I just anthropomorphizing?

It’s human nature to see humanity in everything. In the behavior of animals, for example. Or seeing a smiling face in the headlights and grille of a car. Mostly this human trait is harmless and endearing. We see humanity in the way our dogs look at us, and we read meaning into every expression our children share with us.

When does seeing humanity in everything lead us astray?

There are times when the anthropomorphic view is misleading - particularly with respect to AI, or Artificial Intelligence. You expect to see soft-news and common-interest shows cover AI this way - with some mystique and breathless descriptions of the computer “doing the thinking for me.” But this is for the consumption of the masses, and relayed by people who frankly don’t understand the technology (I can still hear that Good Morning America show where they tried to describe what the Internet was back in… 1995? 1996?… it was dreadful, but exactly how you’d expect mainstream media to misunderstand new technologies).

Hey, Good Morning America is at it again with AI! (apologies I don’t have a good way to embed this one).

What alarms me is when people who should know better engage in the same kind of sophistry. Increasingly, AI startups and their key staff and executives are using such “human” terms to describe what their AI algorithms are doing: anthropomorphizing. It isn’t healthy, nor does it accurately describe what their various AI products do.

When you hear these executives talk about their AI algorithms’ “reasoning” or “thinking”, you know they’ve crossed the line into fiction. They might defend the behavior as using the same word with a different meaning. As an example we would say we “call” an API (Application Public Interface) - but there is no phone call placed or audible sound transmitted in modern applications. But I would argue that that is different - in that everyone understands that the use of the word “call” means something different.

Why is conflating what LLMs do with reasoning and thinking problematic? First, because even these latest LLM techniques are not examples of “reasoning”, and the experts in the field know it. These LLMs simply generate tokens based on probabilities in their training data set and the guidance of the prompts. Second, because it misrepresents what an LLM (or other AI algorithm) is doing behind the scenes. It invokes thoughtfulness and intent, where there is none. An LLM does not “consider” an answer, it “produces” (as in production) and answer. Third, we create fear that these LLMs are a concrete step toward general intelligence.

What does the research say?

Don’t take my word for it. Some researchers from Apple have gone ahead to prove the obvious, and published their results in this paper. They do not come across as LLM-skeptics so much as wanting to define what LLMs are capable of doing:

Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, including natural language processing, question answering, and creative tasks (Gunter et al., 2024; OpenAI, 2023; Dubey et al., 2024; Anil et al., 2023; Abdin et al., 2024; Rivière et al., 2024). Their potential to perform complex reasoning tasks, particularly in coding and mathematics, has garnered significant attention from researchers and practitioners. However, the question of whether current LLMs are genuinely capable of true logical reasoning remains an important research focus. While some studies highlight impressive capabilities, a closer examination reveals substantial limitations. Literature suggests that the reasoning process in LLMs is probabilistic pattern-matching rather than formal reasoning (Jiang et al., 2024). Although LLMs can match more abstract reasoning patterns, they fall short of true logical reasoning.

As the authors point out, introducing additional irrelevant input tokens can drastically alter model outputs, and that the likelihood of focusing on the right inputs decreases exponentially with an increase in the number of tokens. Using only questions on an 8th grade math test standard, the researches proceed to prove that LLMs are not reasoning how to solve the problem, they are only automating the production of the most likely output tokens, which may or may not represent a correct answer.

AI is software Automation, not Intelligence

The simplest way to not be fooled by the idea that AI is thinking for you - is to simply substitute the word automation. The LLM isn’t reasoning the answer, it is automating the production of a probable answer based on its ability to pattern match its existing training data.

Think about it this way. The LLM isn’t reasoning the answers to your law exam. It is instead automating the process of selecting most probable tokens (answers) from its canon of training data. If law exam questions and test results are in that training data, it is likely to do well on the test. If the laws called into question by the exam are in its corpus of training data it is also likely to perform well on the test.

LLMs are amazing. They simulate language production and art production that humans would do - or exceed human production in certain dimensions. They are, in effect, automating the production of words, paragraphs, and outlines. They are, in effect, automating the generation of images based on a description. LLMs can automate the revision of words in a particular style, or images in a particular style. The fact that it is automation doesn’t take away from how amazing the technique is nor from the amazing results it can produce.

The techniques of LLM automation are relatively new, but this kind of use case for automation is not a new. idea. For example, we’ve had filters you could apply to photos that would render them to look like a comic book illustration. Or a charcoal sketch. LLMs take a different tack but the use case is very similar. Today you can have an LLM apply Shakespeare’s style to your work email, but some 20 years ago we had the Snoop Dogg Shizzlelator which would automate making it sound like something Snoop Dogg would say.

What do we actually need to worry about?

There’s a lot of concern about the development of general artificial intelligence. And many felt like LLMs were the key to unlock a generally intelligent AI. It hasn’t worked out that way (as we can see from the research above). When discussing autonomous driving a year ago, and why it didn’t have mass adoption yet, I made a joke that feels like it has some truth:

“I think General Intelligence comes before fully autonomous driving, because it is probably an easier problem to solve. After all, we have billions of generally intelligent people and a lot of them can’t drive”

There are a few reasons why we jump to conclusions like this too soon:

We anthropomorphize AI tools and inanimate things way too easily
New AI techniques are super narrow in scope. They aren’t “general” capabilities - they’re super specific capabilities.
LLMs, in particular, are good at simulating human production of words and phrases, which tends to trick us into thinking it does more than it really does. I’ll note that no one seems confused that LLMs ability to generate images and videos is *actual* intelligence or creativity - we all accept that that is rather a form of automation and probability.

I would propose that it is the simulation of human intelligence that we should worry about. We have almost always worried about AI reaching “the singularity” as the event to worry about because this kind of AI could surpass us. But the simulation of human intelligence and communication - and the simulation of empathy and emotion and anger - may well be enough to cause humanity to destroy itself.

The real damage to humanity will be done by our fellow humans.

We already know that humans are vulnerable to The Big Lie (see: WWII, we can ignore for now the more recent examples around election integrity). We know that propaganda delivered the right way shapes our opinions. We can see that when a Democrat is president, Republicans think the economy is doing poorly - and this shift happens within days (not nearly long enough to account for the economy changing). And we can see that when a Republic is president, their impression of the economy shifts positively within days - and Democrats’ view of the economy begins to decline. The fact set is the same, but our views of it change based on which group we affiliate ourselves with.

So we know humans can be manipulated to do terrible things to each other, to dislike each other, to riot, and to start wars with each other. The only question is when will AI be sufficiently skilled to automate all of that work for us and start wars for us, without all the hard work. Sadly, I believe we have already reached that point. The real damage to humanity will be done by our fellow humans.

We will see groups of people as either “in” or “out” with respect to our own affiliations, and increasingly out groups will be pushed further away, their humanity criticized and undermined, until we no longer have empathy when something bad befalls them. LLMs will speed this along and make Fox News look slow by comparison. LLMs can already help bad actors today, to propagate the memes that propel this sort of division forward. I won’t be surprised when LLMs are used to train “Lone Wolf” terrorists to act out. I think this future is grim if we don’t get a grip on it. General Intelligence may be a fear - but imitating human interaction is already the threshold for how we can destroy ourselves.

When I was in college, I took a class “AI applications in Prolog” from Yoav Shoham - and in that class he illustrated for us how we can simulate intelligent behavior by layering very simplistic motives or imperatives on a “robot”. This was demonstrated with a robot that had limited computer vision capabilities (it was 1993 after all). Its first motive was to maximize its distance from any objects it could see and remember. Its second motive was to get from A to B. When moving through a room it would glide to the open spaces, and if a person approached it, it would skitter away to create equal spacing. It wasn’t “intelligent” but it was rather like a nervous animal keeping a safe distance.

Yoav had a great point - intelligence isn’t *just* the high level cognition that we associate with humans - it is all of the layers AND the cognition. We must be able to walk and talk and think and listen - all at the same time. It also taught me that we as humans will interpret actions that simulate intelligence to be intelligecnce and we will act accordingly.

Is it possible to use LLMs for good? Of course. There are plans to use LLMs in therapy situations to try to help people who are mentally ill, perhaps LLMs will be key to deprogram people who have been brainwashed to be terrorists. I don’t know the use cases, but I know that LLMs can be trained on these interactions over time, and learn to be more persuasive and more convincing through sophisticated scoring and probability, and feedback loops.

In conclusion, I’d ask that we spend less time talking about machines and programs as “thinking” or “reasoning” and more time talking about what activities they will automate, and what that might mean for our lifestyle, jobs, economy, education.

Thank you for reading - please sign up below to get this in your inbox every week (or so) as I work on writing about the things that are on my mind. If you disagree with the above I’m interested in your take - and your insights.

Turning Coffee into Process

Discussion about this post