AI Chatbots on Causality

Having recently encountered some major causality train wrecks, I got curious about LLM “understanding” of causality. If AI chatbots are trained on the web corpus, and the web doesn’t “get” causality, there’s no reason to think that AI will make sense either.

TLDR; ChatGPT and Bing utterly fail this test, for reasons that are evident in Google Bard’s surprisingly smart answer.

ChatGPT: FAIL

Bing: FAIL

Google Bard: PASS

Google gets strong marks for mentioning a bunch of reasons to expect that we might not find a correlation, even though x is known to cause y. I’d probably only give it a B+, because it neglected integration and feedback, but it’s a good answer that properly raises lots of doubts about simplistic views of causality.

AI doesn’t help modelers

Large language model AI doesn’t help with modeling. At least, that’s my experience so far.


DALL-E images from Bing image creator.

On the ACM blog, Bertrand Meyer argues that AI doesn’t help programmers either. I think his reasons are very much compatible with what I found attempting to get ChatGPT to discuss dynamics:

Here is my experience so far. As a programmer, I know where to go to solve a problem. But I am fallible; I would love to have an assistant who keeps me in check, alerting me to pitfalls and correcting me when I err. A effective pair-programmer. But that is not what I get. Instead, I have the equivalent of a cocky graduate student, smart and widely read, also polite and quick to apologize, but thoroughly, invariably, sloppy and unreliable. I have little use for such  supposed help.

He goes on to illustrate by coding a binary search. The conversation is strongly reminiscent of our attempt to get ChatGPT to model jumping through the moon.

And then I stopped.

Not that I had succumbed to the flattery. In fact, I would have no idea where to go next. What use do I have for a sloppy assistant? I can be sloppy just by myself, thanks, and an assistant who is even more sloppy than I is not welcome. The basic quality that I would expect from a supposedly intelligent  assistant—any other is insignificant in comparison —is to be right.

It is also the only quality that the ChatGPT class of automated assistants cannot promise.

I think the fundamental problem is that LLMs aren’t “reasoning” about dynamics per se (though I used the word in my previous posts). What they know is derived from the training corpus, and there’s no reason to think that it reflects a solid understanding of dynamic systems. In fact there are presumably lots of examples in the corpus of failures to reason correctly about dynamic causality, even in the scientific literature.

This is similar to the reason AI image creators hallucinate legs and fingers: they know what the parts look like, but they don’t know how the parts work together to make the whole.

To paraphrase Meyer, LLM AI is the equivalent of a polite, well-read assistant who lacks an appreciation for complex systems, and aggressively indulges in laundry-list, dead-buffalo thinking about all but the simplest problems. I have no use for that until the situation improves (and there’s certainly hope for that). Worse, the tools are very articulate and confident in their clueless pronouncements, which is a deadly mix of attributes.

Related: On scientific understanding with artificial intelligence | Nature Reviews Physics

ChatGPT struggles with pandemics

I decided to try out a trickier problem on ChatGPT: epidemiology.

This is tougher, because it requires some domain knowledge about terminology as well as some math. R0 itself is a slippery concept. It appears that ChatGPT is essentially equating R0 and the transmission rate; perhaps the result would be different had I used a different concept like force of infection.

Notice how ChatGPT is partly responding to my prodding, but stubbornly refuses to give up on the idea that the transmission rate needs to be less than R0, even though the two are not comparable.

Well, we got there in the end.

ChatGPT and the Department Store Problem

Continuing with the theme, I tried the department store problem out on ChatGPT. This is a common test of stock-flow reasoning, in which participants assess the peak stock of people in a store from data on the inflow and outflow.

I posed a simplified version of the problem:

Interestingly, I had intended to have 6 people enter at 8am, but I made a typo. ChatGPT did a remarkable job of organizing my data into exactly the form I’d doodled in my notebook, but then happily integrated to wind up with -2 people in the store at the end.

This is pretty cool, but it’s interesting that ChatGPT was happy to correct the number of people in the room, without making the corresponding correction to people leaving. That makes the table inconsistent.

We got there in the end, but I think ChatGPT’s enthusiasm for reality checks may be a little weak. Overall though I’d still say this is a pretty good demonstration of stock-flow reasoning. I’d be curious how humans would perform on the same problem.

Can ChatGPT generalize Bathtub Dynamics?

Research indicates that insights about stock-flow management don’t necessarily generalize from one situation to another. People can fill their bathtubs without comprehending the federal debt or COVID prevalence.

ChatGPT struggles a bit with the climate bathtub, so I wondered if it could reason successfully about real bathtubs.

The last sentence is a little tricky, but I think ChatGPT is assuming that the drain might not be at the bottom of the tub. Overall, I’d say the AI nailed this one.

ChatGPT does the Climate Bathtub

Following up on our earlier foray into AI conversations about dynamics, I decided to follow up on ChatGPT’s understanding of bathtub dynamics. First I repeated our earlier question about climate:

This is close, but note that it’s suggesting that a decrease in emissions corresponds with a decrease in concentration. This is not necessarily true in general, due to the importance of emissions relative to removals. ChatGPT seems to recognize the issue, but fails to account for it completely in its answer. My parameter choice turned out to be a little unfortunate, because a 50% reduction in CO2 emissions is fairly close to the boundary between rising and falling CO2 concentrations in the future.

I asked again with a smaller reduction in emissions. This should have an unambiguous effect: emissions would remain above removals, so the CO2 concentration would continue to rise, but at a slower rate.

This time the answer is a little better, but it’s not clear whether “lead to a reduction in the concentration of CO2 in the atmosphere” means a reduction relative to what would have happened otherwise, or relative to today’s concentration. Interestingly, ChatGPT does get that the emissions reduction doesn’t reduce temperature directly; it just slows the rate of increase.

Modeling with ChatGPT

A couple weeks ago my wife started probing ChatGPT’s abilities. An early foray suggested that it didn’t entirely appreciate climate bathtub dynamics. She decided to start with a less controversial topic:

If there was a hole that went through the center of the moon, and I jumped in, how long would it take for me to come out the other side?

Initially, it’s spectacularly wrong. It gets the time-to-distance formula with linear acceleration right, but it has misapplied it. The answer is wrong by orders of magnitude, so it must be making a unit error or something. To us, the error is obvious. The moon is thousands of kilometers across, so how could you possibly traverse it in seconds, with only the moon’s tiny gravity to accelerate you?

At the end here, we ask for the moon’s diameter, because we started a race – I was building a Vensim model and my son was writing down the equations by hand, looking for a closed form solution and (when the integral looked ugly), repeating the calculation in Matlab. ChatGPT proved to be a very quick way to look up things like the diameter of the moon – faster even than googling up the Wikipedia page.

Since it was clear that non-constant acceleration was wrong, we tried to get it to correct. We hoped it would come up with F = m(me)*a = G*m(moon)*m(me)/R^2 and solve that.

Ahh … so the gigantic scale error is from assuming a generic 100-meter hole, rather than a hole all the way through to the other side. Also, 9.8 m/s^2 is Earth’s surface gravity.

Finally, it has arrived at the key concept needed to solve the problem: nonconstant acceleration, a = G*M(moon)/R^2 (where R varies with the jumper’s position in the hole).

Disappointingly, it crashed right at the crucial endpoint, but it’s already done most of the work to lay out the equations and collect the mass, radius and gravitational constant needed. It’s still stubbornly applying the constant acceleration formula at the end, but I must say that we were pretty impressed at this point.

In the same time, the Vensim model was nearly done, with a bit of assistance on the input numbers from Chat GPT. There were initially a few glitches, like forgetting to reverse the sign of the gravitational force at the center of the moon. But once it worked, it was easily extensible to variations in planet size, starting above or below the surface, etc. Puzzlingly the hand calculation was yielding a different answer (some kind of trivial hand computation error), but Matlab agreed with Vensim. Matlab was faster to code, but less interactive, and less safe because it didn’t permit checking units.

I’d hesitate to call this a success for the AI. It was a useful adjunct to a modeler who knew what they were doing. It was impressively fast at laying out the structure of the problem. But it was even faster at blurting out the wrong answer with an air of confidence. I would not want to fly in a plane designed by ChatGPT yet. To be fair, the system isn’t really designed to do physics, but a lot of reasoning about things like the economy or COVID requires some skills that it apparently doesn’t yet have.

Can AI do SD?

There are lots of interesting developments in machine learning for time series modeling, but can an AI write about systems sensibly? I decided to try a few free online AI content generators.

Using system dynamics modeling for project delay and disruption

1. System Dynamics Modeling (SDM)

System dynamics modeling is a mathematical model that helps us understand how systems work and how they interact with each other. SDM is useful for understanding complex systems and their behavior over time. In addition, it can help predict future outcomes based on current conditions.

2. Project Delay & Disruption

Project delays and disruptions occur frequently in projects. These events cause significant problems for organizations and individuals alike. A project delay occurs when a project is delayed past its original completion date. A project disruption occurs when a project is disrupted before it reaches its original completion date. Both types of project delays and disruptions can have negative effects on the success of a project.

3. Causes of Project Delays & Disruptions

There are many causes of project delays and disruptions. Some of these causes include:

• Poor planning

• Unclear requirements

• Lack of communication

• Poor management

• Miscommunication

• Poor leadership

• Poor team coordination

• Poor resource allocation

Well, we didn’t get the rework cycle or any feedback effects, but not an entirely useless effort.

Do climate tipping points signal the onset of Limits to Growth?

Climate change is a serious human problem that causes concern among many people. Some believe that climate change is caused by human activity and that it will bring about a catastrophic future. However, others believe that it is natural and that there are ways to prevent it from causing further damage. Regardless of whether climate change is caused by human activity or not, it has the potential to cause major damage to the planet. Many believe that we are approaching climate tipping points, which could have catastrophic consequences for humanity.

One reason why climate tipping points signal the onset of limits to growth is because they are a harbinger of future environmental disasters. These disasters would occur because of the way we have altered the climate and what we have done to the planet. For example, the Larsen B ice shelf in Antarctica broke away in 2002 and disintegrated within four years due to global warming. Climate tipping points signal the onset of limits to growth because they are a harbinger of future environmental disasters. Governments and environmental groups should prepare for future environmental disasters by investing in renewable resources and making changes in agriculture to reduce food insecurity.

Another reason why climate tipping points signal the onset of limits to growth is because they are a harbinger of future economic problems. Scientists have warned us about manmade climate change for years; however, most people ignore their warnings and keep on using fossil fuels anyway. This is causing major problems for our planet, including droughts, rising sea levels and more frequent natural disasters. Unfortunately, many people are paying little attention to the warnings; they think they can ignore them and avoid problems later. Governments and environmental groups should prepare for future economic problems by investing in renewable resources and making changes in agriculture to reduce food insecurity.

I have to say, the AI is making more sense than the South Dakota legislature.

Should hunters try to harvest more deer to help slow the growth of chronic wasting disease? Or is the cure worse than the disease, so we are merely killing deer now to avoid them dying later?

This is a case of a deer that had been found dead by an early hunter, then found by another hunter on the same day. These hunters were in the woods for several days, hunting deer. Then, for a week or more, the deer had eaten enough grass to cover itself, leaving the deer lying in its own feces. In the United States, this is called wasting. Chronic wasting is an infectious disease that, in humans, is caused by bacteria. When a person is infected with chronic wasting, they begin losing weight and become weak. They begin to smell bad, have sores on their body, lose their appetite, feel tired, listless and weak, having trouble standing up. Some people lose consciousness. Their physical…

I guess we still need modelers for some things.

Detecting the inconsistency of BS

DARPA put out a request for a BS detector for science. I responded with a strategy for combining the results of multiple models (using Rahmandad, Jalali & Paynabar’s generalized meta-analysis with some supporting infrastructure like data archiving) to establish whether new findings are consistent with an existing body of knowledge.

DARPA didn’t bite. I have no idea why, but could speculate from the RFC that they had in mind something more like a big data approach that would use text analysis to evaluate claims. Hopefully not, because a text-only approach will have limited power. Here’s why.

Continue reading “Detecting the inconsistency of BS”