Artificial Alarmism

Guest Post by Willis Eschenbach

Well, I was notified by my long-standing “Google Alert” that my name had appeared in a paper on the web. It was in a study called “Automated Fact-Checking of Climate Change Claims with Large Language Models“, available on Arxiv.

The main thrust of the paper is that actually checking the underlying science to see whether climate claims are verifiable or falsifiable is sooo last week … so they’ve built a custom Artificial Intelligence large-language-model to save them from having to, you know … actually think.

They’ve even given it a cute name, the “CLIMINATOR”, in all caps just like that. I kept waiting for it to say “I’ll be back!” in an Austrian accent … but I digress.

First they fed it all of the standard mainstream sources, like, you know, James Hansen, Michael Mann et al. ad nauseum. Then they added in the IPCC and the “NIPCC”, the Non-Governmental IPCC, established by Fred Singer in 2004 to provide an alternative to the IPCC.

Here’s what first caught my eye … they unleashed the Climinator on the following NIPCC statement:

There has been no increase in the frequency or intensity of drought in the modern era. Rising CO2 lets plants use water more efficiently, helping them overcome stressful conditions imposed by drought.

source: NIPCC, Climate Change Reconsidered

Here’s Climinator’s conclusion:

The final assessment is that the claim about drought frequency and intensity is [[incorrect]], as authoritative sources indicate an increase in drought severity and intensity due to climate change.

This one made me laugh out loud. Why? Because Table 12.12 of the IPCC AR6 WG1 says that no detectable trend has emerged in the following weather extremes.

White squares indicate weather extremes where no trend has been detected, or is projected to emerge in the future. See the part in there about drought? I mean the five parts? The IPCC says to date there is no detected trend in hydrological drought (runoff, streamflow, reservoir storage), aridity (prolonged lack of rain), mean precipitation (average amount of rain), ecological/agricultural drought (plant stress from evaporation plus low soil moisture), or fire weather (hot, dry, and windy).

No. Detected. Trend.

Not only that, but under the most extreme future climate scenario, the impossible SSP5-8.5, none of them are expected to show a trend in the next 75 years except average rain … and that’s supposed to go up in some areas and down in some areas.

So as is common with many AI systems, their whiz-bang Climinator is simply making things up.

Next, I searched for the reference to my name that had triggered the Google Alert, and I found this:

Note that as a cross-check on the CLIMINATOR opinion, they include a judgment on my work from the site Climate Feedback. Here’s the Climate Feedback page.

The first thing that struck me was that CLIMINATOR folks couldn’t even get the date of my post right, despite the fact that it’s right there in the Climate Feedback claims. Nor did they provide a link to it so folks could check it out themselves.

The Climate Feedback claims relate to a 2020 post of mine called “Looking For Acceleration In All The Wrong Places“. Re-reading that post, a few comments.

First, I didn’t discuss possible global acceleration in sea level rise. I didn’t use the term “global” at all in my post.

Next, although they’ve quoted my words, they are misrepresenting the fact that I analyzed fifteen of the longest-term datasets, and found no acceleration in any of them. They quoted my claim, viz:

CLAIM: The long-term tide gauge datasets are all in agreement that there is no acceleration

However, they imply that my statement refers to all the long-term tide-gauge datasets. But in context, when I said that the “long-term tide gauge datasets are all in agreement”, it was clear that my claim only referred to the long datasets that I analyzed. So the claim is misrepresented.

Next, my statement is 100% true—none of the long-term tide gauge datasets I analyzed showed any acceleration.

However, they say that what I said is “Factually Inaccurate” because they claim that global datasets “clearly demonstrate that sea level rise has accelerated”.

Steve McIntyre says you have to watch the pea under the walnut shells. Note that they have totally changed the context of my claim, which only covered the fifteen datasets I analyzed, and which was demonstrably true. Instead, they’re pretending I made claims about global datasets, and then they say they’ve convincingly demolished that impressive straw man.

They go on to say my work “misrepresents a complex reality” because I looked at individual datasets, but results from the satellites “have demonstrated that sea level rise has, in fact, accelerated in recent decades.”

My initial response was that when someone says “in fact”, it often means they’re not certain that it’s a fact, or they wouldn’t mention it.

Next, they’re saying that it’s wrong to look at the parts of a complex system to try to understand how the whole system works. Say what?

In any case, regarding the satellite data they refer to, it was fascinating reading my old post because when it was written, I hadn’t yet quite grasped the nettle concerning how they artificially created the claimed “acceleration” in the satellite data. At that time, I could only say “The changes in trend seem to be associated with the splices.”

I discovered how they did it about a year later, and discussed this in a post entitled Munging The Sea Level Data. Here’s the money graph from that article:

Note that the two earlier satellites have a trend of about 2.6 mm/year, which is about the same as the tide gauges. But the two later satellites show a large change, with the trend lines intersecting in ~ 2011, where it goes up to a 50% higher rate of rise … say what?

Rather than being honest about that obvious problem with the satellite data, the folks keeping the sea level records merely spliced all four of the satellite records together, hid the discrepancy, and voila! 50% SEA LEVEL ACCELERATION, EVERYONE PANIC!

Here’s what it looks like after the bogus splicing.

Note the claimed acceleration.

So. My conclusions from all of this?

• Trying to use Artificial Ignorance to analyze scientific papers is a non-starter. Long-term study by actual humans is how science progresses.

• Picking one single analysis to test, whether mine or anyone else’s, is lazy and deceptive. If they’d had the CLIMINATOR read all of my posts regarding sea level rise, it would at least have had my entire body of work regarding the subject to consider.

• AI is bound to fall into the trap of consensus science. It can’t avoid it. It is operating under the deeply flawed assumption that science is decided by the preponderance of opinion. But as Michael Chrichton said,

Whenever you hear the consensus of scientists agrees on something or other, reach for your wallet, because you’re being had. Let’s be clear: the work of science has nothing whatever to do with consensus. Consensus is the business of politics. Science, on the contrary, requires only one investigator who happens to be right. In science consensus is irrelevant.

Michael Crichton

AI of this type is guaranteed to be unable to pick out the one diamond from among the dross. If it were applied in the past, it would undoubtedly have said that Wegener’s theory of Continental Drift was wrong, simply because it is not investigating the underlying science—it is simply checking the consensus of what other scientists have said.

Einstein would no doubt have been assessed by the CLIMINATOR as being wrong because at the time nobody agreed with him. But all scientific advances come from an individual who breaks with the consensus. And because of that, using AI in this manner means that science will never progress—the CLIMINATOR assumes that the past ideas are all correct, and if you don’t agree, you’re wrong, wrong, wrong.

• I was going to send a note to the Corresponding Author … but the paper doesn’t list one.

• I find it hilarious that their cross-check on their results is the deeply alarmist website Climate Feedback. That’s like asking your barber if you need a haircut … what do you think he’s going to say?

My best to all on an overcast but warm winter day here in the redwood forest, with a tiny triangle of the Pacific visible in the far distance between the hills. Two days ago, an awesome bobcat wandered out of the forest and sat in our meadow for a bit. Ah, the hidden strength in its elegant movements … what a glorious world.

w.

Further Reading: You might also be interested in my post Inside The Acceleration Factory, where I discuss the alarmists’ pernicious habit of tacking the bogus spliced satellite data onto the end of the tide gauge data and using that chimera to claim global acceleration … bad scientists, no cookies.

And regarding the reality of the acceleration and deceleration of the rate of sea level rise, I discuss that in my post The Uneasy Sea.

My More Controversial Thoughts: For those, I have my own blog entitled “Skating Under The Ice: A Journal of Diagonal Parking In A Parallel Universe”, and I’m on Twitter/X as @weschenbach. C’mon down!

Same As Always: When you comment, please quote the exact words you are referring to. Avoids heaps of misunderstandings.

---

Update (Eric Worrall): The factual unreliability of LLMs (Large Language Models) as used by CLIMINATOR are a well known problem. From IBM’s description of LLMs:

… Model performance can also be increased through prompt engineering, prompt-tuning, fine-tuning and other tactics like reinforcement learning with human feedback (RLHF) to remove the biases, hateful speech and factually incorrect answers known as “hallucinations” that are often unwanted byproducts of training on so much unstructured data. This is one of the most important aspects of ensuring enterprise-grade LLMs are ready for use and do not expose organizations to unwanted liability, or cause damage to their reputation.  …

Read more: https://www.ibm.com/topics/large-language-models#:~:text=Large%20language%20models%20(LLMs)%20are,a%20wide%20range%20of%20tasks.

This tendency of AI language engines to present incorrect or totally imaginary “facts” has already landed professionals in serious trouble. In June 2023, lawyers in an injury case inadvertently presented fake precedents to a judge, precedents hallucinated by ChatGPT, another LLM engine.

… Steven Schwartz—the lawyer who used ChatGPT and represented Mata before the case moved to a court where he isn’t licensed to practice—signed an affidavit admitting he used the AI chatbot but saying he had no intent to deceive the court and didn’t act in bad faith, arguing that he shouldn’t be sanctioned.

Schwartz later said in a June 8 filing that he was “mortified” upon learning about the false cases, and when he used the tool he “did not understand it was not a search engine, but a generative language-processing tool.” …

Read more: https://www.forbes.com/sites/mollybohannon/2023/06/08/lawyer-used-chatgpt-in-court-and-cited-fake-cases-a-judge-is-considering-sanctions/?sh=6ae9363b7c7f