Today I did a Google search for the term "JUNK DNA" and, as usual, the first thing I saw was the Google AI description of junk DNA. It's wrong, but that's not the scary part. The most frightening thing about the AI description is that it promotes three videos that misrepresent science and two of them are from well known kooks.
What does this tell you about current versions of AI? It tells you that it is not intelligent in any meaningful sense of the word. It tells you that Google AI is incapable of distinguishing between scientific facts and ignorance. It tilts toward the loudest voices on the internet and, as we all know, those voices are frequently wrong.
Here's what I saw ...
"Junk DNA" is a historical term for non-coding DNA, which was once thought useless, but is now known to have crucial roles in regulating gene activity, structuring the genome, and influencing development, with much of it (around 98% in humans) involved in complex biological functions, though some fragments (pseudogenes, viral remnants) may still be functionally neutral.
This video explains how the term 'junk DNA' originated and why scientists now believe it has important functions:
This is a video by the well-known kook, Bruce Lipton. Read the Wikipedia article to see who I'm talking about. Apparently Google AI is incapable of checking its sources.
What It Was Thought To Be
Non-Coding: The term emerged because only about 2% of our DNA codes for proteins, the rest was deemed "junk".
Evolutionary Remnants: It included old viral sequences, transposons (jumping genes), and pseudogenes (broken gene copies) that didn't seem to do anything beneficial.
Watch this video to understand the origins of the 'junk DNA' theory and why it is now considered a myth:
This is a video from the Discovery Institute, a creationist organization. I've already explained why it is wrong [Why do Intelligent Design Creationists still lie about junk DNA?]. The important point is that Google AI cannot tell the difference between good science and bad science.
What We Now Know It Does (Functions)
- Gene Regulation: Acts as enhancers and switches, controlling when and how much other genes are turned on or off, even from far away.
- Genome Structure: Helps fold DNA into specific 3D shapes, crucial for proper cell function.
- Cellular Processes: Regulates cell development, differentiation, and response to environmental cues.
- Evolutionary Role: Transposons, once "junk," have been co-opted for vital developmental roles.
This video provides a detailed look into the important functions of what was once called 'junk DNA':
This is a video published by PBS. It's slightly better than the first two videos but still full of misinformation about junk DNA and the history of genomics.
Why the Term is Problematic
- Misleading: The term "junk" downplays its importance, as many regions once labeled as such have essential functions.
- Context-Dependent: Some DNA might be neutral (neither helpful nor harmful) in one context (like "Spam DNA"), but "junk" implies a total lack of purpose, which is often incorrect.
In Summary: "Junk DNA" is largely a historical label for the vast amount of non-protein-coding DNA that scientists are now discovering plays critical, complex roles in our biology, from orchestrating gene expression to shaping the genome's structure.
The good news is that the current versions of AI are not very intelligent so it's not going to take the place of knowledgeable scientists.
The bad news is that there are many AI advocates who are saying that AI is already more intelligent that the average high school graduate and it won't be long before it's smarter that someone with a Ph.D. Because of the hype, more and more people are beginning to believe that AI is actually intelligent. It will be a disaster if the average person thinks that Google AI, for example, is actually giving them accurate information.
11 comments :
Bruce Lipton has an entry in the online Encyclopedia of American Loons, he's #1899.
-César
Larry wrote: "The important point is that Google AI cannot tell the difference between good science and bad science."
There is no absolute dichotomy between good and bad science. Since there are many aspects of what makes 'good science', the difference is gradual. There is continuity between good and bad science.
As I pointed out in Appendix 1 of my review of your book
https://korthof.blogspot.com/2023/06/scientists-say-90-of-your-genome-is.html
in the top science journals Science and Nature, there are many anti junk DNA articles. So, the dividing line between 'good' and 'bad' science runs right through the top scientific journals. If that is true, how can you blame AI?
It seems to me that the biggest misconception about junk DNA is that it applies to *categories* of DNA, like all transposons, ERVs, or even all noncoding DNA, rather than specific instances. That idea, along with the second biggest misconception that natural selection would necessarily purge all nonfunctional sequences from a genome, seems to be the basis of the popular strawman of junk DNA that is alarmingly ubiquitous even among actual biologists. Creationists just swap the existing misconception about natural selection for the idea that God must have created a perfect genome to begin with.
I think a big part of the problem is a pattern I've noticed where even actual journalists and scientists, along with creationists, are incentived to push the idea that scientists were hopelessly clueless or confused about some concept in order to promote their clickbait article or their new paper to an increasingly anti-intellectual and anti-academia audience. They're effectively throwing the rest of the scientific community under the bus for their own sake, and then people further popularize these ideas to stick it to the "ivory tower academics". These are almost always things that scientists actually understood well or at least had reasonable disagreements about, which the article just strawmans.
I just now tested this with Grok as an alternative AI source. The result was NOT any better!
@gert korthof : It's true that a lot of bad science gets published in leading journals like Science and Nature. It's even more common to see bad science promoted in press releases and science articles aimed at the general public.
What we need is better peer review in order to eliminate bad science from the journals. Meanwhile, we need good science writers who can sift through the scientific literature and recognize what's good and what's bad. Barring that, we can hope that decent science writers can at least recognize what's controversial and what is well established.
I expect that a good artificial intelligence program should be able to do a better job than the average popular science writer. I expect it to, at the very least, be able to recognize the existence of good science and take it into account.
In this case, there are many publications in the scientific literature defending junk DNA and providing a good definition of junk. There's also a Wikipedia article on junk DNA that could have been consulted. Google AI ignored all the evidence for junk DNA and accepted without question some videos from very questionable sources.
None of that is meant to excuse the behavior of scientists who publish bad science. They should lose their grants. And I feel just as strongly about the leading journals such as Nature and Science. I think Nature should apologize for the ENCODE publicity disaster and Science should apologize for giving Elizabeth Pennisi a forum to spread misinformation.
Larry, I agree. But you shouldn't generalise, you should not lump all AI applications together. AI in the hands of qualified scientists delivers valuable and unique results. For example: AlphaFold and several others.
@gert korthof: That depends on how you define intelligence as in artificial "intelligence."
AlphaFold is a sophisticated algorithm that builds on secondary structure predictions that were developed 50 years ago. It only works because of the many thousands of 3D structures that were deposited in public databases since the late 1970s. The algorithm was created by humans to compare amino acid sequences to those structures much faster than any human could do.
Larry, you don't seem very impressed by the achievements of AI such as AlphaFold. Your opinion contradicts the following facts: "John J. Hopfield and Geoffrey E. Hinton were awarded the 2024 Nobel Prize in Physics for developing machine learning technology using artificial neural networks. In Chemistry it was awarded to Demis Hassabis and John M. Jumper for developing an AI algorithm that solved the 50-year protein structure prediction challenge. This highlights AI’s impact on science, medicine and society;"
https://www.nature.com/articles/s41746-024-01345-9
and:
The Nobel Prize in Chemistry went to an AI model (and rightly so)
https://www.universiteitleiden.nl/en/news/2024/10/the-nobel-prize-in-chemistry-went-to-an-ai-model-and-rightly-so
It seems that so many AI "failures" result from using the models as single ask-response systems (which Google search results encourage) rather than as a dialog. I asked ChatGPT 5.2 this question, but kind of warned it by mentioning Larry's name. The response was kind of somewhere in the middle. I then asked it why LLM responses are so biased toward the wrong answer. It's response was quite good, explaining why the term "junk DNA" makes such a reliable LLM pothole: "massive asymmetry in textual volume versus epistemic weight".
Essentially, the topic is kind of uninteresting to most scientists, and just kind of there as a fact in the background. As ChatGPT says: "Nobody writes manifestos titled 'Still Mostly Junk, Guys'." But, "...junk DNA is existentially inconvenient for special creation." So, there is a mass volume of noise compared to real information.
The end result is: "From an LLM perspective, this is toxic in a very specific way..." But, the modern versions _are_ capable of reasoning about it, but they have to be used as intended, interactively, to explore the answers and reasons behind them.
The above comment is mine. It seems that selecting "Comment as: Google account" doesn't actually work, and it comes out as anonymous.
@gert korthof: I know Geoffrey Hinton. We even worked together briefly in the 1990s in order to develop a bioinformatics program in our two departments (Biochemistry and Computer Science). I'm familiar with machine learning algorithms and all the advances in bioinformatics that have been made in my field.
This includes AlphaFold and all the programs that preceded it over the past 50 years. I know people who worked on predicting alpha helices and beta structures in the early 1970s.
I don't consider that "intelligence" in the same way that you do. It's just fast calculations using large databases and very sophisticated algorithms written by humans. Most of the advances in bioinformatics were/are driven by the development of better computers and the availability of large amounts of data produced by humans.
We'll know that true artificial intelligence has arrived when the Nobel Committee awards a Nobel Prize to a computer program instead of to smart programmers at Google like Demis Hassabis and John Jumper.
Real AI (i.e intelligent) would be a program that could write a biochemistry textbook, or even a Wikipedia entry, and do a better job than a average informed expert. Real "intelligence" is when a program can examine all available evidence and come up with a consistent model of how something works. It requires a deep understanding of the underlying concepts and the ability to sort the wheat from the chaff.
In the short term, I'm looking forward to the day when we have intelligent AI that can replace all the talking heads on CNN. It won't require any form of super intelligence, just the average intelligence of high school students who have been taught how to think critically.
Post a Comment