AI May Soon Replace Pushover Science Journalists, But Not Analysts of Scientist Truth Claims

 So-called artificial intelligence (AI) is making progress, and quite a few workers are worried that their jobs may be replaced by AI. One type of worker who might reasonably have such worries is the typical person calling himself a science journalist. In recent years the work of those calling themselves science journalists has been largely a kind of formulaic work that AI programs might be able to reproduce. Typical science journalists act like cheerleaders to promote the latest scientific papers.  

To understand the current role of the science journalist, you need to understand the economic motivations behind misleading exaggerations about scientific papers and the uncritical regurgitation and amplification of scientist boasts. In my post "Why the Academia Cyberspace Profit Complex Keeps Giving Misleading Brain Research Reports" I discussed the economic reasons why we keep getting misleading research about brains, and misleading headlines about brain research. The analysis in that post holds true not just for brain research, but for scientific research in general. We live in an economy in which misleading stories about scientific research and groundless but interesting-sounding scientific speculation are highly incentivized, so that there are strong motives for such things. To give a short synopsis of what I discussed at much greater length in that post, the economic motivations are something like this:

(1) Scientists are judged by how many papers they publish and how many citations such papers get.

(2) Because of publication bias (in which papers reporting positive results and particularly interesting-sounding positive results are more likely to be published), scientists are strongly motivated to publish papers claiming positive results and also claiming interesting-sounding results.

(3) Wishing to make themselves appear like sources of important research breakthroughs to help justify their exorbitant tuition, universities are motivated to produce press releases exaggerating the importance of research papers published by their professors.

(4) Since science news is published on web pages with ads that generate revenue for the people running or funding the web pages, with revenue proportional to how interesting-sounding a story is, those running science news web sites or science analysis web sites have an enormous economic motivation to create clickbait headlines that generate higher numbers of page views, and more advertising revenue. Science news sites these days are almost always built in the form of headlines that you must click to read the story; and each time this causes a web page with ads to appear, the people running or funding the site get money from views of the ads displayed on the page you opened up.

The result of all of this is a very wacky world we might call the world of scitainment, to use a word that combines the words "science" and "entertainment." Scitainment is a part of the internet that blends science and entertainment. Very much of what we read in this strange world of scitainment is true, and very much of it is false. The world of scitainment blends fact and fantasy, always trying its best to produce entertaining stories and clickbait headlines. It's all about luring you in to click on the headlines that lead to articles, so that you go to pages that generate ad revenue for the people running the web sites. 

Scitainment is very big business these days.  A vital part of scitainment is what we can call the pushover science journalist.  The pushover science journalist is typically someone ever-ready to promote and hype and sing the praises of low-quality science research.  

The economic ecosystem that incentivizes unfounded boasts about weak-quality scientific research shows no signs of going away any time soon. 

Part of the reason such an ecosystem persists is that researchers have a "wink and a nod" toleration for the lies and exaggerations of university press releases, knowing that such hype increases the likelihood that researchers will get an increase in their cherished goal of higher citation counts. 

But while the work role of "pom-pom" pushover science journalists shows no sign of disappearing, there is a reason to think that trying to build a long-term career working in such a role may be a hazardous bet. The reason is that it may be likely that the role of pushover science journalist can be replaced by AI programs (so-called artificial intelligence). 

Nowadays one of the main features of so-called artificial intelligence programs (AI) are what are called large-language models (LLM).  Such programs are trained on very large bodies of text such as billions of pages that can be accessed by programs roaming the internet. Such programs are able to detect certain types of patterns, and mimic such patterns. 

So, for example, suppose you give a large-language model 10,000 cases of science-related press releases appearing online. Such press releases will typically have a link to a scientific paper being promoted. A large language model can train using some algorithm in which the scientific papers are regarded as inputs and the press releases are regarded as outputs.  From such training, a LLM may be able to pick up outputs that tend to be produced when scientific papers are being promoted. This does not involve any real understanding of the content. The large language model simply learns that certain types of scientific papers become associated with certain types of promotional documents that use language in a particular way. 

We can imagine a day not very far in the future in which the role of pushover science journalists is largely taken over by AI programs such as large-language models.  A few years from now, it might be that a university wishing to promote its latest science study may simply provide the study to some AI program that produces a promotional press release. We can also imagine that in a few years AI programs may largely take over the job of writing clickbait science news stories that are based on a particular press release and a particular scientific study. 

Part of the reason why the work of pushover science journalists may be achievable by AI programs is that such work is so formulaic and tends to repeat the same old catchphrases, the same old boasts, and the same old recitations of the speech customs of materialists. The more formulaic some type of writing is, the easier it is for an AI program to duplicate it.  An example of a formulaic statement constantly appearing in puff piece press releases are statements that sound like:  "This new research sheds light on the long-standing mystery of ____________________."

But there's a type of writer very different from the pushover science journalist.  Using an acronym, you might call this type of writer an ASCOT. The acronym ASCOT stands for Analyst of Scientist Claims of Truth. 

An ASCOT has no goal of providing boastful accounts of some scientific paper. An ASCOT has no goal of writing online stories that provide clickbait headlines.  An ASCOT has the goal of analyzing the truth claims made by scientists, and to judge whether such claims are well-founded or poorly supported, and whether such claims are contradicted by other things that humans have observed.  Part of the job of being a good ASCOT is the job of being a reality detective, one willing to ponder and classify all of the clues that reality gives us, without throwing away clues after giving lame justifications such as "professors don't believe in that." 

There is no way to automate most of the work of an ASCOT, because such work requires analyzing the claims of scientists from a perspective that considers all observations that humans have made and all facts that humans have learned, with a large degree of understanding and insight (something so-called AI systems do not actually have).  Any good ASCOT needs to be a diligent scholar of many different topics and many different subjects.  A good ASCOT needs to have a deep understanding of many broad topics, an understanding that is beyond the reach of AI systems, which do not really understand anything. 

Let us consider a hypothetical example in which some scientific study claims to have determined that Little Neural Effect X offers a solution to what the authors call "the problem of consciousness." Analyzing whether such a claim is a groundless boast first requires some understanding of whether the matter of explaining human minds is equivalent to solving a "problem of consciousness." But an AI system could not have such an understanding.  To decently understand such a thing, you would need to understand how human mentality and human mental experiences and human mental capabilities are something a thousand times bigger or a million times bigger than mere "consciousness." The diagram below illustrates a bit of the understanding you might need. 

If a person has studied all of the items listed above, and the person has some understanding of such topics, he may have done quite a bit of the work needed to be a good ASCOT. But no AI system is likely any time soon to have such understanding.  

A study of all the topics mentioned would produce only some of the knowledge and understanding needed to skillfully analyze scientist truth claims. Very much more study would be needed, including a study of the topics that are discussed on this blog and my blog here. 

AI systems using large language models (LLM) work by web crawling and frequency-of-assertion ranking. This works well when the AI system is called on to answer factual questions such as how did Abraham Lincoln die.  Web crawling and frequency-of-assertion ranking can also produce good results when it comes to opinions that are widely held and true. But web crawling and frequency-of-assertion ranking can fail very badly when it comes to assertions that are frequently made but are untrue or not well justified. Very many of the claims in today's science literature are such untrue claims, claims that are massively made but not justified by observations, and often contradicted by other observations that have received insufficient attention. When unjustified claims are being massively made, AI systems and large language models can aid, abet and amplify the "echo chamber" effect by which some erroneous idea "went viral." 

How can you sort out which commonly made claims in science literature are well-justified, and which are mainly just speech customs and social constructs of a conformist belief community with headquarters in academia? That requires understanding, insight and the most careful and thoughtful study of many very deep subject matters.  That is the type of work that cannot be done by AI large-language models that work based on web crawling to tally up which assertions are most commonly made. To properly sort out which commonly made claims in science literature are well-justified and which are not, it requires insight and philosophical judgment beyond the reach of AI systems. 

An ASCOT will often be someone taking the long, hard, winding path that leads to a destination of philosophical insight after the most laborious study and pondering of many deep topics and many strange and surprising observational reports. There's no way to get to such a destination by unleashing some horde of web-crawling software programs that engage in programmatic screen-scraping and frequency-of-assertion ranking.