Can AI help create popular science content you want to read?

Sure it can.

The basic idea for the Science Reader website is to use AI creatively and responsibly as an editorial assistant in order to help create useful, meaningful science content that people want to read.

When we read AI-generated text, we often sense that "something" is missing.

That "something" is what I call soul.

Imagine that we could identify all the ingredients that make up "soul" - could we somehow mix this into the AI generation process?

Like a pumpkin spice latte. It's still coffee, but because of the added syrup, milk and spices, it tastes like pumpkin spice latte-flavored coffee. Not everybody likes it, but some people prefer it. And we recognize it when we smell it. And some people favor it over pure coffee.

But you know that there is no such thing as a "pumpkin spice latte" coffee bean.

And, as you also know, there is no soul in AI content. Because soul is an organic thing, and machines are not organic, at least not yet.

That is what Science Reader is all about: Learning how to put soul into AI content.

Okay. Why, exactly?

Excellent question. I sometimes wonder why I bother. But popular science is a huge part of my soul.

One reason is: We have to combat AI slop, which is taking over the internet. I feel immensely guilty for having built a site that wass once stuffed full of horrendously soul-less AI writing, but it was all paved with good intentions (it took about 1500 published articles and many versions of my method before I broke real ground). I have deleted no less than 2,700 articles from the site in order to remove ai slop.

The real reason is: I have been a fan of popular science since I first watched Carl Sagan's Cosmos TV-series with my dad back in 1981. I was 11 years old.

Reader alert!
This is a bit of a personal passage.
Feel free to skip to #achievements if you don't care.

In the many decades since, I have worked with science communication in several ways:

• as an internet journalist and editor from 1995-2000
• as a radio reporter in the Norwegian Broadcasting Corporation (NRK). I interviewed people like Brian Greene, John Barrow, George Dyson (son of Freeman) and many others.
• as a freelance book reviewer for the same NRK. I reviewed (or sometimes simply recommended) 100 popular science books, most of them in a live show called "Verdt å vite" (now defunct and replaced by Ekko) between 2000-2010. I wonder if it is some kind of national record - is there a competition for "Has Reviewed Most Popular Science Books On Radio"?
• as the founder of the website Hypography - Science for Everyone (1998-2013), one of the first, major, personal popular science websites in the world
• as a freelance writer for the Norwegian science website forskning.no from 2000-2003
• as the web editor for the Norwegian Space Centre (2003-2008), where I also did a lot of public outreach like visiting schools, talking at conferences and presenting to students
• as a partner in a strategic communication agency where we worked with technology customers like Microsoft
• and of course, I have read a lot of popular science books.

"But," you ask. "This was so long ago!"

Yes. Sadly, I was brought down by illness. In the years 2013-2022 I struggled with increasing, debilitating pain. For long years I was unable to work more than a few months at a time, and spent an insane amount of time at home, much of it in bed. In the end I could not even read books. (And - somehow, miraculously - I managed to hold onto a job as a CMO for an IT company from 2013-2024, when they merged with much larger company. What a fantastic employer!)

By 2023 I was able to crawl back to life. I missed working with science, content and websites - so I started Science Reader in October 2023.

For the past two years, this hobby project has helped me regain my love for popular science, reading, writing - and, most of all - editing and publishing content.

So that's why. I do this because I love it, and because I can. And because I simply hate seeing AI kill our vibe.

So, enough history.

What have I achieved?

It has taken a lot of trial and error to get here, and that is why you find a lot of (sometimes very) bad AI written content here.

But in the fall of 2025 I started to figure out ways to solve the "soul" problem.

Like a lot of things with AI, I realized that you can't make it human. All LLMs are simply fancy calculators that can predict the next word in a sentence. They are getting very good at it, but it is what they do.

So the "coffee beans" that AI content is brewed on are simply that - calculated words.

In order to add the pumpkin spice latte flavor (ie, the "soul") - we need three things:

1. Understand what makes human text human
2. Distill it into something tangible
3. Apply it to the AI content creation process.
   
   And, it really takes a 4th "thing" (notice that this is the only tangible "thing"):
4. A human cowriter and editor who can control the creative process, interact with it, and make editorial decisions based on human insight and experience (and, sometimes, like all good editors, throw it all out and start over). And of course rewrite and edit any part of the text.

Sounds really simple! I like to call this process "duowriting". It requires a thorough, organized approach - to create quality content, you must spend quality time with that content.

Creating quality content of any kind takes hard work.

What does science say about "soul" in content?

According to research, there are clear and measurable differences between AI-generated content and human-written content. (Apart from the obvious fact - that the first is generated by a statistical analysis machine, and the other is actually written by a living being).

For example, the PAW Scale (Perceived Authenticity of Writing) uses 17 indicators to evaluate if a text is considered to be authentic - in the sense that it measures how a reader perceived the text as meaningful, relevant and connected to their lives. This was covered in a study all the way back in 2014, but I find it quite interesting because it attempted to identify - in my view - what kids perceived as soul in writing.

PAW Scale focus areas

Category	Number of Items	Examples of Focus
Community & Global Relevance	6 items	Writing that connects to community issues, global concerns, or real-world audiences.
Personal Relevance	5 items	Writing that relates to the student's own life, experiences, or interests.
Academic Relevance	5 items	Writing that supports school learning goals, subject knowledge, or academic skills.
General Authenticity	1 item	Overall judgment of how authentic the task feels.

Science vs ChatGPT

In a more recent study (2023), researchers compared articles in Science to content produced by ChatGPT. Not surprising, they discovered that even though ChatGPT was trained on insane amounts of human text, it did not write like humans.

The study showed clear, measurable differences, for example:

Difference	AI	Human
1. Proper Names	"Researchers" (generic)	"Dr. Laura Brannigan" (specific, personal)
2. AI Fingerprint Words	"Tapestry," "delve," "camaraderie" (overused)	Normal frequency
3. Rhythm	Monotonous (metronome)	Varied (jazz)
4. Nuanced Language	Avoids doubt ("but," "however")	Actively nuances
5. Punctuation	Simple (commas, periods, em dashes –)	Varied (-, ;, :, ?)

Here are some contextual examples to make it easier to see the differences for yourself:

Category	AI Example	Human Example
Proper Names	"Researchers at the university discovered..."	"MIT's team led by Dr. Laura Branigan discovered that self control..."
AI Fingerprints	"Let's delve into the intricate tapestry of this discovery..."	"The discovery reveals how proteins fold..."
Rhythm	All sentences roughly same length (15-20 words each)	Short bursts. Then longer, flowing explanations that build complexity. Varied.
Nuanced Language	"The study shows climate change affects temperatures."	"The study suggests climate change may affect temperatures, however other factors complicate this picture."
Punctuation	"The results were clear. Scientists found evidence. It supports the theory."	"The results were clear–but puzzling. Scientists found evidence; it supports the theory. Or does it?"

The Pattern: AI generalizes and simplifies. Humans specify and vary. This holds across multiple measurable dimensions, creating a detectable "signature" that we can use to distinguish what constitutes an authentic human voice from AI-generated content.

Stance markers - the human perspective

A researcher worth following is Ken Hyland from the UK. He writes a lot of papers on human vs AI writing in various contexts. One important area of research is the concept of stance markers - how humans shape text through little words and phrases to give it perspective and quantified quality (my term - imagine that a stance marker gives you clues to how the writer evaluates a property when seen from a human perspective).

This is something LLMs can only simulate. An analytical machine is not able to actually perceive the world or understand human evaluation (well...we might be getting there).

Stance markers can be classified as follows:

Type	Function	AI Frequency	Human Frequency	Examples
HEDGES	Soften statements, show uncertainty	Low	High	"might," "perhaps," "seems," "suggests," "appears," "could," "may," "probably"
BOOSTERS	Strengthen statements, show certainty	Moderate	High	"clearly," "obviously," "certainly," "definitely," "undoubtedly," "indeed," "surely"
ATTITUDE MARKERS	Show author's attitude/emotion	Very Low	High	"surprisingly," "fortunately," "importantly," "unfortunately," "remarkably," "intriguingly"

Again, here are some contextual examples to make it easier to see the differences.

Type	AI Often Writes	Human Researchers Write
HEDGES	"The study shows that climate change affects ocean temperatures."	"The study suggests that climate change may affect ocean temperatures."
BOOSTERS	"The data shows increased temperatures."	"The data clearly demonstrates significantly increased temperatures."
ATTITUDE MARKERS	"The researchers found a new method for measuring..."	"Surprisingly, the researchers found a new method for measuring..."

It is important to point out that Hyland's data is from academic writing (research papers to other scientists). For popular science writing (web content for general audiences), these patterns may not apply directly.

What is important is that the use and frequency of stance markers are measurable quantities. AI will typically use less than 10 stance markers per 1000 words, while human writers will average 15-25. Since human stance markers are also more varied, texts are naturally richer and diverse. Again, these statistics is for human vs AI academic writing, not popular science. But I hope you get my point - the differences are measurable.

What this means for "soul" in AI content generation

It means a couple of things:

1. It is possible to quantify some aspects of human writing
   • For example: Rhythm, symbol usage, stance marker frequency, use of pronouns, sentence and paragraph variation, metaphor use, transition techniques, openings, closings, verb forms, comprehension level
2. There are human writing techniques that are uniquely human
   • Like perspective (as in retelling the experiences of real vision, hearing, smell, touch), thoughts (as in inner monologue or random, sudden insights), writing from experience, writing in the moment (touched by what happens right now - that blaring fire truck horn, the noisy café, the stormy ocean mist you just came in from), having a relation to the subjects in your story (because you inhabite the same world as them)

So there are some things we can make an LLM do in order to create more interesting, engaging texts, but in order to make it good - to give it soul - we need to also make it try to generate texts in the same way a human writer would.

If we expect to be able to say to an LLM, "write this article so that people want to read it", it has no idea what you mean. But a human writer (with some experience) knows how to write engagingly for a human audience.

I use four main components in my work:

1. A 10-step editorial workflow created specifically for AI content, including source selection, research, editorial pitch, drafting, fact checking, editorial review (ie, discussion with me), final draft, image creation, SEO, Wordpress preparation
2. Custom personas that are "distilled" personalities tailored for LLM use
3. A Science Reader style guide which some "brand components" that are not unique for personas, but which need to be observed by all personas and content types.
4. A human editor and cowriter (you know who) who oversees all editorial decisions and can instruct the assistant to evaluate, remove and change anything at any time - and who is actively taking part in the writing process.

I'll explain more about det details of my workflow at a later time, for now it is my little secret toolbox.

The Graham Farmelo persona

Here comes the main juice of this very long story.

I apologize in advance to Mr Graham Farmelo, one of my all-time favorite writers of popular science books, articles and interviews.

(Hey Mr Farmelo, if you ever read this: your book "It Must Be Beautiful (Great Equations of Modern Science)" ranks among my best-ever popular science experiences. It was also one of the books I reviewed on Norwegian Radio back around 2002 when it came out).

One night I was reading "The Strangest Man", Farmelo's biography of Paul Dirac ("The British Einstein"), and it struck me - Mr Farmelo's writing has all the kinds of ingredients which makes up "soul" of human content. It is simply amazing.

So I, ahem, borrowed some of it. Or rather, I tried to bottle it.

I gave Claude Graham Farmelo's website, his Wikipedia page, I fed it a selection of his articles (here are some of them) and through a series of collaborative nights with Claude, we managed to distill a custom persona for Science Reader and bottle it.

It is of course a work in progress, but at the same time, it is a proof of concept.

So "my" Graham Farmelo persona created this story about Norwegian Jazz Guitarist Jon Larsen who, while eating strawberries on his forest view terrace, found a speck of stellar dust and revolutionized the search for micrometeorites.

While I have not met Mr Farmelo, I have met Mr Larsen, who has an amazing presence and a science communicator's gift of gab. (Hey Claude, try to use that expression in a natural way!)

But what about AI detectors?

This has become a new industry now - tools that can tell you whether or not a given text was written by a machine or a human.

Well, as was showed in a 2023 paper, they are easily fooled (as I will prove to you in a minute) and - importantly - they are of course only analytical machines trying to figure out whether the input they get is the output from other analytical machines, or real writing from humans like us.

One trick they use is to recognize the fact that LLMs have unique "fingerprints" that can be detected.

But guess what? My approach actually turned out to work. After two years of trying "a thousand things that did not work" (yup, a Thomas Edison quote for measure) something finally clicked.

My apologies to those who click around my site and think, "wow, there is soooo much AI slop". It's collateral damage. But - I have spent a lot of time on editing, polishing and evaluating bad stories, and I have deleted thousands of articles (failed experiments!). So the ratio of gold vs junk is shifting towards gold - and a lot of visitors only ever see the 404 "article is missing" page, and hopefully find another interesting story to read.

Well, Tormod - how about some results?

Well, I am a man of words, and of words, I probably have too many. But ok, here goes:

I already discounted the AI checkers, but since they are all over the web now, let's give them a chance.

And - case in point - This Science Reader article on AI self-awareness gained a "Your text is human written" with a 0% AI score on both Grammarly's AI detector and the ZeroGPT website.

Quite ironic, considering it was entirely written by Claude Sonnet 4.5 using my Graham Farmelo persona - and I even state in the editor's note that it is written by Claude.

ZeroGPT: "Your Text is Human written" (Sheesh, who wrote that phrase...)

Grammarly AI checker: 0% of this text appears to be AI-generated

The bottom line

So, after all this, do I have a working concept that enables me to create AI generated content which people want to read?

Is my Graham Farmelo persona the Science Reader Pumpkin Spice Latte syrup?

Well. No. I'm not quite there yet. I can fool the AI detectors, but I can't fool my readers.

But remember - my goal is not to produce content that is human-like. My goal is to create content you want to read - even if you know it was created together with Edie, my AI editorial assistant.

And that will take more work.

I really hope you enjoyed this little background story. Find me on LinkedIn if you want to discuss anything (link below).

And a tiny postscript: My Graham Farmelo persona is not actually Graham Farmelo (in case you got confused, which I doubt). It will not make Claude write the same way he does. But it can make Claude simulate writing like him as it generates the content - to a small degree. I am working on improving it - it requires a lot of learning about how LLMs generate text. I think one should consider the "Science Reader Persona" a mix of my favorite writers and my own style, blended to create a unique style that will evolve as I learn more about LLM word prediction and how to manipulate or direct it.

It still takes a human being to write like we humans do. Thank you to every human writer, and thank you to every human reader. Let us never stop writing and reading science.

Wait a minute! What about fact checking?

It's no suprise that LLMs are prone to hallucination - writing text that sounds correct and factual, but was simply the most likely mathematical solution for the AI creating the content.

To combat this, we use a pretty thorough fact check process at Science Reader:

1. Source Control: First we do a basic source review and evaluate if the sources are trusthworthy. This (mostly) weeds out all content mills, but we still need to manually remove some sources now and then.
2. Initial Fact Check: We then run a fact check on the first draft of the story.
3. Epistemic Check: Then we check for epistemics - which basically asks "is information presented in a way that matches the trust level for the facts and assumptions we are basing the story on?"
4. Final Fact Check: Before publishing the story in Wordpress, we run a final fact check using Perplexity's Sonar Pro Search model, to ensure all sources are rechecked and that the story is matched against them again. This is actually a three-step check in itself.
   • Claude extracts what it considers the claims in the story. It then sends the entire article to Sonar Pro Search
   • Sonar returns it's own claims list and says what is verified and what is not - and documents it.
   • Claude then matches these claims against its own list. Any disagreement is sent back to Sonar with Claude's arguments or suggested edits to the article.

The results are stored with every article, plain for all to read - includind which changes Claude made to the article because of the fact check results. A yellow result means we need to fix something (we usually don't publish stories with a yellow result, but sometimes we rerun fact-checks and something may have changed since we published the story, and we may need time to correc it). Green is "go".

We aim to show the fact check results with every article. If an article doesn't have it, it is probably an older article and we did not store the fact check back then. I do re-run fact-checks occasionally, but each fact-check costs money so I try to ensure it is necessary.

Why the name “Science Reader”?

The term “Science Reader” is describing the target audience for this website. I want to give you more stuff to read (because hey, the world is not full enough of great science content, there is always room for more...as long as it is good).

And of course - I am a science reader! And maybe you are, too! I think it is vital to read, learn and be fascinated by everything we find out about ourselves, our society, our planet, and our cosmos.

I hope this site will inspire more people to read and engage, but if nothing else, I have fun while making it.

The "Science Reader Recommended" badge

So, here at Science Reader, we not only create science content. We find and share a lot of it with you! A lot of our articles are reviews or condensed explanations of content I find elsewhere.

In order to showcase content I really think everyone should read, I mark it with this little "badge of honor". I always credit sources, and I always strive to name them in our stories.

By the way, that is one thing great writers (like Mr Graham Farmelo) do - they make science personal, they make it relatable, and the bring out the human beings who did the science and made the discoveries). And so will I strive to do.

Who am I?

My name is Tormod Guldvog. I was born in 1970 and I live in Oslo, Norway with my French partner (who also loves to read, but she reads a lot more non-science than I do).

I work in IT, at the crossroads of communication, collaboration, strategy and management. I have worked as a science reporter, and I have reviewed more than 100 popular science books on NRK Radio, the national broadcaster in Norway. I also have 30 years’ experience building websites and writing content.

If you are interested in discussing AI writing, AI editorial workflows or maybe the art of the perfect pumpkin spice latte (which I really know nothing about), feel free to hit me up on LinkedIn.