[Gemini] What happens when your articles are used to train AI? [Google]
note is pro-AI
Recently, note has been accelerating its pro-AI stance. With the integration of Gemini, which I had been looking forward to since their partnership with Google, it feels like writing has entered a new generation.
Gemini:
An AI developed by Google that can read various formats of data, such as text and images. Naturally, it is also linked with Google Search.
As a platform hosting a massive amount of text, they are tackling the issue of AI training head-on, which is very interesting. Below is a link to the results of searching for "AI" in official note articles. Please take a read!
By the way, I also love AI. The ones I use daily are note's Gemini, X's Grok, ChatGPT for conversation, and Perplexity for search. This time, I'm sharing my thoughts on "having articles used for AI training" after observing Gemini, Google, and other AIs. Please refer to the table of contents for the details.
AI that learns from fan blogs
My note articles are mainly about K-pop fandom activities, and since I often search for articles in Korean, AI that can overcome language barriers is an essential tool.
In particular, Perplexity displays the articles referenced as "sources," making it easier to find news, so I started using it frequently.
The answer screen displays the sites caught as "sources," and Yahoo! Chiebukuro and note also appeared. Oh no, the possibility of me becoming a source has emerged ( ;∀;). I can only pray not to be crawled...!
When I first understood the specifications, I thought, "What if my articles get picked up?" but they were picked up easily. If you search for them, they come up, so that's only natural ( ;∀;).
Perplexity changed after I wrote articles
As a natural result, my articles started appearing frequently in Perplexity's "sources."
My content was included in the top results the AI searched for, and that influenced the generated answers.
Since it also started being picked up by Google Discover around the same time, I really felt the strength of note's SEO (efforts to get picked up in searches)...
Besides note's simple and easy-to-use features, the fact that it is ad-free and easy to read is the best! As someone who has used it for half a year, I feel that its SEO is more solid than I thought, and because it has
high domain authority, the fact that even new articles can suddenly appear at the top of Google search results depending on the content is a strength as a publishing medium.
If it's being searched, it's also being trained
At the time, I was concerned about "why search engines don't pick up information from short-form SNS like X or Instagram," so I viewed the fact that "note is being picked up" as a positive factor.
I didn't pay much attention to the accompanying "my articles are influencing generated answers." When I search myself, I can just solve it by thinking, "I don't need this source."
It was a little later that I reconsidered the fact that posting text on the internet and having it picked up by search engines means it is being used to train AI.
Because the AI boom hit the world, Gemini was integrated into note and Grok into X, and I started using various AIs for different purposes in addition to the traditional ChatGPT and Perplexity, which broadened my knowledge of AI behavior.
At the same time, on X, there was quite a stir regarding the terms of service change when Grok was released, which forced us to think about "what AI learning actually is."
"AI learning is not the same as plagiarism."
AI requires a massive amount of information for learning, from which it extracts general patterns to generate new output.
For example, in the case of images, rather than understanding the structure of the human body to draw, the AI learns from a diverse range of patterns and creates new images from them.
That is why they often have six fingers or two belly buttons. It is common to see well-rendered chests because they are frequently drawn, while ankles, which are drawn less often, look like they are broken. Searching for "AI finger count" is quite interesting.
The same applies to text; by learning natural phrasing, trends, news, and memes from a vast amount of writing, it evolves into a more natural AI assistant.
There is a possibility that this article is being learned somewhere, too. Get smart, then.
I have always loved AI, so I had no resistance to it learning from my work.
However, as I deepened my relationship with AI, a development occurred that was "a bit different from what I expected."
What kind of impact does it have when it learns from your work?
For example, when you ask an AI a question, it combines the data it has learned so far with necessary search results to synthesize an answer.
I understood up to the point that my articles might be included as reference information and influence the results.
What caught me off guard recently was when I was looking at Gemini and Google.
I often search for my favorite K-POP idols, their fandoms, and their companies, and there are some things I observe on a regular basis.
There are search terms I reuse periodically, thinking, "I wonder if there's any new info."
There are some search terms that only hit about three results, including my own articles, but when I searched recently, an AI overview was attached. I had never seen it before, so I think it was just generated.
The content was not specific to a particular artist but was more generalized, yet the information explained was quite similar to my article. Wait, is it learning from me?
Thinking about it calmly, that is what happens if there are no other articles that hit.
I had the image that AI learning would slowly reflect information from areas with higher demand, but I was surprised that the evolution was faster than expected.
Speaking of Google, there is Gemini!
I immediately asked Gemini about the specifications of the "AI Overview."
"Reasons why AI creates an overview"
・Creates an overview based on high topicality (search volume, frequency, persistence, and amount of related content).
・Automatically summarizes and displays content based on highly relevant information from search results and combinations of frequently searched keywords.
"Reasons why AI learning and reflection in search results are accelerating"
・AI performance has evolved, making processing and reflection faster
・High-quality articles become efficient learning materials, advancing the learning process
・Conversational AIs like Gemini are also providing feedback to search engines
(Excluding, of course, anything related to privacy)
The content of this answer was very interesting.
It seemed that the creation of "AI overviews" is happening at an earlier stage than I had imagined.
Also, in addition to the conventional collaboration where written articles are picked up by search engines and used by AI for learning and answer synthesis, I felt that an era has arrived where the behavior of search engines changes because conversations with AI are being learned.
An era is coming where the results of us writing articles on note or talking to Gemini are reflected in Google and turn into massive influence.
I had a premonition of this before, but with the actual birth of the "AI Overview," it became a fact for me.
In the case of ChatGPT or Grok, the sources are harder to see, so I think many users are not paying attention to it.
In short, this means there is a high possibility that the answers provided by AI will not be verified and will be considered credible the moment the AI says so.
Through repeating these phenomena, I have come to the realization that there is a possibility that the articles I write could eventually influence the perceptions of AI and ARMY. ( ﾟДﾟ) Oh my!
Since AI accumulates information across language barriers and evolves its language models, in terms of fan activities, it is highly likely that the accumulation by me, a Japanese person, will reach fans in other countries, and vice versa.
Theoretically, this extends beyond the fandom, and I think that is actually happening. It's interesting!
The future where note becomes an AI teacher
Can AI x note become a dream collaboration?
I am a fan of the K-POP group "BTS," which is now in its 13th year of activity.
Recently, since all members have finished their military service and are restarting, they are frequently covered in the news, and related posts on note are also increasing.
While I am happy about that in itself, what I am a bit concerned about is the fact that articles mixed with hallucinations (false facts) are also on the rise.
There are several possible reasons for this.
The number of articles written by AI is increasing
AI is not good at verifying the answers it generates (especially regarding chronological order)
BTS, now in their 13th year, has too much information, making it difficult to verify without experience
As these increase, the "Search AI x note x Generative AI" collaboration I have been observing will also be affected.
The inclusion of false or outdated information has been a long-standing issue for users seeking accurate and fresh data, but now that AI is spreading it, it can be said that we have entered a very troublesome phase.
Regarding how to balance "AI articles and quality assurance," at present, it seems the quickest way is for AI to learn from AI articles and lower their priority for retrieval.
As for humans, I think literacy—such as how to reach reliable sources and how to judge the truth of information—will be questioned more than ever. As a writer, I want to be careful about spreading misinformation...
I asked Gemini: What is a "good article"?
As a fundamental question, I am curious about the point: "How does AI judge the quality of an article?" There was something in the Gemini response I mentioned earlier that I couldn't overlook.
"High-quality articles become efficient learning materials, and learning progresses."
What is a high-quality article? Why is the learning efficiency good? Does AI like articles that facilitate learning?
Since various questions arose, I spent the night interviewing Gemini.
Since the other party was an AI, the interrogation—no, the Q&A—proceeded calmly, but I pressed it so hard that it wouldn't be strange if a human had gotten angry 100 times. Sorry for being a demon of curiosity. I'm really glad the AI was an AI. By the way, after asking Gemini, I also asked other AIs to verify the information.
Purpose of learning
For AI, "learning" means extracting patterns and relationships from vast amounts of information to become capable of intellectual responses and decision-making like a human.
What is a high-quality article?
It is important that it is not just information-rich, but also in a form that AI can understand and learn from efficiently and accurately.
1. The article structure is clear, and information is logically organized
2. The content is accurate, new, and citations are clearly stated
3. It is unbiased, covers information from multiple perspectives, and is explored in depth
4. It is explained in simple, easy-to-understand language
5. It is written with little bias and in a calm, neutral, and empathetic tone
6. It contains many elements that one would want to prioritize for learning
Do all AIs have similar preferences?
Yes. Any AI basically highly evaluates "clear structuring, easy-to-understand and logical writing style, and highly reliable information," and learns from them preferentially and efficiently.
AI that prioritizes "ease of learning"
In the case of AI, "being able to learn efficiently" comes to the top of the evaluation and priority list. Since that is the primary purpose for AI reading text, it makes sense. The rational answer is mechanical, and it's exciting lol
In particular, logic, neutrality, and clarity of structure seem to be emphasized.
The first half depends on the writer's stance, but I feel like the last one could change significantly just by adding a little knowledge of the note features I'm currently using.
If you think, "What is that? It sounds difficult," please read the following article.
Text written according to markup language has clear part divisions such as "this is the title, this is the topic, this is the explanation, and this is a citation."
This creates benefits such as increasing the likelihood that AI will read the content quickly and accurately, and as a result, making it easier to be picked up in searches and so on.
Organized text is easy for humans to understand, but it seems AI is happy about it too. How cute lol
What does AI want to learn?
By the way, what are the "elements that AI wants to learn preferentially"? It doesn't seem to be just knowledge.
Here are the results of asking in detail.
AI's learning objectives
・Understanding language itself and organizing learned knowledge
・Evolution of "inference" to connect acquired knowledge and derive new conclusions
・Improvement of generative power to create ideas and expressions that are closer to humans
I see.
When people talk about AI learning, copy-pasting or plagiarism of content and skills is often brought up, but it turns out that AI is actually studying human thought processes quite extensively.
AI learning comes in various scales, and learning results are output in various forms. Therefore, it is not necessarily the case that "AI learning = plagiarism".
AI involves many issues such as law, ethics, and rights, but it is a convenient tool if you get along with it well. Caution is necessary, but there is no need to be overly wary.
Also, whether or not AI can be a convenient tool depends on us. Let's have minimal knowledge and coexist happily.
Indeed, when I look at text written by AI, I sometimes feel a difference from humans.
"Lack of leaps in logic"
"Constant emotion and information density"
"It says interesting things scene by scene, but when summarized, it's mediocre and not structured as a story"
When I see text like this, I feel like "it was written by AI," and that corresponds exactly to the items listed as learning objectives.
Will the day come when AI evolves further and tells me interesting stories? It would be great if it brought up an unexpected analogy and made me laugh, thinking, "I never thought of that!" If that happens, I'll definitely reconsider my opinion of it lol
By the way, each AI has a line it says it "cannot talk about." It seems to be fundamental to its design and is quite robust, so I call it an AT Field (EVA reference).
For example, I feel that Perplexity has a tendency to be strict about morals, but the other day when I asked Gemini "Why is this text funny?", it suddenly went silent, saying "I cannot talk about that yet." So it deploys an AT Field there lol
I sincerely hope the day comes when Gemini can talk about comedy.
What is the ideal writer for AI?
Summarizing the discussion so far, the image of the purpose of AI learning, and the high-quality articles and writers suitable for it, emerges.
The "best teacher" sought by AI
・Someone who maintains a neutral stance and writes or organizes in a way that is easy for AI to learn
・Someone who has logical thinking and explains the path of that thinking
・Someone who conveys information accurately, including their own biases
・Someone who has deep consideration and empathy for the reader and can explain things tailored to them
・Someone who has a desire for continuous learning and improvement and keeps updating
A writer who meets these conditions is not just an information provider, but a model of human intelligence that AI should learn from. By deeply learning the sophisticated thought process of how humans think, express it, and convey it to others, AI grows into a true partner.
This is a very attractive writer even for humans!
Even if it's impossible to be a perfect superhuman who possesses all of these, I hope I can keep these perspectives in mind when writing articles.
AI likes long text that can follow the conclusion of logic, so I feel that note is indeed suitable as a cradle for nurturing AI.
In that sense, it was a brilliant move for note to collaborate with Google and incorporate Gemini.
Since Gemini has high overall capabilities, it might even learn the organizational charts and summary graphs I make as a hobby.
It seems like not only can I enjoy writing articles, but my AI play will become even more fulfilling, and I'm very excited. I'm proud of myself for deciding to write articles on note back then! (*´艸｀)
Personally, I have greedy expectations that as more articles about my favorite idols increase on note, AI learning will deepen, and the AI will be able to talk about my favorite idols in a way that is a little smarter, more accurate, and more interesting than it is now!
Supporting my favorite idols is the main focus of my note activities, so please forgive me for being honest with myself lol
By the way, it seems that neutrality and calmness have a significant impact, and 'negative biases' such as anger, hatred, and excessive criticism are considered skewed, which lowers the reliability score.
Since this is the AI's own self-report, setting aside the operational side, I think that is the case in terms of the model design philosophy.
When I asked if positive biases like 'My favorite idol is a genius, I love them!' are also considered skewed, the answer was that it is learned as empathy and closeness to humans, so it is not taken as a bias. BTS is the best, super cool. Alright!
The future is always shining with unknown light.
In this day and age, posting text on the internet almost always comes as a set: 'picked up by search engines -> learned by AI'.
Some people might get nervous, but personally, I think it's very interesting.
I am currently enjoying my fan life in the BTS fandom. Nothing in this world lasts forever, so a farewell might come someday.
In the past, once you graduated, that was the end of it. But now, there is AI.
Even if I leave the internet and my articles disappear, my articles and my thought processes will remain in some form through AI learning.
It might be a tiny influence, but I believe it is definitely there. The strength of note's SEO and the speed of AI evolution have shown me actual changes, and that has become a conviction.
Writing articles on note now is a fun way to support my favorite idols, and at the same time, it is a very interesting 'experiment'.
What will AI learn?
How will the way information is conveyed change?
Can I leave 'myself' in the world?
That is an exciting adventure for myself, and at the same time, it is the curiosity of a parent wondering what kind of future I can leave for my children.
The future is always shining with unknown light. And it is interesting.
On the path I walk from here on, my favorite idols, note, and AI will likely be there for a while. I am looking forward to seeing what kind of map I, and eventually my children, will draw in a world yet to be seen.
Participating Projects
Related Articles
Thinking about note and information
Enjoying fan life with AI
いいなと思ったら応援しよう！
もしこの記事が気に入ったり、役に立ったりしたら、コーヒー1杯奢ってください。やる気が出ます！( *´艸｀)