Bounties attract serious brainpower to the challenge.
Why?
to always be able to prove that the text is AI-generated
to define "digital handwriting" as a category
to know which AI generated that text (AI-specific signatures)
to prevent fraud (people earning money on AI-generated stories, novels, etc)
How would it work?
In written text, handwriting is much more distinguishable between individuals. That's not the case in digital text. Here all the text seems the same, in a font you choose, that can be easily changed. However, every open-source AI algorithm for text generation has a certain logic behind the text generation strategy. It shouldn't be hard to implement a part of code that defines certain signature patterns of word uses or frequencies of certain word types, letters, etc.
Ideas:
specific word type (e.g. verb) appears in text always after a certain sentence of comparison
comma is never used after a sentence where inversion happened ("Every time we came, he wasn't there.")
colours are always represented in shape like "shoes of colour blue"
many of traits like mentioned above combined into a "handwriting signature"
What do you think? Is there a better way how to create this "digital handwriting"? Do you have some ideas on how to improve the idea? Is there any linguist which could think of a cool writing patterns that could be used?
Based on the comments from Marco Agudelo, I came to an idea of an AI that will be specialized in checking if the input text is AI-generated or not. It would be a sort of digital cop. I even have thought of the name - ForgeAI.
Why?
to prevent forgery in text-based industries
to be used as a check tool for seminars, master's or doctoral thesis and homework
How would it work?
The AI would read all the texts on the internet and categorize them as either human-written or AI-written. Then it would search for differences and traits that can help recognise the AI-generated texts. If the signature patterns I mentioned in the idea text were used, the job of an AI would be easier. If not, the AI would search for a code-specific patterns of written text and try to decipher forgery in text-based industries like poetry, prosis, movies, theaters, news, etc.
Please leave the feedback on this idea
Marco Agudelo2 years ago
Even though this could sound viable, it should be analyzed with detail at least to use it directly within a specific website (e.g text-based portals), since it may fall into the category of Men in the Middle and should be verified with TLS 1.3
Wouldn't it be enough to embed codes in the text, similarly to watermarks on copyrighted images? This could be done in many ways. The simplest one would be to insert code numbers (like 45XYC239K) here and there in the text - at the end of the paragraph, at the end of some sentences, etc.
To make this minimally visible, barcode-like stamps could be used. They could be extremely tiny and appear as superscript (basically dots) over some words.
Such codes/stamps could contain information about the ID of an AI and the time when the text was generated. Maybe something else?
On one hand, codes might be easier to remove from the text to hide the fact that it is AI-generated, but on the other, you might as well use a program to detect signature patterns (that are, in fact, a sort of code as well ) in the text and then rewrite it to eliminate them.
Another way I can think of is to use letters in the text to form a code (every this and that letter is counted). This would be more complicated and more cryptography-like, but at the same time more secure.
Please leave the feedback on this idea
jnikola2 years ago
Great idea, but I am not sure that all text-editing software (both online and offline) have the function to implement barcodes or "tiny dots" you mention. On the other hand, "45XYC239K" could work, although aesthetics sometimes matters and not everybody would be happy to implement that. But we are on the right track.
The fourth paragraph problem could be solved by having a closed-code AI cop that would dynamically change to adopt all the current trends in signature pattern recognition.
In the last paragraph, we are back on the right track. Instead of words, certain letters could create a barcode that can be read by software. Nice! The sequence between "key letters" should be defined by the length of the paragraph/text, type of text, or? How could we make it simple for the algorithm to compute and do it seemingly and not computationally exhausting, but still secretly and fake-proof?
Please leave the feedback on this idea
Povilas S2 years ago
J. Nikola Nice thoughts. I'm not sure how AI text generation works. Can they just be hooked up to a text editor like MS Word or do they use their own software and then repost/insert the texts?
The only text-generating AI I've ever dealt with is chatbots. And they do use their own chat apps. I've never seen one hooked up to messenger or instagram. This is perhaps for security reasons and maybe it's a bad example, but the same could be the case with text editing software.
In any case, if such AI-identification systems became widely accepted, usual text editing programs could be adapted to include the required codes in the text.
About aesthetics - I agree, but a rather strange text structure/syntax is also a dubious case regarding aesthetics. Inserting codes leaves freedom for the text to be more freely written (generated).
Yes, the challenge with text-embedded code is not easy. This is because you need to both write a logical and specific text and have certain letters in it at certain places, so you'd either need to make mistakes (insert wrong letters) here and there which would reduce security level making it easier to spot or chose certain words that would naturally contain certain letters in certain places. The second case would make the text look a bit off. But since it's AI-generated this is perhaps better than inserting wrong letters in the right words.
Please leave the feedback on this idea
copyrights on AI matters is important
Marco AgudeloOct 17, 2022
On how to manage AI assistance when creating creative processes.
After reflecting on the subject with present tools and my comprehension of the subject I propose to start the AI copyright infringement booster by concatenation of events similar to what was work on this session rather with one created from scratch, considering the capacities of these systems .
Why
Writing is a creative process which can unifies in various ways a similar statement. Is like having a function F(x) ā Y, such that multiple inputs are having the same output; take for example square function, you canāt tell if a number was brought into the function by its positive or negative values, both will have the same output. In a similar way, when a person request to a AI syntax motor to write, the given output text could happen to be generated by any person. It could be compared to if a person requests someone else to write for. So by analogy if a person requests another person to write, the style will vary from person to person. Since it is harder to generate content with different style every time (considering a human is writing) focusing on the long term analyzing, Iām certain style would be a key concept to analyze AI copyright fraud as well as a tool to contribute to public recognition of AI tools as well. In the way AI increases their capacity each model will defer from others showing a style concept. If one could dare to extrapolate things, the situation could go as far as the subject is attained in the movie ghost in the shell.
How it works
The ability to concatenate various writings, if a machine learning system could train to determine style and classify by user, then each new writing will give to the system more data to determine if a given user is retaining the style or is switching. Also businesses that concern this problem could benefit from deep understanding AI generating natural language models styles . Indeed AI models already have things that distinguish between them and the user can fine-tune some parameters (those fine tuning could be turned into a back door to trick style recognition).
If there is an AI style interpreter that has been trained with selected information from the creators of distinct AI generating natural language models, then could it be biased to recognize the fine-tune users' adjustment. It would be something similar to talk to the same inteligible thing but with variations; this subject requires further analysis of implications, but it is similar to talking to a friend; they can change but we find somethings deep in their behavior that lets us recognize them
To the businesses that concerns this issue, each user generated content will sum to generate their style, which will maybe let pass some first tries to cheat the system, but in the long term may offer a stable solution to the challenge and if AI models count already with style pattern, even first tries could be catched.
Also in my opinion it is important that terms of use are actualized to include this new reality. Start with trust but keep on the long term with proven results.
If you see this route as an alternative, the next step would be to analyzed with detail term of use and capabilities of actual AI models to adjust this proposition.
[1]Shubhankar Kulkarni, 2020, What are better alternatives to using a password to log in?, https://brainstorming.com/r/s192
J. Nikola Absolutely in accordance. To begin we can ourselves practice over those AI playgrounds (I do it quite often myself), it could be noticed that AI builders trying to avoid this kind of copyrights infringements suggest to (1) clearly state that the code was aided through AI. (2) claims as yours the generated text, since AI will not generate the written content out of nothing instead uses input information (3) third users can ask to AI platform to verify a given content (Iām not sure how does it works, but I read it on the platform I use)
I have noticed that AI written code is more opinion based and will not point out concrete facts on its own, only when clearly requested to do it. So here are two things in which I will be working out to generate a creative contribution into this brainstorm.
AI platforms try to minimize user abuse and this subject seems to me a very important matter to them.
Should AI syntaxis be human recognizable?
I believe AI platforms can have the power to search over the internet for copyright infringements, but that would be so intrusive that maybe shouldnāt be the way to manage.
Because the platform generates written text (this is pure txt code) only regulating syntaxis will determine the wanted condition here, but by doing so, this can also decrease the quality of the developing technology.
A possible way would be to send the text to evaluation when doubted.
Update terms of use of potential affected platforms, by clearly stating how to manage this kind of situation.
This is a very needed and challenging question J. Nikola has put out here to test!š
Take a look at what one AI says about the subjectš
Marco Agudelo Thank you for the in-depth comment. AI builders, the same as virologists or nuclear scientists, suggest using their inventions and findings for good purposes. However, here I want to deal with a problem that will undoubtedly emerge - using AI as a tool to generate more content and earn more money.
Also, I do not 100% believe that it should be strictly regulated because AI could probably serve us as a tool or an add-on to Word for faster writing or content generation (as you said, "AI will not generate the written content out of nothing instead uses input information"). Therefore, yes, AI development mustn't be disrupted but should be controlled somehow.
Comment on your bullet points:
I'm happy to hear that AI platforms try to make this problem important. In my opinion AI, syntaxis doesn't need to be human-recognisable. We could use another AI developed just for this purpose. Currently, AI-generated texts could be recognized due to insufficient details regarding grammar, repeating words, etc., but the time will come when these texts will become astonishingly great. That's when we will need a strong AI to help us decipher which one is author-original and which is not.
PS I like how the AI tried to avoid the question :D
Please leave the feedback on this idea
Marco Agudelo2 years ago
J. Nikola did you notice the sarcasm it uses to aboard the subject. āThe quick brown fox jumps over the lazy dogā, remarcable. LoL
I find the subject very encouraging, since I consider AI grammar and synthesis good enough to trick anyone. In my opinion, the key point to look after would be the depth and writing style. I have read some writings (mostly opinions) that I doubt were written by persons only, but arenāt explicitly cited AI aids as itās required. Basically were those two criterias that alert me:
Non in-depth detail and lack of cited references of argued key concepts, that arenāt subject to opinion only.
Writing style without taking a position.
We should consider too, that you define the amount of words you want your AI to answer with, so it would be easy to try to trick a reader with AI supported paragraphs constructed, rather than just bulk a 1024 tokens-like-word paragraph. That would be even harder to notice. Wouldnāt you agree with me?
Also, when asked to AI with a non-closed information question the answer is vague too, and based on the given information the quality of the answer enriches (this is a key concept!!).
So, if (1) AI will generate non conclusive answers with non in-depth criterias when requested for long answers, because (2) taking a position implies short answers, I think the method for forgery AI detection should start with (-->) styling review, request for cited references when applied, and further criterias in this same line of thought.
jnikola Sep 10, 2022
[1]https://developers.cloudflare.com/ssl/edge-certificates/additional-options/tls-13/
[2]https://www.fortinet.com/blog/business-and-technology/tls-is-here-what-this-means-for-you
Please leave the feedback on this idea