As we speak, AI can write poems, essays, analysis and scientific papers and likewise scripts for films with such sublimity that it typically turns into tough to evaluate whether or not an AI or a human authored the content material.
OpenAI, the corporate spearheading a lot of the AI growth and innovation in current occasions, has an answer to this—watermarking.
Inventory picture firms akin to Getty Photographs typically defend their photos with watermarks. A watermark may very well be a brand or textual content superimposed on a picture. Whereas it’s straightforward to watermark photograph or video content material, how does one watermark AI-generated textual content? Scott Aaronson, a researcher at OpenAI, has the reply.
Watermarking AI-generated content material
One of many predominant tasks Aaronson is engaged on at OpenAI is a software for statistically watermarking the outputs of a textual content mannequin like GPT.
“We wish it to be a lot more durable to take a GPT output and go it off as if it got here from a human,” he revealed whereas presenting a lecture on the College of Texas at Austin.
“For GPT, each enter and output is a string of tokens, which may very well be phrases but additionally punctuation marks, elements of phrases, or extra—there are about 100,000 tokens in complete. At its core, GPT is consistently producing a likelihood distribution over the following token to generate, conditional on the string of earlier tokens,” he mentioned in a weblog submit documenting his lecture.
So, each time an AI is producing textual content, the software that Aaronson is engaged on would embed an “unnoticeable secret sign” which might point out the origin of the textual content.
“We even have a working prototype of the watermarking scheme, constructed by OpenAI engineer Hendrik Kirchner.”
When you and I would nonetheless be scratching our heads about whether or not the content material is written by an AI or a human, OpenAI—who may have entry to a cryptographic key—would have the ability to uncover a watermark, Aaronson revealed.
Why is it the precise strategy?
On the finish of November 2022, OpenAI launched ChatGPT, a chatbot that interacts with people utilizing pure language. This new mannequin from OpenAI makes use of a novel coaching technique and relies on the GPT-3.5 structure.
We at Analytics India Journal have been making an attempt our hand as properly with the most recent software from OpenAI. Whereas there isn’t a doubt that it’s spectacular; in sure instances, it does present solutions that are factually incorrect or fully made up.
Earlier this month, Stack Overflow, the favored programming discussion board, banned all solutions created by ChatGPT, citing a excessive diploma of inaccuracy within the bot’s responses.
Regardless of being inaccurate at occasions, ChatGPT may be used to jot down participating essays, educational papers and even scripts. Lately, a Twitter consumer Ammaar Reshi posted that he mixed ChatGPT, Midjourney and different AI instruments to give you a kids’s e-book co-written by AI.
It then turns into vital for finish customers such as you and I to know whether or not the content material we’re consuming on the web or on social media websites is definitely written by people and which of them by an AI, and eat AI-generated content material with a pinch of salt.
Aaronson additionally admits that watermarking may very well be useful in stopping educational plagiarism. It might additionally show to be a software towards the mass technology of propaganda.
“, spamming each weblog with seemingly on-topic feedback supporting Russia’s invasion of Ukraine with out even a constructing stuffed with trolls in Moscow, or impersonating somebody’s writing model so as to incriminate them,” Aaronson mentioned.
OpenAI can be set to launch GPT-4 subsequent 12 months which is anticipated to push giant language fashions to their limits. Inside just a few years, half of the content material we learn on the web may very well be written by an AI, thereby making a watermarking software completely important.
Entry for all is crucial
Because it seems, solely OpenAI would have entry to the cryptographic keys which might assist decide whether or not the content material traces its origins to an AI or a human.
Nonetheless, it’s equally, if no more, crucial that most people additionally has entry to those keys to find out for themselves who the creator of the content material they’re participating with is.
Such a key would assist academics and professors decide whether or not the essays submitted by their college students are, in truth, penned by them and never by an AI. It might additionally assist scan emails for phishing assaults and social media websites for propaganda content material.
Nonetheless, freely giving the important thing, which solely OpenAI has entry to, without cost, would imply OpenAI would miss the possibility to make a revenue out of it. Additional, giving everybody entry to the keys would additionally imply the keys may very well be used to bypass or eliminate the watermark, which does depart OpenAI in a quandary.
Whereas it’s noteworthy that watermarking is among the many numerous alternate options OpenAI is exploring to take care of this concern, we must wait and see if OpenAI or another person manages to search out a solution to it that works properly for everybody concerned.
