How The ChatGPT Watermark Works And Why It Could Be Defeated

Posted by

OpenAI’s ChatGPT presented a way to immediately produce content but prepares to present a watermarking feature to make it easy to spot are making some people nervous. This is how ChatGPT watermarking works and why there may be a way to defeat it.

ChatGPT is an amazing tool that online publishers, affiliates and SEOs concurrently like and dread.

Some online marketers like it because they’re finding new ways to use it to produce content briefs, details and complex short articles.

Online publishers are afraid of the possibility of AI material flooding the search engine result, supplanting specialist short articles written by people.

As a result, news of a watermarking function that opens detection of ChatGPT-authored content is similarly prepared for with stress and anxiety and hope.

Cryptographic Watermark

A watermark is a semi-transparent mark (a logo or text) that is embedded onto an image. The watermark signals who is the initial author of the work.

It’s mainly seen in photos and increasingly in videos.

Watermarking text in ChatGPT includes cryptography in the form of embedding a pattern of words, letters and punctiation in the type of a secret code.

Scott Aaronson and ChatGPT Watermarking

An influential computer system researcher named Scott Aaronson was worked with by OpenAI in June 2022 to deal with AI Security and Positioning.

AI Security is a research field worried about studying ways that AI might present a damage to human beings and developing methods to prevent that sort of unfavorable disruption.

The Distill clinical journal, featuring authors associated with OpenAI, specifies AI Safety like this:

“The goal of long-lasting artificial intelligence (AI) security is to guarantee that sophisticated AI systems are reliably aligned with human values– that they reliably do things that people want them to do.”

AI Positioning is the artificial intelligence field worried about making sure that the AI is lined up with the desired goals.

A big language design (LLM) like ChatGPT can be utilized in such a way that may go contrary to the objectives of AI Positioning as specified by OpenAI, which is to create AI that advantages humanity.

Accordingly, the factor for watermarking is to avoid the misuse of AI in such a way that hurts humanity.

Aaronson described the reason for watermarking ChatGPT output:

“This could be helpful for preventing scholastic plagiarism, clearly, but likewise, for instance, mass generation of propaganda …”

How Does ChatGPT Watermarking Work?

ChatGPT watermarking is a system that embeds a statistical pattern, a code, into the choices of words and even punctuation marks.

Material created by expert system is created with a fairly predictable pattern of word choice.

The words written by human beings and AI follow an analytical pattern.

Altering the pattern of the words used in created material is a method to “watermark” the text to make it simple for a system to discover if it was the product of an AI text generator.

The trick that makes AI material watermarking undetectable is that the distribution of words still have a random appearance comparable to typical AI created text.

This is described as a pseudorandom circulation of words.

Pseudorandomness is a statistically random series of words or numbers that are not in fact random.

ChatGPT watermarking is not presently in use. However Scott Aaronson at OpenAI is on record stating that it is planned.

Today ChatGPT is in sneak peeks, which permits OpenAI to discover “misalignment” through real-world use.

Presumably watermarking may be introduced in a final variation of ChatGPT or earlier than that.

Scott Aaronson blogged about how watermarking works:

“My primary task so far has actually been a tool for statistically watermarking the outputs of a text model like GPT.

Essentially, whenever GPT produces some long text, we desire there to be an otherwise unnoticeable secret signal in its choices of words, which you can utilize to prove later on that, yes, this originated from GPT.”

Aaronson discussed further how ChatGPT watermarking works. However initially, it’s important to understand the idea of tokenization.

Tokenization is a step that takes place in natural language processing where the device takes the words in a document and breaks them down into semantic systems like words and sentences.

Tokenization changes text into a structured type that can be utilized in machine learning.

The process of text generation is the maker thinking which token follows based on the previous token.

This is made with a mathematical function that figures out the likelihood of what the next token will be, what’s called a possibility circulation.

What word is next is predicted but it’s random.

The watermarking itself is what Aaron describes as pseudorandom, in that there’s a mathematical reason for a particular word or punctuation mark to be there but it is still statistically random.

Here is the technical explanation of GPT watermarking:

“For GPT, every input and output is a string of tokens, which could be words however also punctuation marks, parts of words, or more– there are about 100,000 tokens in overall.

At its core, GPT is continuously generating a likelihood distribution over the next token to generate, conditional on the string of previous tokens.

After the neural net produces the circulation, the OpenAI server then really samples a token according to that circulation– or some customized version of the distribution, depending on a criterion called ‘temperature.’

As long as the temperature level is nonzero, though, there will generally be some randomness in the option of the next token: you might run over and over with the same prompt, and get a different conclusion (i.e., string of output tokens) each time.

So then to watermark, instead of picking the next token arbitrarily, the idea will be to choose it pseudorandomly, using a cryptographic pseudorandom function, whose secret is understood just to OpenAI.”

The watermark looks entirely natural to those checking out the text due to the fact that the choice of words is imitating the randomness of all the other words.

But that randomness contains a predisposition that can only be discovered by someone with the secret to decode it.

This is the technical description:

“To show, in the diplomatic immunity that GPT had a bunch of possible tokens that it evaluated equally possible, you could just select whichever token made the most of g. The option would look uniformly random to someone who didn’t understand the secret, however somebody who did know the secret could later sum g over all n-grams and see that it was anomalously large.”

Watermarking is a Privacy-first Solution

I’ve seen discussions on social networks where some people recommended that OpenAI could keep a record of every output it creates and utilize that for detection.

Scott Aaronson confirms that OpenAI could do that but that doing so postures a privacy concern. The possible exception is for law enforcement situation, which he didn’t elaborate on.

How to Discover ChatGPT or GPT Watermarking

Something intriguing that seems to not be popular yet is that Scott Aaronson noted that there is a way to defeat the watermarking.

He didn’t state it’s possible to defeat the watermarking, he said that it can be defeated.

“Now, this can all be defeated with adequate effort.

For instance, if you utilized another AI to paraphrase GPT’s output– well fine, we’re not going to be able to find that.”

It looks like the watermarking can be beat, at least in from November when the above declarations were made.

There is no indication that the watermarking is currently in usage. However when it does enter usage, it might be unidentified if this loophole was closed.

Citation

Check out Scott Aaronson’s post here.

Included image by SMM Panel/RealPeopleStudio