Generated by Human or AI: OpenAI to Watermark its Content

Today, AI can write poems, essays, research and scientific papers and also scripts for movies with such sublimity that it often becomes difficult to judge whether an AI or a human authored the content.

OpenAI, the company spearheading most of the AI development and innovation in recent times, has a solution to this—watermarking.

Stock image companies such as Getty Images often protect their images with watermarks. A watermark could be a logo or text superimposed on an image. While it is easy to watermark photo or video content, how does one watermark AI-generated text? Scott Aaronson, a researcher at OpenAI, has the answer.

Watermarking AI-generated content

One of the main projects Aaronson is working on at OpenAI is a tool for statistically watermarking the outputs of a text model like GPT.

“We want it to be much harder to take a GPT output and pass it off as if it came from a human,” he revealed while presenting a lecture at the University of Texas at Austin. 

Download our Mobile App

“For GPT, every input and output is a string of tokens, which could be words but also punctuation marks, parts of words, or more—there are about 100,000 tokens in total. At its core, GPT is constantly generating a probability distribution over the next token to generate, conditional on the string of previous tokens,” he said in a blog post documenting his lecture. 

So, whenever an AI is generating text, the tool that Aaronson is working on would embed an “unnoticeable secret signal” which would indicate the origin of the text.

“We actually have a working prototype of the watermarking scheme, built by OpenAI engineer Hendrik Kirchner.”

While you and I might still be scratching our heads about whether the content is written by an AI or a human, OpenAI—who will have access to a cryptographic key—would be able to uncover a watermark, Aaronson revealed.

Why is it the right approach?

At the end of November 2022, OpenAI released ChatGPT, a chatbot that interacts with humans using natural language. This new model from OpenAI uses a novel training method and is based on the GPT-3.5 architecture.

We at Analytics India Magazine have been trying our hand as well with the newest tool from OpenAI. While there is no doubt that it is impressive; in certain cases, it does provide answers which are factually incorrect or entirely made up. 

Earlier this month, Stack Overflow, the popular programming forum, banned all answers created by ChatGPT, citing a high degree of inaccuracy in the bot’s responses.

Despite being inaccurate at times, ChatGPT could also be used to write engaging essays, academic papers and even scripts. Recently, a Twitter user Ammaar Reshi posted that he combined ChatGPT, Midjourney and other AI tools to come up with a children’s book co-written by AI.

It then becomes important for end users like you and I to know whether the content we are consuming on the internet or on social media sites is actually written by humans and which ones by an AI, and consume AI-generated content with a pinch of salt. 

Aaronson also admits that watermarking could be helpful in preventing academic plagiarism. It could also prove to be a tool against the mass generation of propaganda. 

“You know, spamming every blog with seemingly on-topic comments supporting Russia’s invasion of Ukraine without even a building full of trolls in Moscow, or impersonating someone’s writing style in order to incriminate them,” Aaronson said.

OpenAI is also set to release GPT-4 next year which is expected to push large language models to their limits. Within a few years, half of the content we read on the internet could be written by an AI, thereby making a watermarking tool absolutely essential.

Access for all is imperative 

As it appears, only OpenAI would have access to the cryptographic keys which could help determine whether the content traces its origins to an AI or a human. 

However, it is equally, if not more, imperative that the general public also has access to these keys to determine for themselves who the creator of the content they are engaging with is.

Such a key would help teachers and professors determine whether the essays submitted by their students are, in fact, penned by them and not by an AI. It would also help scan emails for phishing attacks and social media sites for propaganda content.

However, giving away the key, which only OpenAI has access to, for free, would mean OpenAI would miss the chance to make a profit out of it. Further, giving everyone access to the keys would also mean the keys could be used to bypass or get rid of the watermark, which does leave OpenAI in a quandary.

While it is noteworthy that watermarking is among the various alternatives OpenAI is exploring to deal with this issue, we will have to wait and see if OpenAI or someone else manages to find an answer to it that works well for everyone involved.