Did a Robot Write This? We Need Watermarks to Spot AI

The better AI is used in tools like ChatGPT, the harder it is to differentiate between human text and machine-generated text. 

Highlight

1. OpenAI's ChatGPT can write articles, scripts and jokes in seconds
2. The next iteration of the AI tool is expected to launch in the coming months.
3. OpenAI's ChatGPT policy states that users must share content with a disclaimer

A talented scribe with impressive creative skills makes a sensational debut. ChatGPT, a San Francisco-based OpenAI text generator system, has been writing essays, scripts and limericks since its recent public release, usually in seconds and often at high rates. His jokes are also funny. Many scientists in the field of artificial intelligence are amazed at how human-like it is.

The better AI is used in tools like ChatGPT, the harder it is to differentiate between human text and machine-generated text.

But all these improvements come at a price. The better the AI, the harder it is to distinguish between human text and machine-generated text. OpenAI must prioritize its efforts to label the work of machines, otherwise we may soon be inundated with a confusing mix of real and fake information online.

For now, he asked people to be honest. According to OpenAI's policy for ChatGPT, when users share content from their system, they must clearly indicate that it is AI-generated in a way that "no reader can miss" or misinterpret.

Did you call that good luck

AI will almost certainly help kill the college essay. (A New Zealand student once admitted to using it to boost his grades.) Governments flood social media with propaganda, spammers write fake Amazon reviews, and ransomware gangs write emails. More reliable phishing attacks. No one points out the machine behind the truth.

And you'll have to take my word for it that this column is written entirely by a human.

AI-generated text needs some form of watermark, similar to how photo companies protect their images and movie studios discourage piracy. OpenAI already has a method for identifying another content generation tool called DALL-E with a signature embedded in every frame it generates. But finding the source of the text is very difficult. How do you put a secret, hard-to-remove tag on words?

In fact, he said, anyone using a GPT tool would find it difficult to remove the watermark signal, even by rearranging words or removing punctuation marks. The best way to overcome that is to use another AI system to paraphrase the output of the GPT tool. But it takes effort, and not everyone does it. In his speech, Aronson said he has a model that works.

But even assuming your method works outside of a lab environment, OpenAI is still confusing. Are you publishing the watermark keys or keeping them private?

If the keys were made public, teachers around the world could run their students' essays through special software to make sure they weren't machine-generated, as many are now doing the same verification. But it also allows bad actors to detect the watermark and remove it.

Meanwhile, keeping keys private creates a powerful business model for OpenAI: charging users for access. IT administrators can pay a subscription to scan incoming emails for phishing attacks, but universities can pay their professors pool fees and the price should be high enough to use the tool to discourage ransomware gangs and campaign writers. OpenAI essentially makes money by stopping the abuse of its own creations.

We also need to remember that tech companies don't have the best track records for preventing abuse of their systems, especially when they are unregulated and for-profit. (OpenAI says it's a for-profit/non-profit hybrid company, which limits future revenue.) But the strict filters OpenAI already has in place to prevent its text and image tools from generating objectionable content are a good start.

Now OpenAI should prefer a watermark system for your text. Our future looks set to be filled with machine-generated information, not only from the increasingly popular OpenAI tools, but also from the proliferation of pseudo-“synthetic” data used to train AI models and replace human-generated data. Images, videos, music and more are artificially produced to suit our hyper-personalized tastes.

However, our future selves may not care if a catchy song or cartoon comes from an AI. Human values ​​change with time; We care less about memorizing facts and means now than we did 20 years ago, for example. So, at some point, watermarks may not seem so necessary.