Did a robot write this? We need watermarks to spot AI

A talented scribe with amazing creative abilities makes a sensational debut. ChatGPT, a San Francisco-based OpenAI text-generating system, wrote essays, scenarios, and limericks after its recent public release, usually within seconds and often at a high level. Even his jokes can be funny. Many scientists in the field of artificial intelligence have marveled at how human-like it looks.

And remarkably, it will get better soon. OpenAI is expected to release its next iteration known as GPT-4 in the coming months, and early testers say it’s better than anything that came before it.

But all these improvements come at a price. The more AI improves, the harder it will be to distinguish between human text and machine-generated text. OpenAI must prioritize its efforts to label the work of machines, or we may soon be overwhelmed by a confusing mishmash of real and fake information online.

For now, the onus is on people to be honest. OpenAI’s policy for ChatGPT states that when sharing content from its system, users must clearly indicate that it is AI-generated “in a way that no reader could miss” or misunderstand. .

To that I say 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 will use it to flood social media with propaganda, spammers to write fake Amazon reviews, and ransomware gangs to write emails. more convincing phishing emails. None will point to the machine behind the curtain.

And you just have to take my word for it that this column was also written entirely by a human.

AI-generated text desperately needs some sort of watermark, similar to how photo companies protect their images and movie studios discourage piracy. OpenAI already has a method for flagging another content generation tool called DALL-E with an embedded signature in every image it generates. But it is much more difficult to trace the provenance of the text. How to put a secret and difficult to remove label on words?

The most promising approach is cryptography. In a guest lecture last month at the University of Texas at Austin, OpenAI research scientist Scott Aaronson gave a rare insight into how the company could distinguish text generated by the tool even further. more human GPT-4.

Aaronson, who was hired by OpenAI this year to complete the provenance challenge, explained that words could be converted into a string of tokens, representing punctuation marks, letters, or parts of words, or about 100,000 tokens in total. The GPT system would then decide on the layout of these tokens (reflecting the text itself) in such a way that they could be detected using a cryptographic key known only to OpenAI. “It won’t make any detectable difference to the end user,” Aaronson said.

In fact, anyone using a GPT tool would have a hard time erasing the watermark signal, even by rearranging words or removing punctuation marks, he said. The best way to beat it would be to use another AI system to paraphrase the output of the GPT tool. But it takes effort, and not everyone would do it. In his talk, Aaronson said he had a working prototype.

But even assuming its method works outside of a lab environment, OpenAI still has a dilemma. Does he release the watermark keys to the public or hold them privately?

If the keys are made public, professors around the world could run their students’ essays through special software to make sure they’re not machine-generated, the same way many are doing now to check the plagiarism. But it would also allow bad actors to detect the watermark and remove it.

Meanwhile, keeping the keys private creates a potentially powerful business model for OpenAI: charging users for access. IT administrators could pay a subscription to scan incoming emails for phishing attacks, while colleges could pay group fees for their professors – and the price for using the tool should be enough high to discourage ransomware gangs and propaganda writers. OpenAI would essentially make money by ending the misuse of its own creation.

We also have to keep in mind that tech companies don’t have the best track records for preventing misuse of their systems, especially when they’re unregulated and profit-driven. (OpenAI says it’s a hybrid for-profit and not-for-profit company that will cap future revenue.) But the strict filters OpenAI already has in place to prevent its text and image to generate offensive content are a good start.

Now OpenAI needs to prioritize a watermark system for its text. Our future appears to be awash with machine-generated information, not just from OpenAI’s increasingly popular tools, but from a wider rise in fake “synthetic” data used to train computer models. AI and replace human-created data. Images, videos, music and more will increasingly be artificially generated to suit our hyper-personalized tastes.

It’s of course possible that our future selves don’t care whether a catchy song or cartoon is from the AI. Human values ​​change over time; we care much less now about memorizing facts and driving directions than we did 20 years ago, for example. So, at some point, watermarks might not seem so necessary.

But for now, with tangible value placed on human ingenuity that others pay for or rate, and with the near certainty that OpenAI’s tool will be misused, we need to know where the brain stops. human and where the machines begin. A watermark would be a good start.

© 2022 Bloomberg L.P.

---