Hacker News Reader

Comment by chrsstrm | Hacker News Reader

In some alternate timeline an enterprising person at OpenAI has written a memo in which they requested a new service to be implemented. This service hashes all generated responses before they are sent to the end user and saves each hash in a db somewhere. After OpenAI’s products have become popular enough that plagiarism is a valid concern, OpenAI releases their (paid) generated response history API which allows you to provide a hash and get a response as to whether or not it exists in their database. This puts a stop to 85% of all low effort plagiarism and earns OpenAI a decent additional revenue stream.

View on HN

A minor edit in the document completely changes the hash. And if it doesn't, we agan have the question if this is just a false positive.

There are similarity based hashes.

There are non-cryptographic digest hashes that do not exhibit an avalanche effects.

But that would immediately scare away all the future potential OpenAI customers… why would OpenAI do that? If customers know that it’s possible to determine that a text came from ChatGPT with a 100% accuracy, then no one would use ChatGPT.

Why? This assumes the only market is for cheaters and those seeking ghost writers.

It is.

The world, society, is not prepared for a scenario where computers generate human quality copy. This is hugely disruptive and will be used by workers and students to cheat their classes and do their work for them for a while.

It's a disruptive technology.

Not the only market, but a big part of it. People who work on social media will definitely make use of ChatGPT… unless Twitter, FB, et. al start to label tweets and messages with “Written by ChatGPT”.

Same goes for anything that has to do with selling written words. Not a small market precisely.

Because it is for ghost writers or to act as the ghost writer. What good is a text generating AI otherwise if you're not generating text to present as though it's your own? Even if the text is accurate, it's still ghost writing unless the whole point is to specifically present an AI-generated piece of text as AI-generated (aka pure novelty). AI written articles, for example, aren't doing that.

There are many, many uses for a text generating AI other than for nefarious purposes. It can give suggestions for things to do, it can write short stories to amuse you, it can act as an RPG game master, it can help with language learning, and so on. Assuming that it is only for cheaters is like how the music and film industry assumed that the only purpose for ripping CDs and DVDs onto your computer was to pirate them, ignoring that many people wanted to have a media server to stream their legally purchased media within their own homes.

The use of generated content for professional or academic dishonesty is a very tiny subset of all potential use cases. Neither OpenAI or schools and corporations want to face the backlash from their customers if it is discovered that they allow generated content to be used deceptively. To say that OpenAI would scare away all their potential customers if there is no chance for cheating or deception with generated content is a gross overstatement. Being able to trace the origin of content (or anything for that matter, in the virtual or real world) is a cornerstone of how the world operates and removing the anonymity of generated content is no different.

You'd have to have the hashes based on small portions of the responses, in order to catch copies where the recipient made minor changes throughout.

Sounds like a reason to partner with turnitin.

Shazam but for text.

You read a good article so you wonder which bot made this.

You've heard it here first.

This sounds like a good opportunity for fuzzy hashes like SSDEEP. This ought to give enough flexibility to detect minor tweaks to output.

https://www.microsoft.com/en-us/security/blog/2021/07/27/com...

And then you have the issue of those small parts being so common that they could match even normal English. It's not an easy problem to solve :(

Changing a single letter in the response will defeat such a system.

Reply on news.ycombinator.com