Google detects and downranks pages that have mostly AI copy. And from what I can tell, OpenAI are working on watermarking their generated content so other platforms can detect it automatically.
> OpenAI are working on watermarking their generated content so other platforms can detect it automatically
So when is this technology gonna be advanced enough for us mere mortals to run it on our own computers with our own models free from pointless restrictions that exist purely to appease threatened stakeholders?
Trained on what though? Procuring training material ain't no trivial feat., in-fact it might be even more elaborate to do. This is where Google has home advantage. They have their own copy of the internet.
To be fair, training is a lot more intensive than inference. Though if open source models are any indication, the big issue is actually VRAM requirements.
The restrictions are not pointless. People shouldn't have to endure AI generated spam, no matter how well-formatted.
I can opt out of robocalls, email spam, programmatic texts, even take legal action against people who make them outside the law. Similar legal protections should exist to protect people from AI spam. If there's a good technical solution to this problem then the laws banning it become less necessary.
> This is some real text. It was not written by a robot. It was written by a human. If you don't believe me, ask the guy who wrote it. He will tell you that he wrote it using his brain and fingers. Do you believe me?
The page says it needs 50 tokens to start getting accurate results and that the above text has 54. It also rates it as 99.93% fake.
Spotting robots may be easy. Spotting humans I think is the hard part.
Also, in many use cases that have been described thus far it's a collaboration between the AI and a human (i.e. the AI writes the first copy, the human edits). That blurs the line even further.
I read a blog post by Scott Aaronson saying they can watermark by using a cryptographically biased random number generator when selecting which token to output from the stream. I didn't entirely follow how that helps but a little bias per token adds up pretty quickly.
I'd say the arms race of AI- detection/detection-avoidance is inevitable.
But it seems we're a way off given how bad common spam detection tools are today (I mean, it's likely that a lot of spam is written using detection-avoidance tech but that seems to make it even more obviously spam to a human reader...)
I think gmail's spam filter is so close that it feels like a solved problem. But of course it has a lot of other indicators to look at than just the email body text and spam tends to revolve around a fairly predictable range of topics. None of that would apply in a "was this webpage written by a robot?" algorithm
I just logged in to check my gmail account (I don't use it that often) - and literally every single message in the last few months is spam, none were flagged as such. I didn't even need to open the message to determine that they were spam, simply the combination of subject line and sender name/address were enough.
So why is AI so bad at this?
(The sender name for virtually every single message is either "Lowe's winner", "Kohl's winner", CVS or some big US chain that doesn't even operate in my country. The actual email addresses are even more obviously dodgy. There is one genuine message from DELL asking about a survey related to a product I ordered from them many many months ago. I basically use gmail as my spam trap...I get far far less spam on my hotmail account that's my regular one, but the ones that do get through don't even pass the most basic tests, which I don't understand).
Edit: actually the bigger problem with hotmail is false positives. I just tried ChatGPT out on one such example, and it definitely did a better job:
'It is unlikely that an email beginning with the text "Hi All, Here are a few things you need to know for the Christmas concert" is spam.'
I tried a few others (both genuine and incorrectly-identified spam) and for each attempt ChatGPT got it right just from the subject/first paragraph.
With a bit more questioning I could also get it to identify messages that were "genuine" marketing promotions (from companies I've bought products from and agreed to receiving such messages - which hotmail still identifies as junk)
I assume most people would not be that sophisticated, but I understand the point. This seems like it would be an ongoing battle no matter what OpenAI does though.
That will almost certainly preserve the invisible characters. Most invisible characters are used for some kind of in-line formatting in Unicode, so it's not desirable to remove them.
I've already got a script running every 2500 milliseconds to strip leading and trailing whitespace, HTML, and non-ASCII characters except for the UTF-8 characters of our local language.
There are several linguistic patterns/repetitions that might seem completely normal to us, but are easily detected as deliberate if you search for them using software.
To me it makes sense that openai would store all input and output and could make something like a plagiarism checker page/API to search and see if openai generated that or very similar content... Another service they could sell to teachers and such.
So when is this technology gonna be advanced enough for us mere mortals to run it on our own computers with our own models free from pointless restrictions that exist purely to appease threatened stakeholders?