Ok /scitards/ I got an idea but I'm borderline retarded so can someone actually tell me if it'll work. Basically after seeing the 100th tiktok of a student crying after being falsely accused of using AI for an assignment I came up with this idea>letters in bodies of text likely follow certain distribution or frequency based on how common they are>create software that generative AI must use when generating any piece of text>software works by generating a certain skew in the distribution/frequency of letters in the text it generates>this unique skew will be attached to a key>license out software to colleges that will scan assessments submitted and match distribution of every submission to any keys that were generated using AIThe only way this will work though is with government support that enforces AI companies to use such software to offer their services in the country.
>ai detectionis fundamentally an arms raceyour idea is also on multiple levels retarded.for instance, how could it be enforced that everything goes through your filter?
>>16867790First movers advantage, by being the first to develop the software, big AI corps either gotta license it from you, acquire it, or try and develop it in house while being significantly behind you. This only works if the government can strong arm AI companies with regulation enforcing someway to detect outputs
>>16867793that's like untenable problem #943 out of #48790
>>16867793Sure and there should be trouble at all getting every government on earth to agree about a regulation.
The only true empiric work around is to do away with homework and online essays and make them write out their thoughts on paper with pen, in-class. God I’m glad I graduated before this shit came out and made everyone’s degrees essentially worthless, all post-2022 degrees will be scrutinized when it comes to employment. And I actually learned shit back then
>>16867856Degrees were always worthless, its always been the nepotists who make up 80% of the high paying jobs, 10% people with sob stories and/or a bunch of dependents, and 10% extremely hard working no lifers who make up the rest.
>>16867789>uses a local modelnothing personnel prof
>>16867789Scott Aaronson worked for OpenAI for a couple of years implementing such an AI statistical steganographic watermark, but if I recall correctly, they haven't implemented it yet. And even if they implement it, there are alternatives that (maybe) are not watermarked.Pic related, 4chan is an anime board
>>16867789>100th tiktok of a student crying after being falsely accused of using AInever had a student cry about being falsely accused. instead, here's what i've noticed>i only call out cases i'm extremely confident on>the whiners use ai to whineon the rare offchance i falsely accuse a student, we've had calm interactions where they challenge the accusation (rightfully so) and prove the work is their own. meanwhile, the students who passively cry are the ones who cannot prove innocence and seek empathy instead of justice. because they know they're dead to rights.
>>16867983You likely would not need a watermark, there is certainly some intrinsic fingerprint from the model itself in there. Only question is what word count you would need to reliably identify it.
>>16867789Literally every one of my university exams required these.
>>16867790more ai bro obviously, good luck selling that one to anti-ai retards
