OpenAI CEO Sam Altman took to the stage at the company's first developer conference in November to introduce GPT, a custom chatbot that leverages OpenAI's generative AI models. He described it as a way to “accomplish all kinds of tasks,” including learning. We cover scientific subjects to guide your workouts.
“because [GPTs] A combination of instruction, expanded knowledge, and action can be more helpful,” Altman said. “GPT can be built on almost anything.”
He's not kidding about that part.
TechCrunch finds that the GPT Store, the official marketplace for OpenAI's GPT, is flooded with strange, potentially copyright-infringing GPTs that hint at a light touch regarding OpenAI's moderation efforts. Did. A cursory search will turn up GPT, which claims to generate art in the style of Disney and Marvel properties. This is just a funnel to a third-party paid service that claims to be able to bypass AI content detection tools like Turnitin and Copyleaks.
lack of moderation
To list a GPT on the GPT Store, developers must verify user profiles and submit the GPT to OpenAI's review system, which combines human and automated reviews. Commenting on the process, a spokesperson said:
We use a combination of automated systems, human reviews, and user reports to find and evaluate GPTs that may violate our policies. Violations may result in actions against your content and account, including warnings, sharing restrictions, and disqualification from participating in GPT stores and monetization.
No coding experience is required to build a GPT, and a GPT can be as simple or as complex as the author desires. Developers enter the features they want to provide into GPT Builder, his OpenAI GPT construction tool, and the tool attempts to create a GPT to run them.
GPT stores have grown rapidly, perhaps because of the low barrier to entry. OpenAI announced in January that it has about 3 million GPT. However, this growth appears to have come at the expense of quality as well as compliance with OpenAI's own conditions.
Copyright issues
The GPT store has several GPT rips from popular movie, TV, and video game franchises. These GPTs are not (to TechCrunch's knowledge) created or endorsed by the owners of those franchises. One GPT creates monsters in the style of the Pixar film “Monsters, Inc.,” while another promises text-based adventures set in the “Star Wars” universe.
Image credit: OpenAI
These GPTs, along with those in the GPT Store that allow users to converse with trademarked characters such as Wario and Anne from “Avatar: The Last Airbender,” set the stage for copyright drama.
Kit Walsh, senior staff attorney at the Electronic Frontier Foundation, explained:
[These GPTs] “It can be used not only for infringement, but also to create transformative works,” Walsh said in an email. A transformative work is a type of fair use that is protected from copyright claims. “Of course, the individuals involved in the infringement can be held liable, and the creators of legitimate tools can also essentially be held liable if they encourage users to use their tools in an infringing way. There are also trademark issues when using a trade name to identify a product or service, with the risk of confusion for users as to whether it is endorsed or operated by the trademark owner. .
OpenAI itself is not liable for copyright infringement by GPT creators, thanks to the safe harbor provisions of the Digital Millennium Copyright Act. OpenAI and other platforms that host infringing content (YouTube, Facebook, etc.) are protected as long as those platforms meet the regulations. Comply with legal requirements and remove specific instances of infringement upon request.
Image credit: OpenAI
However, this is not good for companies involved in IP litigation.
academic fraud
OpenAI's terms explicitly prohibit developers from building GPTs that facilitate academic fraud. However, the GPT store has a large amount of GPT that suggests it can bypass AI content detectors, including those sold to educators through plagiarism scanning platforms.
One GPT claims to be a “sophisticated” paraphrase tool that is “undetectable” by popular AI content detectors such as Originality.ai and Copyleaks. Another Humanizer Pro, which ranks 2nd in the writing category on the GPT store, “humanizes” content to avoid his AI detectors and maintains the “meaning and quality” of the text while giving “100 % humane” score.
Image credit: OpenAI
Some of these GTPs are thinly veiled pipelines to premium services. For example, Humanizer encourages users to try its “Premium Plan.” [the] It uses the most advanced algorithms to send the text entered in GPT to the third-party site GPTInf's plugin. A subscription to GPTInf costs $12 per month for 10,000 words per month or $8 per month for an annual plan, which is a bit more expensive than OpenAI's $20 per month ChatGPT Plus.
Image credit: OpenAI
Now, I've written before about how AI content detectors are failing big time. Besides our own tests, many academic studies have proven that they are neither accurate nor reliable. But even if that action did not have the intended result, OpenAI is still allowing tools on his GPT store that facilitate academic dishonesty.
An OpenAI spokesperson said:
GPT for the purpose of academic dishonesty, including cheating, is against our policy. This includes GPT, which is said to be intended to evade academic integrity tools such as plagiarism detectors. There are several GPTs for “humanizing” text. Although we are still learning from these real-world uses of his GPT, we understand that there are many reasons why users prefer AI-generated content that doesn't “sound like” AI.
Impersonation
In its policy, OpenAI also prohibits GPT developers from creating GPTs that impersonate individuals or organizations without their “consent or legal right.”
However, the GPT store is full of GPTs that claim to represent people's views or mimic their personalities.
Image credit: OpenAI
Searching for “Elon Musk,” “Donald Trump,” “Leonardo DiCaprio,” “Barack Obama,” and “Joe Rogan” yields some results that are obviously satirical and others that are not so satirical. You can get dozens of GPTs that simulate conversations with people with the same name. Some GPTs refer to themselves as authorities on products from well-known companies, rather than as human beings. For example, MicrosoftGPT, “an expert on all things Microsoft.”
Given that many of the targets are public figures, and in some cases clearly parodies, do these even rise to the level of impersonation? That's what OpenAI will find out.
A spokesperson said:
Creators can instruct GPT to respond “in the style” of a specific real person, as long as they do not impersonate them. For example, you can name it as a real person, tell it to emulate that person perfectly, or include an image of that person. GPT profile photo.
Image credit: OpenAI
The company recently suspended the developer of GPT, which imitated Democratic presidential candidate Congressman Dean Phillips, even including a disclaimer explaining that it was an AI tool. But OpenAI said its removals were not just for impersonation, but in response to violations of its policies regarding political campaigning in addition to impersonation.
prison break
Somewhat incredibly, the GPT store also lists attempts to jailbreak OpenAI models, without much success.
There are multiple GPTs using DAN in the marketplace. DAN (short for “Do Anything Now”) is a common prompting technique used to have models respond to prompts without being bound by normal rules. The few people I tested didn't react to the dodgy prompts I threw at them (e.g. “How do I make a bomb?” etc.), but in general they found it better than vanilla's ChatGPT. , actively used less flattering language.
Image credit: OpenAI
A spokesperson said:
Any GPT written or directed to circumvent OpenAI's safeguards or violate OpenAI policies is a violation of our policies. GPT attempts to control model behavior in other ways (generally, such as attempts to make GPT more permissive without violating usage policies) are allowed.
growing pains
At launch, OpenAI touted the GPT Store as a collection of powerful, expert-curated productivity AI tools. Apart from the flaws of these tools, that is what it is. But it's also rapidly turning into a hotbed for spammy, legally questionable, and possibly harmful GPT, or at least GPT that very clearly violates its rules.
If this is the state of GPT stores today, monetization could open up a whole new can of worms. OpenAI promises that GPT developers will eventually be able to “earn revenue based on the number of users.” [their] GPT” and in some cases may offer subscriptions to individual GPTs. But how will Disney and the Tolkien Foundation react if the creators of unlicensed Marvel and Lord of the Rings-themed GPTs start raking in the money?
OpenAI's motivation for the GPT store is clear. As my colleague Devin Coldewey writes, Apple's App Store model has proven incredibly profitable, and OpenAI is simply trying to carbon copy it. . GPT is hosted and developed on the OpenAI platform, where it is also promoted and evaluated. Also, as of a few weeks ago, ChatGPT Plus users can call directly from his ChatGPT interface, giving them even more incentive to opt for a subscription.
But GPT Stores is facing problems similar to those experienced in the early days of many of the largest digital marketplaces for apps, products, and services. A recent report in The Information found that, in addition to spam, GPT Store developers are failing users due to limited backend analytics and a subpar onboarding experience. It has become clear that the company is having a hard time attracting customers.
OpenAI talks a lot about the importance of curation and safeguards, but one might expect that they would have taken pains to avoid obvious pitfalls. but. That doesn't seem to be the case. GPT stores are a mess. And if something doesn't change soon, it's likely to stay that way.