I am going to show how we can easily and for free detect if a text has been written by an artificial intelligence or by a human, and if so, which artificial intelligence. For this we will use the GPTZero tool, which I will explain later.
The problem: AI is breaking institutions🏚️
Artificial intelligence is breaking institutions, but not all types of AIs, image and video generative AIs are not posing the same institutional problem as text generative AIs in the NLP field. In the near future it is likely that sound-generative AI’s will pose a similar problem, but at the moment the biggest institutional challenge are text-generative AI’s like ChatGPT, GPT4, Alpaca….
How are institutions breaking down?
In many ways, some of them are:
- 🎓 Educational centers: Universities, colleges and all kinds of educational institutions are seeing how assignments and homework are often being done by AIs, so that it is not possible to know if the student has cheated since the text is not copied but generated. Although the underlying problem or debate is how education should be reshaped based on the new paradigm(tasks that become productively useless such as basic cognitive tasks and memorization, etc. Although that does not mean that memorization should be completely abolished).
- 📝 Claims: AI-managed claims are being used on a massive scale, congesting the system.
- 📚 Books: Sites like Amazon are filling up with AI-written books, mostly with ChatGPT. This might not be a problem, but it would need to be analyzed at the very least.
- 👷Works: Many workers are using these tools in secret from their bosses. This in principle would not be something negative, because the important thing is the job well done, if a worker does it better this way, that’s it. The problem is when it is not corrected because ChatGPT makes mistakes.
- 💬 Chats: Typical chat scams, on chat websites but also on platforms like Facebook, are increasing, as it is much more scalable thousands of intelligent bots able to impersonate a human and understand the conversation having targets in “mind”. For example, you can ask ChatGPT to get a person’s number by pretending to be a woman, a policeman, etc. There are many cases of extortion by this means, with ChatGPT this can increase a lot, even so, ChatGPT is tied in short by the company OpenAI, but, what will happen when these tools are open source and free and instead of having 3 or 4 institutions or companies controlling these intelligences, they will run in tens of thousands of computers of individuals, organizations, etc.? ? Time is running out to find solutions.
Clearly the probem is served, being for the moment, ChatGPT the biggest threat and although it is limited, for example, OpenAI will not allow it to be used for scams when they detect it, they are hardly going to know if a student is using it to do homework and assignments.
💡There is a logical solution to all these problems: Detect texts written by AI. And the most famous tool in this field is GPTZero, an app capable of detecting texts generated by artificial intelligence.
How to detect text by an AI: GPTZero 🔍
The way to use it is very simple, we go to their website, and there we can either copy in a textarea the text we want to analyze or upload the PDF or the document to analyze.
But it will only let us use it a couple of times, then we can register with e-mail and password or with our Google account, etc. And then we will be able to use it for free in an almost unlimited way in terms of the number of analyses we can do. And with a maximum of 5000 characters to analyze each time.
In the next point I leave the payment plans, because if we pay a small monthly fee we can analyze up to 50 000 characters at a time, unlimited files, that is, instead of uploading a PDF to be analyzed, we can upload for example 30 PDF to be analyzed, saving us time. And most importantly, GPTZero’s payment plans allow us to use this fine-tuned tool to detect AI-generated text in the field of education, which as I said in the previous point, is currently the biggest institutional enforcement problem.
This fine-tuning has been achieved by training the tool in a final touch-up on exclusively educational texts.
GPTZero Features 🧰
Here are the different features of this tool according to the plan chosen.
Prices 💸
| Features | GPTZero Classic | GPTZero Educator | GPTZero Pro |
|---|---|---|---|
| Price | Free | $9.99/mo (discount starting at $19.99/mo) | $19.99/mo (discount starting at $29.99/mo) |
| For | All | Educators | Advanced users |
| Character limit per document | 5K | 50K | 50K |
| File limit per batch upload | 3 | unlimited | unlimited |
| Words per month | N/A | Up to 1 million | Up to 2 million |
| Detection Model | Free GPTZero detection model | Fine-tuned detection model for education | Access to GPTZero’s premium detection model with high limits through our dashboard application |
| API Access | No | No | No |
Sign up for GPTZero to use it for free or view plans.
Please note that these plans do not include API access. For API subscription plans, you must click on the link provided in your account once you have registered.
And yes, it has API, so it is very useful for developers.
GPTZero API 🍣
With the API you will be able to have integrated in your desktop or web software, calls to an app capable of detecting text generated artificially by AI. This can have many utilities, here are some examples:
- Chats💬: A chat app could integrate this tool to detect users registered as humans and not as bots, which are actually bots.
- Forums🗨️: Forums could detect fraudulent users who are SPAMming using artificial intelligences such as ChatGPT, using this tool.
- Online exams📝: It could also be integrated directly into the platforms that support exams or educational platforms such as moodle, or other variants created with plugins in CMS such as WordPress.
- Comments and General Spam✉️. Any WordPress user knows the amount of SPAM comments we receive from minute one. Plugins for WordPress and other platforms could integrate GPTZero or some similar tool to stop SPAM comments becoming more and more realistic and invisible to the moderator’s eye. The same with platforms that serve emails.
My opinion and limitations of GPTZero🤓
In my opinion GPTZero is a useful tool but only temporarily. Let me explain, we can use it to stop cheating in public and private educational institutions, and also to stop fraudulent uses on certain platforms. But this wall, this tool and any other similar tool, has its days numbered. Why? For three reasons:
- Economic interests: The battle against drugs has never ever been won, the incentive to receive the merchandise of millions of customers gives more economic power to the traffickers than we can give to their prosecution. The incentive to develop tools like GPTZero is great, yes, but nothing like the incentive to develop ever more powerful AIs, more difficult to distinguish from human handwriting, since we are talking about a business that includes: video game characters, websites and blogs, customer support tasks, books, etc. We are talking about billions of dollars, billions in English.
- Cat vs. Mouse: Right now, GPTZero has “signed up” AIs that can be problematic due to their power: ChatGPT, GPT-3, Bard, etc. But little by little, open source AIs are becoming more powerful, thanks in part to Facebook, which has open-sourced its models. Increasingly, every person, company, organization, etc. will be able to create their own modified models. It is going to be possible to create their own modified models in an increasingly simple way. And that would be a period to this type of tools that first need to analyze texts of each AI to detect it, if there are thousands and some internal ones that are not published, how it is going to detect it, it cannot.de.
- Final point: the moment an AI is able to imitate human handwriting, the game is over, no matter what tool we develop, we will not be able to detect it.
My experience using GPTZero
After using GPTZero I have found that sometimes it manages to detect that I have created the text with ChatGPT and also the function tuned to education is known to have a higher detection rate.
But what happens when I pass it a text generated by GPT-4? It doesn’t detect it, it marks 100% human every time. At the moment GPT4 can only be used under approval and with a monthly payment, so not all students, in fact almost none, use it, most use ChatGPT which is free, but it is a matter of maybe a year, I don’t know, that they end up using it, or that it ends up being free and used more, etc.
So my opinion, as I said at the beginning is that GPTZero is a useful tool temporarily.
