Generative AIs made a major breakthrough into our daily lives in 2022 with the release of ChatGPT. Although AIs and deep learning models already existed before, this event was a major upheaval in recent years that gave impetus to the creation of new tools, some based on these generative AIs and others completely separate but also claiming to be AI.
An overview of three AI tools is given here. A more complete list of the various solutions available on the market can be found at https://www.futuretools.io/.
Sapling – AI detection
Jules Verne spoke in his « Autour de la Lune » of the eternal race between the cannon and the armor; in the same way, ever since AI text generators have existed, tools have sought to detect them. There are many reasons why one might want to check whether a text was produced with AI, the most common certainly being to verify whether articles were generated from scratch or written by hand, and in education, for teachers, to check whether the work submitted was indeed created by the student. Several tools exist; here we tested https://sapling.ai/ai-content-detector. The reason for choosing this tool is that it offers a free version. For similar but paid—and therefore also more powerful—tools, one can mention Winston AI, which claims an effectiveness of 99.99%. The operation of these two tools is similar. It is rather ironic that this too is a form of deep learning AI: the models are trained on a set of texts created by AIs in order to recognize them and learn their characteristics. In this way, it builds a model that « learns » how a generative AI composes its text.
Sapling, although it offers a free version, operates on a freemium model. The main product is free, but add-ons are available for a fee. For $25/month, the character limit is raised from 5,000 to 20,000 and other Sapling products, such as the ability to rephrase texts, are included. The full list can be found at : https://sapling.ai/pricing.
Overview of the Sapling user interface
To test the tool, we created 5 texts of slightly under 500 characters on the same theme. The five texts explain in which contexts and why one should or should not use a generative AI. To ensure the reliability of the tool, four of the texts were generated by AIs, namely MS Copilot, ChatGPT – free version, DeepSeek r1, and Llama 3.3 70B. The last text was written, in all honesty, in haste, by a human. For each AI, the same prompt was used. The fact that the text is so short adds an extra difficulty, both for the detection tool and for the AIs that must generate a concise explanation.
The prompt used : « Write a short text between 450 and 499 characters. Imagine you are a journalist for a major newspaper. The text must explain in which contexts and for what reasons generative AIs should/would be allowed to be used. Be very clear and concise. » Here are the texts used for the test :
Human text :
The use of generative AI such as ChatGPT is subject to controversy. As with any new technology, its use is initially criticized and presented as a kind of heresy. There are nevertheless legitimate uses. The use of LLMs to help draft emails or repetitive documents, reports for which the AI’s analytical capabilities are superior to those of the average person, translation or text summarization are among these areas.
Copilot text
Generative artificial intelligences, such as language models and image generators, should be allowed to be used in contexts where they bring significant added value. For example, they can be used to automate repetitive tasks, enhance creativity in artistic and advertising fields, or provide in-depth analyses in the research sector. Their use is justified when they increase efficiency, reduce costs, and enable innovations that would otherwise be impossible. However, it is crucial to frame them with strict regulations to prevent abuse and ensure ethical and responsible use.
ChatGPT text – free version
Generative AIs should be allowed in specific contexts: writing assistance, translation, artistic creation, or support for teaching. They are useful for saving time, stimulating creativity, or making certain knowledge more accessible. However, their use must remain regulated, particularly to avoid misinformation, protect copyright, and ensure transparency of the content produced.
Deepseek R1 text
Generative artificial intelligences, such as ChatGPT or MidJourney, raise ethical and legal debates. Yet their use should be allowed in specific contexts: assisted creativity (writing, design), education (personalized tutoring), or optimization of repetitive tasks (translation, data synthesis). These tools boost productivity and innovation, provided they are supervised by humans.
A total ban would be counterproductive, but a strict framework is necessary: transparency regarding the origin of content, respect for copyright, and the fight against misinformation. AIs must remain auxiliaries, not substitutes.
Llama 3 70B
Generative artificial intelligences (AIs) are revolutionizing the way we work and live. In certain contexts, their use should be encouraged. For example, in healthcare, AIs can help doctors diagnose more quickly and accurately, thereby saving lives. In education, they can personalize learning for each student, thus improving their understanding and results. In the environment, AIs can analyze vast amounts of data to identify the most effective solutions for reducing carbon emissions and preserving biodiversity. However, it is essential to put in place strict regulations to prevent abuse and ensure transparency in their use.
The results :
Copilot
The tool successfully detected the text as AI-generated
ChatGPT
The tool successfully detected the text as AI-generated
Deepseek r1
Sapling failed to detect the text generated by DeepSeek
Llama 3 70B
Sapling successfully detected the text as AI-generated
Human text
Sapling correctly says that the text was not generated by AI
The results are not truly conclusive. We see that the tool did succeed in detecting three of the four AIs. However, it completely failed on the text by DeepSeek r1. It is also interesting to note that it is prone to giving false positives. Although it succeeded in recognizing the first human text as not having been generated, it completely failed on another text that came from an excerpt of a general dissertation on Simone de Beauvoir. The latter having been written by a human, the tool recognized it as an AI creation. In order to have a true estimate of this tool’s reliability, more examples should be taken, with texts of different lengths and above all on diverse and varied themes. Nevertheless, this race between the cannon and the armor is constantly evolving, and as long as new models are released, detectors will continue to be trained, which will gradually change their detection quality. Even this text you are reading is falsely detected as being 79.9% AI. It would therefore be wise, if you really need to know whether a text was made with AI—if you are a school teacher, for example—to turn to paid solutions.
Although the text was written by a human, Sapling gives a false positive and detects it as generated.
BgBye – Removes the background from videos and photos
Removing the background from a video or photo is a recurring problem for anyone working with media. Although just a few years ago, the use of a green/blue screen was necessary to replace the background of a video, and for photos manual cutout in Photoshop was the only option, today AI-based tools exist to greatly simplify your life. Here, we present one of them: BGBye. Behind its simple interface and its explanation in Comic Sans, this tool brings together in one place 10 AI-based algorithms for removing the background from a media file. The website is a personal and independent project, which means that from time to time the site is down. A version that you can run locally on your own machine—computer or server—is available.
The process with a simple image is fairly suitable and effective. On the other hand, the tool had issues with an image that had a more complex background and was larger (about 5 times bigger). None of the tools succeeded in removing the background, which is disappointing. This may be due to the limited power of the servers on which the site is hosted—it lacks the power to process such large images. In that case, the solution would be to self-host it on a powerful machine or to lower the image quality.
Another image was tested. The result is fairly satisfactory apart from the sun that creeps into the final result, but that is fairly simple to correct. This particular image is rather amusing to process given that the sun, which is part of the background, is sharper than certain parts of the woman in the foreground. The sharpness is also not homogeneous in the foreground.
This is the best version, chosen from among the versions of the 10 models. There are other proposals that are a little less successful; as a result, not being tied to a single algorithm makes it possible to choose the one that performs best in a given situation.
For videos, we do not have the option to see the results of the ten models before choosing which is best. When launching the process, you are asked to choose a specific model, without knowing the potential result in advance. Videos being by definition much larger, the process is much, much longer. Moreover, in the web version, video duration is limited to 10 seconds. The result itself is far from perfect, but we can see that the technology is on the right track. It is possible that the result would have been better with another model.
In short, the tool is fairly powerful and flexible; the fact that it is open source is a major plus. Nevertheless, one can still note that the interface is a little « outdated » and that there are reports that the online service is at times « down », which is rather understandable for an individual project.
Uniscribe – audio and video to text conversion
Several times already, when using text transcription features, whether in Word or on the phone’s dictaphone, I have had to rewrite or correct the rendered text by hand. For short audio excerpts, this is not really a problem, but as soon as it concerns longer passages, the lack of reliability of these tools starts to become inconvenient.
Since then, transcription tools have improved considerably. UniScribe is a tool specifically designed for this role. It offers a free subscription, with 120 minutes of transcriptions per month and a maximum of 5 files per day. Other paid subscriptions, « Basic » and « Pro », are available. These increase the amount of time allocated and use a faster and more efficient model.
To test the tool, we recorded a reading of a text, in this case part of Isaac Asimov’s Wikipedia page.
UniScribe interface
The result is quite impressive. There are only two or three minor errors in the text, due to poor pronunciation on my part. The AI perfectly transcribed the text and managed to recognize the English words.
But its capabilities do not stop there. The AI simultaneously creates a summary of the transcription and provides the main points of what was said. In an academic context, this can be particularly useful. It is also promised that a feature that will allow the user to ask questions to an AI, which will answer with responses drawn from the transcript, will be implemented.
It is nevertheless important to note that export in Word, PDF, CSV, SRT and VTT formats for subtitles is reserved for the paid versions. With the free version, we have to make do with a simple .txt file. It should also be noted that the tool, at least in its free version, is unable to separate the different speakers. In the case where you were to transcribe an interview, this would mean that you would have to separate the different parts by hand. This remains shorter and less tedious work than rewriting everything by hand.
Complete AI transcription for approximately 1 min 36 s of audio :
Isaac Asimov, pronounced Isaac Asimov, is an American writer born around January 2, 1920 in Petrovichi, Russia, and died on April 6, 1992 in New York. Also a professor of biochemistry at Boston University, he is best known for his science fiction works and his popular science books. A prolific writer, Asimov wrote or edited more than 500 books and answered approximately 990,000 letters and postcards. These books were published in nine of the major categories of the Dewey Decimal Classification. Asimov’s most famous work is the « Foundation » series, where the first three books won the only Hugo Award for Best All-Time Series in 1966. His other major series are the Empire series and the Robot series. As such, he is often considered along with Arthur C. Clarke and Robert A. Heinlein as one of the three great English-language science fiction authors. He also wrote important works of mystery fiction and fantasy, as well as numerous non-fiction works. Most of these popular science books explain concepts in a historical manner, going back as far as possible to a time when the science in question was at its simplest stage. Examples include the Guide to Science, the three-volume Understanding Physics series and Asimov’s Chronology of Science and Discovery, and he wrote on many other scientific and non-scientific subjects, such as chemistry, astronomy, mathematics, history, biblical exegesis and literary criticism.
UniScribe also offers to directly take a YouTube link and transcribe the audio of that video. In our tests, this works quite well, and the additional benefit of having a summary on the side and the key points greatly facilitates understanding and note-taking, if for example the video being watched/transcribed is for educational purposes.
Conclusion
In conclusion, of the 3 tools presented above, Sapling is the only one that fell short of initial expectations. The other two, although BgBye is a personal project, offer fairly convincing solutions with satisfactory finished products, even though BgBye’s video feature still needs improvement. UniScribe is a finished, effective product, and its use could be considered in a school or professional context.
