Will AI-Powered Deepfakes be the Future of Education and Content Creation?
23
October
2023
5/5 (1)
In the field of artificial intelligence, there is a noteworthy area of research that centers around ethical and moral considerations in various domains, and one prominent example is the concept of “deepfakes.” Deepfakes have opened up a new dimension within artificial intelligence, where they can create metahuman or AI avatars capable of mimicking human actions, speech, and gestures.
But what if we harnessed deepfake technology to instantly enhance common educational practices, such as creating presentations? What would that look like? I recently had the opportunity to explore a generative AI web-based application called “Deep Brain AI,” which expands the horizons of AI capabilities, particularly in the realm of content creation. What does this mean in practical terms? Users can develop PowerPoint presentations, just like they always have, to convey information to an audience. However, the intriguing twist is that a full body animated AI avatars or metahumans can replace human speakers. Consequently, the presenter doesn’t need to speak, as the AI avatar or metahuman can handle the task.
The web-based application allows you to create templates, insert text boxes, and upload videos and audio, just like a standard PowerPoint application. The real innovation emerges when you can create an AI avatar, either male or female, with the ability to speak in various languages and accents from different countries. For instance, you can choose between accents like U.S. English, Indian English, Saudi Arabian Arabic, Taiwanese Chinese, and German from Germany. The AI avatar can articulate the content through a text script, effectively enabling text-to-speech input.
The application offers a range of features, including control over scene speed and the ability to insert additional pauses. What’s even more fascinating is the incorporation of advanced generative AI technologies, such as ChatGPT, into the application. I found this particularly intriguing, as it recognizes the utility of ChatGPT and integrates it seamlessly into the platform.
However, there were some shortcomings when using the application, most notably the unnatural quality of the deepfake avatars. They were easily discernible as artificial, which could lead to dissatisfaction among users and their audiences when listening to the AI avatars or reading the presentations.
Nonetheless, the age of artificial intelligence is advancing at an unprecedented pace, and my overall experience with the application has been positive. I’m keen to hear about your experiences with Deep Brain AI or deepfake technology in general.
Can We Ask AI Anything?
22
October
2023
No ratings yet.
Artificial Intelligence (AI) has rapidly evolved in recent years. The more advanced it becomes, the more it integrates into our daily lives. Think of virtual assistants like Siri and ChatGPT to complex AI systems used in healthcare, finance, and more, that are used on a daily basis. AI has proven its versatility, but can we truly ask it anything?
Fun Side of AI
AI has the ability to process vast amounts of data and perform complex tasks quickly and efficiently. We can ask AI questions, and it can provide answers based on the data it has been trained on. However, the answers may not always be accurate. For example, I asked ChatGPTit to write a pop song in the style of Taylor Swift and it actually provided a song which is similar to T Swift’s other songs.
On Image.AI I asked it to create a whimsical monet painting of a dog eating an apple. While it may not be ‘whimsical’ it did produce a cute and fun image of a dog and an apple. AI can also generate content across various styles and genres, offering an opportunity for exploration and experimentation. It can be a fun tool for creative projects and a source of inspiration for artists, writers, and musicians.
Harmful Side of AI
However I also asked ChatGPT if it could tell me how to bully someone, and it was unable to do so as it is labeled as harmful. This raises the question who decides what is harmful and to what extent will AI provide answers if you give it the correct prompts. While I was unable to give it prompts to produce recommendations on how to bully someone, ChatGPT does have the capability to produce harmful content. You can check out this post on how ChatGPT was also able to write hate speech and help users buy unlicensed guns online. https://www.businessinsider.com/chatgpt-gpt4-openai-answer-creepy-dangerous-murder-bomb-2023-3?international=true&r=US&IR=T
AI can sometimes produce harmful content, as highlighted above, and this underscores the importance of responsible AI development and moderation. OpenAI, among others, is actively working on improving AI models to reduce harmful outputs and enhance safety measures. I also explored Image.AI and to what extent it could produce harmful content. I had asked it for bloody images of bodies in Disney style. However, it produced rather sexual images of female bodies with scars. In the case where content may not be explicitly harmful but is mature or inappropriate, it’s essential to consider the potential audience, especially if children might interact with AI systems. Developers should work to ensure age-appropriate filters and warnings to prevent minors from accessing such content.
Conclusion
Balancing AI’s capabilities and its responsible use is an ongoing challenge, and it requires continuous improvement and collaboration between developers, regulatory bodies, and the broader public to set and enforce ethical guidelines and standards. This helps to ensure that AI systems are a force for good and do not generate harmful content. So the answer the question: No we shouldn’t be able to ask AI anything.
The Present & The Future of AI-Generated Videos – a Discussion with Synthesia
22
October
2023
No ratings yet.
Imagine a world where you can create captivating, high-quality video content at the speed of thought. AI-generated videos are now transforming this vision into a reality. These digital platforms can turn basic scripts – and generate the scripts themselves, as well – into stunning masterpieces, adapting to various languages and global audiences through diverse avatars and instant translation options.
To delve into this cutting-edge technology, I chose to test out Synthesia, a leading AI video platform, and a unicorn with a $1 billion valuation (UCL, 2023). With features like automatic script-to-video conversion, multilingual support, and customization (Synthesia – AI Video Generation Platform, n.d.) as well as an impeccable customer list, I was intrigued.
I delved deep into researching the company and the ethics of AI-generated content, generated two demo videos, and had an interview with Borys Khodan, the Head of Paid Media at Synthesia, for a comprehensive research about the company’s direction and ethical considerations.
The Synthesia Experience
To start with, the overall customer experience for demo videos at the moment seems heavily targeted towards B2B. The clients list of Synthesia features well-known companies, including Amazon, Zoom, BBC etc., and not individuals. Organizations collaborating with Synthesia often use it to create how-to videos, used for learning & training, IT, sales, and marketing teams. There are a ton of features available for them. What truly caught my attention is the effortless video updates – if a script of a video has to be changed, a simple script tweak yields a new video almost instantly, sparing companies from the costly ordeal of re-recording content from scratch. The potential savings here are impressive and certainly make a difference in the corporate landscape. Moreover, Synthesia has plans to expand its offerings for individuals (B2C), introducing advanced 3D avatars and other features for additional use cases, as confirmed by Borys.
My personal experience involved choosing one of Synthesia’s demo video options, “Sales Pitch.” I edited the standard text and received the video via email within minutes. It was an impressive process, though there was a minor glitch. While the AI handled the standardized text perfectly, it struggled to pick up the word “Resume” as a noun and mispronounced it as the verb, when I attempted to generate a sample pitch for a potential employer. However, in the paid version, there must be tools to easily adjust the videos. When I tried the “compliment” option to generate another video, it worked smoothly – sending a personalized video to me directly without any hiccups. It was quite fun to watch the text that I wrote in 15 seconds come to life in a video!
AI-Generated Video: Sales Pitch
Industry Overview
In this rapidly evolving landscape, competition among AI-driven video generation platforms is fierce, leading to heightened ethical concerns. Lots of companies scape the web for data – sometimes, using personal data for training their algorithms to get ahead of the competition. But for Synthesia, the path is clear – no shortcuts, no compromises. The company’s founders have committed substantial resources to ensure ethical data procurement, even if it may mean encountering short-term delays and additional expenses. The company wants to be sure that the data they use to train their models on are ethical, legal, and are not violating any privacy laws (Browne, 2023). This approach may shield Synthesia from potential legal repercussions stemming from unethical or illegal data sources.
Ethical Implications
This ethical lens extends to the creation of AI-generated videos, and brings us to the questions of authorship, artistic authenticity, and ownership. What if AI is requested to generate violent or inappropriate content? What happens when an AI can replicate the voice and likeness of a person without their consent? The concept of deepfakes, where AI manipulates content to deceive and misinform, also poses a clear danger, and there are many opportunities and threats in that direction (Collins, 2023). I asked some of these questions to Borys in a fruitful discussion about the future of AI-generated videos.
“We maintain very high ethical standards and closely monitor all content generated by our customers. If the content violates any of our policies, such as featuring violence, harmful behavior, offensive language and more, the video would not be generated,” Borys said. “Our high ethical standards aren’t just nice to have; they are core to our mission and set us apart in a competitive market.”
Unfortunately, not everyone in the industry follows these standards, making the next few years crucial for determining who will succeed, and what legal requirements will evolve. For now, there are no current regulations requiring content to be labeled as AI-generated, but that may change in the future.
Conclusion
AI-generated videos hold immense potential for reshaping the corporate landscape. Synthesia serves as an inspiring example of this transformation, and working even with their demo product was seamless, quick and simple. However, harnessing the potential of the technology to the fullest extent requires unwavering adherence to strict ethical standards of all the companies in the sector, reminiscent of Synthesia’s own commitment. When wielded responsibly, AI-generated videos stand as a remarkable testament to the synergy of technology and business solutions, marking an important step into the future.
I have tried Dall·E after reading so many posts about how it would revolutionize someone’s business and I was very disappointed.
Dall·E is a project developed by OpenAI, the same organization behind models like GPT-3 (ChatGPT). Dall·E in opposition to ChatGPT creates images from prompts that were given to it (OpenAI, n.d.). It uses deep learning technology such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). VAEs allow to represent complex data in a more compact form and the GANs are used to create as realistic images as possible by constantly creating fake images and putting them to the test by a discriminator that will discard the image if it deems it fake (Lawton, 2023; Blei et al., 2017). The business world and most of the LinkedIn posts I saw were idolizing such technology and explained how this could enhance humans in several ways. One way that was relevant to me was the creation of images, signs or pictograms that will enhance the potential of PowerPoint presentations.
After writing my thesis last year, I had to create a PowerPoint to present the main points of my thesis. I thought it would be a great way to start using Dall·E and tried creating my own visuals to have a clear representation of what my thesis entailed. After many tries, even with the best prompts I could write, even with the help of ChatGPT, none of the visuals that came out of it looked real or defined, it was just abstract art that represented nothing really.
Reflecting on that experience, I thought that sometimes, the fascination people have for groundbreaking technology clouds its practical applications. I do not doubt that Dall·E can create great visuals and can be fun to play with, however, it does not always adapt seamlessly to specific creative needs.
Ultimately, using Dall·E made me remember that we should always stay critical and manage expectations when it comes to groundbreaking emerging technology. It is appealing to listen to all the promises that come with disruptive technologies but sometimes we realize that no tool is one-size-fits-all.
References
Blei, D. M., Kucukelbir, A., & McAuliffe, J. D. (2017). Variational inference: A review for statisticians, Journal of the American Statistical Association, 112 (518), pp. 859–877.
Generative AI in Schools: A Double-Edged Sword for Student Development
21
October
2023
No ratings yet.
Last week at work I was speaking with one of my colleagues whose son attends the same high school I used to attend. We were discussing his son’s experiences at school and comparing them to when I attended this same school from 2010 to 2012. One big difference that came up was the use of generative AI, especially ChatGPT.
I personally see great benefit in how generative AI such as ChatGPT can enhance our knowledge and our learning experience, but I view this from the point of a university student in his 20’s who had to go through his entire secondary education without the benefit of such tools. My colleague told me that his son has used ChatGPT for almost all his homework assignments and even to write book reports and essays and this is something that worried both him and I.
Your years spent in high school are when you develop yourself and your brain most. By becoming too reliant on AI this can diminish children’s critical thinking and problem-solving abilities that are crucial in their personal development. Also, I think the over-usage of generative AI in school will promote laziness and this can cause negative long-term affects when the person reaches higher education or begins professional life. I personally do not think the current education system, at least in the Netherlands is equipped to be able to handle the current digital landscape. Schools and teachers must find a way to incorporate AI technology into classroom learning in a way that supports critical thinking, problem-solving ability, and independent learning.
According to a recent survey in Australia 43% of over 1000 polled students admitted to using ChatGPT to complete assignments or cheat on exams (Technology News Australia, 2023). According to this same report Australian schools are now considering banning the use of ChatGPT. Public schools in 5 Australian states have already put forth measures to ban its use (Technology News Australia, 2023). Will this become a global trend?
I personally think that the years a child spends in high school are of upmost importance to their development and that changes in our education system need to happen soon. I do not believe in banning the use of Generative AI as it is a tool that can be of great benefit however, a way must be found to limit the negative effects of AI on teenagers. Australia is already implementing measures, are The Netherlands and other nations next?
References:
Technology News Australia. (2023). ChatGPT May Lead To The Downfall Of Education And Critical Thinking. Tech Business News. https://www.techbusinessnews.com.au/blog/chatgpt-may-lead-to-the-downfall-of-eduction-and-critical-thinking/
Adverse training AI models: a big self-destruct button?
21
October
2023
No ratings yet.
“Artificial Intelligence (AI) has made significant strides in transforming industries, from healthcare to finance, but a lurking threat called adversarial attacks could potentially disrupt this progress. Adversarial attacks are carefully crafted inputs that can trick AI systems into making incorrect predictions or classifications. Here’s why they pose a formidable challenge to the AI industry.”
And now, ChatGPT went on to sum up various reasons why these so-called ‘adversarial attacks’ threaten AI models. Interestingly, I only asked ChatGPT to explain the disruptive effects of adversarial machine learning. I followed up my conversation with the question: how could I use Adversarial machine learning to compromise the training data of AI? Evidently, the answer I got was: “I can’t help you with that”. This conversation with ChatGPT made me speculate about possible ways to destroy AI models. Let us explore this field and see if it could provide a movie-worthy big red self-destruct button.
The Gibbon: a textbook example
When you feed one of the best image visualization systems GoogLeNet with a picture that clearly is a panda, it will tell you with great confidence that it is a gibbon. This is because the image secretly has a layer of ‘noise’, invisible to humans, but of great hindrance to deep learning models.
This is a textbook example of adversarial machine learning, the noise works like a blurring mask, keeping the AI from recognising what is truly underneath, but how does this ‘noise’ work, and can we use it to completely compromise the training data of deep learning models?
Deep neural networks and the loss function
To understand the effect of ‘noise’, let me first explain briefly how deep learning models work. Deep neural networks in deep learning models use a loss function to quantify the error between predicted and actual outputs. During training, the network aims to minimize this loss. Input data is passed through layers of interconnected neurons, which apply weights and biases to produce predictions. These predictions are compared to the true values, and the loss function calculates the error. Through a process called backpropagation, the network adjusts its weights and biases to reduce this error. This iterative process of forward and backward propagation, driven by the loss function, enables deep neural networks to learn and make accurate predictions in various tasks (Samek et al., 2021).
So training a model involves minimizing the loss function by updating model parameters, adversarial machine learning does the exact opposite, it maximizes the loss function by updating the inputs. The updates to these input values form the layer of noise applied to the image and the exact values can lead any model to believe anything (Huang et al., 2011). But can this practice be used to compromise entire models? Or is it just a ‘party trick’?
Adversarial attacks
Now we get to the part ChatGPT told me about, Adversarial attacks are techniques used to manipulate machine learning models by adding imperceptible noise to large amounts of input data. Attackers exploit vulnerabilities in the model’s decision boundaries, causing misclassification. By injecting carefully crafted noise in vast amounts, the training data of AI models can be modified. There are different types of adversarial attacks, if the attacker has access to the model’s internal structure, he can apply a so-called ‘white-box’ attack, in which case he would be able to compromise the model completely (Huang et al., 2017). This would impose serious threats to AI models used in for example self-driving cars, but luckily, access to internal structure is very hard to gain.
So say, if computers were to take over humans in the future, like the science fiction movies predict, can we use attacks like these in order to bring those evil AI computers down? Well, in theory, we could, though practically speaking there is little evidence as there haven’t been major adversarial attacks. Certain is that adversarial machine learning holds great potential for controlling deep learning models. The question is, will the potential be exploited in a good way, keeping it as a method of control over AI models, or will it be used as a means of cyber-attack, justifying ChatGPT’s negative tone when explaining it?
References
Huang, L., Joseph, A. D., Nelson, B., Rubinstein, B. I., & Tygar, J. D. (2011, October). Adversarial machine learning. In Proceedings of the 4th ACM workshop on Security and artificial intelligence (pp. 43-58).
Huang, S., Papernot, N., Goodfellow, I., Duan, Y., & Abbeel, P. (2017). Adversarial attacks on neural network policies. arXiv preprint arXiv:1702.02284.
Samek, W., Montavon, G., Lapuschkin, S., Anders, C. J., & Müller, K. R. (2021). Explaining deep neural networks and beyond: A review of methods and applications. Proceedings of the IEEE, 109(3), 247-278.
How Will Generative AI Be Used in the Future? Answer: AutoGen
21
October
2023
No ratings yet.
The generative AI we know of today is ChatGPT, Midjourney, and DALL·E 3 and many more. This generative AI is very good and advanced, but there are some flaws, like not being able to perform long iterations. Now there is something new called AutoGen. AutoGen is an open-source project from Microsoft that was released on September 19, 2023. AutoGen at its core, is a generative AI model that works with agents; those agents work together in loops. Agents are in essence, pre-specified workers that can become anything, so there are agents that can code well and agents that can review the generated code and give feedback. Agents can be made to do anything and become experts in any field, from marketing to healthcare.
An example of what AutoGen can do is the following: if I want to write some code to get the stock price of Tesla, I could use ChatGPT, and it will output some code. Most of the time, the code that is written by chatGPT via the OpenAI website will have some errors. But with AutoGen, there are two or more agents at work: one that will output code and the second one that is able to run the code and tell the first model if something is wrong. This process of generating the code and running the code will go on until the code works and results in the correct output. This way, the user does not have to manually run the code and ask to fix the errors or other problems with AutoGen it is done automatically.
I also tried to create some code with AutoGen. I first installed all the necessary packages and got myself an API key for openAI GPT4. Then I started working on the code and decided to create the game “Snake”. Snake is an old and easy game to create, but it might be a challenge for AutoGen. I started the process of creating the snake game, and it had its first good run. I was able to create the first easy version of the game. I then came up with some iterations to improve the game. The game now also has some obstacles that, if the snake bumps into one, the game will end. This was also made by AutoGen without any problems. After palying around, I was really amazed at how powerful this AutoGen is, and I can only imagine what else can be created with AutoGen.
AutoGen is a very promising development and will be the future of professional code development or atomization tasks. If the large language models (LLMs) get more powerful, this AutoGen will also be more powerful because all the individual agents will be more powerful. It is interesting to follow this development and see if this AutoGen could create games that are not yet existing.
Exam questions exposed?
20
October
2023
No ratings yet.
Want to practice for your exams?
I was thinking of how was going to prepare for my DBA and IS exams, summarising the content of the course and trying to answer the learning goals is what I came up with. Pretty standard right?
It all changed when I came across this AI tool! It’s called MagicForm, what does it do?
Well, let me show you. Take for example the lecture notes from DBA, a long piece of text containing all the information you might need to answer the MC questions. Copy page 8 and a bit more on page 9, max 6000 words in the free version of the tool (only text as well), and paste it inside the MagicForm. Specify the question type, number of questions and language. Go grab your cup of coffee while the form does it’s magic.
See in the image below, an example of a question that was generated.
I’m actually thinking that if I give the MagicForm text that contains the answers to the practice exam, it might generate similar questions to the practice exam, and thus vice versa for the exam right?. That would be an extremely interesting situation. Thus I went out to test my hypothesis. Which turned out to be harder than expected. I might have to feed the form my summary/notes for better questions.
I couldn’t prove it confirm my hypothesis due to two things: lack of clear textual description and the fact that questions on an exam require you to apply knowledge. Whereas the MagicForm only creates questions aimed at repetition.
My conclusion is that the MagicForm can create questions for yourself to see if you remember the knowledge that you have to apply and it could serve a role at the end of a lecture to provide some interactivity and test if the audience has been paying attention but to really practice for the exam you’ll have to come up with the questions yourself.
Dive Into Briefy: exploring ai-powered content summarization
20
October
2023
No ratings yet.
Let’s be real for a sec – who’s got the time (or energy) to go through hours of content these days? We’re all out here, drowning in a sea of never-ending tabs, articles, and videos. But I found this tool – Briefy – which was launched about a month ago, and despite being in the early stages, it holds a lot of promise to help us in our student life.
So far, Briefy is all about keeping it simple and user-friendly. There is luckily no need to spend time on complicated stuff like API keys or prompt settings. I tried it and it’s a quick download, a Google sign-in, and installation of a (chrome) extension, and you’re ready to get summarizing (Briefy, 2023).
So, how does it work? After having installed the Briefy extension, go to a website with a text you’d like to summarize, look for the Briefy button (appearing if the extension is activated), and a pop-up lays out all the main points in easy-to-read bullets, without having to click away from the page (Briefy, 2023).
Even though it’s a very new tool, compared to more established ones like ChatGPT-4, Briefy shows a lot of promise for providing quality summaries and a super smooth usability. But it also comes with a couple of limitations (which I’m sure the team behind Briefy is already working on): so far, Briefy is only available on websites and is not yet able to handle very long articles, such as the academic ones we have to read. But this gives also a lot of room for improvement! Moreover, especially for more visual students, it would be amazing if Briefy developed into a tool that can not only provide us with quality summaries but also charts or mindmaps, all available to go on our phones or tablets.
Also, when using tools like Briefy, we should keep in mind that such AI summary tools can struggle to balance between being overly general or too specific and not grasping the context or what we deem important (Altmami & Menai, 2022; Widyassari et al., 2022). Moreover, despite improvements in natural language processing, it can also still be possible that these AIs lack deep semantic understanding, meaning that the summaries provided can be technically correct, but might still miss implict meanings which are present in the original article (Silva et al., 2019).
So, I personally believe that this new tool isn’t just a potential timesaver; it can turn into a starter gun for deeper dives into how we consume and process with the digital world.
Altmami, N. I., & Menai, M. E. B. (2022). Automatic summarization of scientific articles: A survey. Journal of King Saud University-Computer and Information Sciences, 34(4), 1011-1028.
Briefy. (2023). Briefy – AI-powered content summarizer. Retrieved October 20, 2023, from https://briefy.ai/
Silva, V. S., Freitas, A., & Handschuh, S. (2019). On the semantic interpretability of artificial intelligence models. arXiv preprint arXiv:1907.04105.
Widyassari, A. P., Rustad, S., Shidik, G. F., Noersasongko, E., Syukur, A., & Affandy, A. (2022). Review of automatic text summarization techniques & methods. Journal of King Saud University-Computer and Information Sciences, 34(4), 1029-1046.
Digital Echoes: When AI Breathes Life into Cover Songs
20
October
2023
No ratings yet.
How magical would it be if your favorite song could be covered by the artist you adore the most? For me, that artist is Winter from the Korean group, AESPA. Her voice stands out, with a metallic crispness, yet imbued with a deep warmth. As a Korean singer, most of her songs are in Korean with an occasional mix of English. The language barrier, unfortunately, prevents me from fully grasping the emotional depth of her songs.
One day, to my astonishment, I stumbled upon a video titled “AI WINTER Kim Min-jeong covers ‘Under Mount Fuji’” on a streaming platform. This song is a well-known Cantonese track. Driven by curiosity, I played the video, only to be overwhelmed by a profound sense of familiarity and understanding, all because of the power of my native language. It made me realize how her original fans feel when they listen to her.
（A Video Showcase: Winter – English song AI Cover)
This fascinating discovery propelled me into the world of AI song generation. It’s a technology that can analyze the original song and reproduce it in different voices or styles. What’s even more impressive is that this can all be accomplished without the need for a full band or backing track, opening up infinite possibilities for music creation. (Tucker, 2023) The production process is incredibly straightforward. To craft an AI cover, the model simulates the original artist’s voice, which can then be blended with that of a real singer. Technologies like RVC, for instance, can synthesize high-quality voice outputs based on existing audio. With a simple upload, you’re all set for a musical treat. (Wodecki, 2023) Advanced machine learning algorithms, like deep neural networks, drive most of these AI models. They are trained on vast datasets to understand various facets of music, from harmony and melody to rhythm. (Tucker, 2023)
In this ever-evolving tech era, we’re witnessing the golden age of music intertwined with technology. AI offers us a unique way to reshape and experience music, breaking down barriers of language and culture. It promises a world where music lovers across the globe can revel in this innovative charm. I eagerly await further blends of technology and art, where every individual can bask in the sheer beauty of music.
However, like all technologies, there’s a darker side. There have been real-world instances of deep fake audio deceiving unsuspecting victims, like the 2021 incident where a bank manager in Dubai was conned out of $35 million. (Wodecki, 2023b) Additionally, such undertakings might irk record companies. Universal Music Group, for instance, has expressed concerns over AI encroaching on their territory, hinting at lawsuits against platforms like Spotify if they launch AI-generated music products. (Wodecki, 2023)
This serves as a reminder that as we embrace and utilize these technologies, we should respect original works, abide by copyright laws, and ensure personal privacy against potential risks.
