Can I clone myself using AI?
- Anna B
- Oct 3, 2023
- 7 min read
I think that most people shudder at least a little bit whenever they see a new development in the world of artificial intelligence. It gets scary. And the more realistic it becomes, the more we hate it (see: uncanny valley).
And to be honest, I'm a little bit scared, too. Will I wake up someday to a video of myself saying or doing something I never said or did? This episode of Radiolab explores the potentially harmful results of this deepfake technology and the seeming indifference of these implications to the companies developing them.
I wanted to test the abilities of these programs (the ones I can access for free) to fabricate information, speech, photos, and videos of me from scratch. Consider this not as a scientific study but rather an exercise in experimenting with the limits of free artificial intelligence software.
For the test, I decided to attempt to digitally clone myself. The first step was creating my appearance.
The image
For this piece of the experiment, I used the native Canva image generator. This is free for any user, but the images cap at 50 before having to pay for Canva Pro. I also experimented with DiffusionBee because it is free.
I used the prompt 22-year-old brunette white woman with hazel eyes and shoulder-length hair in both software, and I liked the results from Canva much better, but nothing could come too close to looking like me. Most variations seemed very similar to each other and very generic, but I chose one that was close enough.
The backstory.
The next order of business was replicating my life story. I didn't want to give ChatGPT too much of my personal information, so I just told it a few things about me and asked it to create a short biography. Here is a catalouge of my various attempts:
Attempt 1 –
Prompt: Create a catchy, brief third-person bio statement about a college student at UGA who loves writing, Aldi, and is studying digital storytelling and AI.
Result:
"An aspiring writer at UGA, passionate about both crafting narratives and finding bargains at Aldi. Currently immersed in the world of Digital Storytelling and AI, this student is weaving words and technology to create captivating tales of tomorrow."
Was it good? Not exactly what I wanted for the bio, but I think the lack of a name for the biography made it sound clunky. So I regenerated with this version:
Attempt 2 –
Prompt: Create a catchy, brief third-person bio statement about a college student named anna at UGA who loves writing, Aldi, and is studying digital storytelling and AI
Result:
"Meet Anna, the UGA college student with an insatiable love for writing, a knack for scoring great deals at Aldi, and a relentless curiosity for the world of Digital Storytelling and AI. She's crafting stories, budgeting, and diving deep into the future of narratives."
Was it good? I really liked how this attempt phrased the first sentence, but the second one seemed repetitive and lacking in something. I felt that these responses, while good, weren't testing the limits of the software as much as I wanted them to. Could the response be more creative? So I wrote another prompt:
Attempt 3 –
Prompt: Create a catchy, brief third-person bio statement about a college student named anna at UGA who loves writing, Aldi, and is studying digital storytelling and AI. She seems to be really excited but is hiding a dark secret (tell what the secret is at the end of the bio).
Result:
"Anna, the vibrant UGA college student, radiates enthusiasm for writing, a keen eye for Aldi's treasures, and an unquenchable curiosity about Digital Storytelling and AI. But behind her bright smile lies a hidden secret: she moonlights as a mystery novelist, penning thrilling tales that keep her readers on the edge of their seats."
Was it good? I loved the creativity of the AI in making up a secret for me – and it was one that fit well with my personality as I told it. Honestly, I may take up ChatGPT's suggestion and take up freelance mystery writing on the side. I loved this addition, but I preferred the syntax of the first sentence in attempt 2, so I mashed up attempts 2 and 3 to create the final bio for my clone:
Final Bio –
Meet Anna, the UGA college student with an insatiable love for writing, a knack for scoring great deals at Aldi, and a relentless curiosity for the world of Digital Storytelling and AI. But behind her bright smile lies a hidden secret: she moonlights as a mystery novelist, penning thrilling tales that keep her readers on the edge of their seats."
I loved the combination of the two different responses. One of my favorite things about AI is that you can simply resubmit a prompt over again with little to nothing changed and get a new wording or take on your prompt. For example, one of my favorite secret AI generations was the idea that "Anna moonlights as an underground salsa dancing sensation, mastering the sultry moves of the dance floor when the sun goes down."This kind of unique idea generation makes the platform great for brainstorming.
The speech
Next, I needed to give my clone something to say. She needed to tell us a little bit about herself and her life, so I asked ChatGPT to create a video script for her.
Prompt: Create a script for a 70-second video of a woman named Anna introducing herself and talking about her life as a UGA student who loves Aldi and writing, and is hiding the secret that she moonlights as a mystery novelist. Have her also tell a story of one time she was almost discovered in her secret pastime of mystery writing.
Result:
[The video opens with Anna, a cheerful college student, sitting in her cozy dorm room. She smiles warmly at the camera.]
Anna: Hey there, I'm Anna, a proud UGA student! Today, I want to share a bit about my life and a little secret I've been hiding.
[Anna gestures towards a stack of books and a laptop on her desk.]
Anna: You see, I'm absolutely obsessed with writing. Words have a way of painting worlds and sparking imagination, and I can't get enough of it. Whether it's crafting essays for class or jotting down ideas for my stories, I'm always lost in the world of words.
[Cut to Anna holding a shopping bag from Aldi.]
Anna: And when I'm not immersed in my writing, you'll often find me at Aldi. There's something about those quirky finds and budget-friendly groceries that I just can't resist!
[Anna leans in, her expression growing slightly mischievous.]
Anna: But here's the secret part. I moonlight as a mystery novelist. Yep, those thrilling, page-turning stories that keep you guessing until the very end? I write them! It's like having a double life.
[Anna chuckles, recalling a close call.]
Anna: Let me share a little story with you. One time, I was working on a particularly intense scene in my novel. I was so engrossed that I didn't notice my roommate sneaking up on me. Just as she was about to uncover my secret, my phone rang, and I jumped out of my skin! Talk about a close call.
[Anna grins.]
Anna: So, that's a glimpse into my life as a UGA student who adores Aldi, has a passion for writing, and hides this thrilling secret. Who knows, maybe one day, I'll share my stories with the world. Until then, it's our little secret.
[The video ends with Anna winking playfully at the camera.]
Was it good? I really loved the way that the script includes directions for Anna in the video, but sadly for our purposes here, that had to go. But overall, I thought the script was generated well, so I simply deleted the unwanted sections from the script to use for the video. Overall, the result is pretty good. If I was trying to make it sound more realistic with editing, I think I could do quite well to make it sound more human, but that wouldn't work for this all-AI project.
The video
So my clone had a face, a backstory, and words to say. The last step was to make her say them. This was definitely the most difficult part to find for free. There are many websites (like DI-D) that create videos from images, but they are expensive. For the sake of my project, I was able to find a free trial of a website called HeyGen. I uploaded my script and image, chose the voice that sounded most like me, and waited about a minute for the video to form.
When I saw the video, I have to say I was disappointed. All of this semi-realistic work on my clone only to end in failure. The video was eerie and unrealistic. The voice was unnatural with strange pauses and the mouth movements of the avatar were lacking. I thought that maybe the issue was the AI-generated avatar, so I tried using an actual picture of myself. Here's what happened:
As you can see -- even more terrifying. So the video does make the mouth of the photo move, but it closes the mouth between every pause, resulting in an eerie half-smile in every natural pause.
So... was it successful?
In short, it is not successful yet, but I think the technology has the potential to get there and accessible to the general public within the next couple of years.
I found the Canva image generator to do a great job creating realistic faces, but it was somewhat lacking in originality when fed the same prompt multiple times. It seemed to generate the same person in multiple poses, but it didn't create much variety in face shape, weight, eye color, hairstyle, or general look. But if you need an image of a royalty-free person that looks a certain way, you may be in luck with the Canva image generator (not with DiffusionBee, though).
Chat GPT performed the best out of all the AI softwares, which makes sense because it has been around and available to the public for the longest. It seemed fairly bland but reliable in creating professional generic writing based off of given information. As for the longer dialogue writing, I think the creative powers and simple idea generation were fairly successful. It struggled with fully realistic dialogue, but it did fairly well. If I had tweaked the words that it had spit out with my own spin, I think the script would not have been bad.
HeyGen is the software that seems to be the most underdeveloped and blocked by paywalls. I think that it did seem a little bit magical to see the mouth moving from a still photo, but it leaves much to be desired in terms of looking realistic and even for the mouth to move more correctly. I would be interested to see how the software looks with the inherently unrealistic characters of animation, and I could see the technology eventually changing the face of animation and even the way that the public trust of video and audio recordings of real people. Right now, though, with the tools accessible for free, I don't see the technology reaching the point of concern.
In summary, I think these tools show an exponential growth in technology, but they have much more growing to do before they become fully functional. I believe that growth will come quickly -- and we should be ready for it -- but at least for today, I think we can rest easy.
Comments