What’s Real Anymore In The Age Of Generative AI? — Discussion & Recommendations

How Generative AI Helped Me Fabricate a Video News Story for $30 in One Afternoon (Part 4)

May 02, 2023

The guiding question: “Would you trust them?”

Recap: The Idea For A Personal Creative Project

Earlier this month, I turned my curiosity and learning about generative AI into a creative project, using publicly available tools. In between running errands, I sat down to try it out. To create the assets with AI, I used nothing but a smartphone where the tools support mobile devices.

Four generative AI tools have helped me create a full video of a fabricated news story in one afternoon. Today, I’ll share the final video and how I’ve used the two additional tools to create it...

Note: On April 25, 2023, MSNBC & others reported about an AI-generated, political video ad, published by the U.S. Republican National Committee. Hence, AI-generated information is no longer a thought experiment of the future and rather an imminent matter of the present.

» Watch the latest episodes on YouTube or listen wherever you get your podcasts. «

Disclaimer

The following is a creative project that I have created using several generative AI tools. The goal of this project is to start a discourse about the potential of generative AI technologies, learning and using the tools, combining the output of one tool with the input of another, and thinking about the ethical implications of AI-generated content at scale. Lastly, this project aims to prepare AI leaders to discuss these aspects with their business peers. The use of specific tools is neither an endorsement nor taking a position on the product. They are rather exemplary for applications for this medium. Finally, the objectives and motives of this project as well as the views expressed in this post and the linked videos are my own.

Despite all the good we can expect generative AI to create, it will also make it a lot easier to create and spread misinformation.

“How will we know what’s real anymore? And how can we tell?”

Here is the final video that I created using four generative AI tools:

Thank you for reading The AI MEMO. Share it with your network.

What can you do with generative AI today?

Overall result: ✅✅✅

As it turns out, a whole lot. The technology can already create very convincing results which can be hard to distinguish from real images, for example. While there could still be minor inaccuracies, you can expect the technology to only get better and cheaper within the next few months and years. In turn, this will mean that it will get harder to determine what’s authentic and what has been generated by AI — for those that create it as well as for those that (dis-)prove it. Since starting the project, Runway have released their application, GEN-2, which lets you create short videos from an image or another short video clip, NVIDIA have published a paper and samples on another text-to-video tool. In addition, on May 1, 2023, The New York Times published an article about Dr Geoffrey Hinton (“The Godfather of AI”) stating his immediate concern is that the internet will be flooded with false photos, videos and text, and the average person will “not be able to know what is true anymore.”

How realistic will the results be?

Quality: ✅✅✅

Most of the images I have created in Midjourney look very realistic (e.g., reporter, resident #1). This makes it hard to tell with a naked eye which ones are real or not. This is even harder for the untrained/ unaware user. After creating the full video, I’ve kept improving certain prompts to see whether the results could become even more realistic.

CEO character created by Midjourney (without an ethnic background in the prompt)

For example, the next iterations of the CEO and the aerial shot of the manufacturing plant look more realistic after dropping the word photorealistic in the prompt.

More realistic aerial shot created by Midjourney of the manufacturing site

How do you use the tools?

Access: ✅✅

ChatGPT and Midjourney are text-based applications. Getting the prompts right, such that they create the desired output, requires some reading, refinement, and multiple iterations. It’s well-documented.

ElevenLabs and D-ID have a user-friendly interface and work via copy & paste of the text you want to be spoken. There is no hurdle to using them.

How long does it take to generate a good result?

Scale: ✅✅

It has taken me roughly one afternoon to create most of the components I needed, making it extremely quick to put together a full deepfake video. Out of all the steps, editing the final video has taken the most time.

What are the current limitations of the technology?

Diversity/ bias: ❌❌

There are two kinds of limitations I’ve noticed: technology and data. The quality of images created with Midjourney has improved immensely within the last year. Current limitations are few, but noticeable imperfections of the characters’ hands/ fingers and any words/ signs that are visible in the image remain. Secondly, lack of diversity appears to be an issue. Unless specifically added, several characters had been created as Caucasians; for example the news anchor: a color photo of a male news anchor wearing a suit and sitting at a news desk. Hands open, concerned look, 2023 --ar 16:9 --v 5

News anchor created by Midjourney (without an ethnic background in the prompt)

Only when explicitly adding the words “African American” has Midjourney provided diverse characters: a color photo of a male African American news anchor wearing a suit and sitting at a news desk. Hands open, concerned look, 2023 --ar 16:9 --v 5

News anchor created by Midjourney (when adding “African American” in the prompt)

In the final, animated scenes created with D-ID, robotic and only partial body movements (e.g. head, eye lids, mouth) make the characters seem unnatural at times — especially, when scenes are longer or have more than one person in them. But for now, this can also be a way to spot an AI-generated video.

What are you going to do with the results?

Ethics: ⚠️

Is it ethical to write in public about creating a deepfake video? I went back and forth on this one for quite some time and discussed it with ethicists in my network. With the knowledge that AI-generated misinformation is going to be created and how this will likely be accomplished, would it be ethical not to write about it? For me, the answer has been a clear: “No. Not writing about it would be unethical.”

Three Factors For AI-generated Misinformation

After posting the trailer on LinkedIn, several people in my networked have asked me what I think is different when it comes to AI-generated misinformation? After all, haven’t we already been living in a “post-factual” world for a number of years (“alternative facts”, “fake news”, etc.)? So, why does it matter?

To me, the answers to these questions are simple. They are much simpler than the answer to the question “what can we do about it?”

In the end, it comes down to three factors that make AI-generated mis- and disinformation such a big concern: Access. Quality. Scale.

1) Access: Anyone Can Create It

I was sitting in my car in the parking lot of my local grocery store and created all the text-based artifacts using ChatGPT and ElevenLabs in my browser. Creating the images was a few swipes away, using Midjourney in the Discord app. It has cost $30 or less!

Anyone can do it — whether with good intentions or bad ones. All you need is a browser and an internet connection.

2) Quality: Hard To Distinguish From Reality

The quality of the generated output is already significantly better than it has been last year around. It is getting harder and harder to distinguish AI- from human-created work. The quality of the results will only get better over time — and so will the capabilities.

3) Scale: Distribute It Instantly — Worldwide

Once it’s been created, anyone can share it — instantly and with nearly anyone in the world. The impact of this level of scale can be far-reaching when the message reaches minds that are receptive for the narrative told.

Conclusion

I’ve been writing about various aspects of generative AI and its social/ societal impact — for example, using outsourced labor for content moderation, the early hype of prompt engineering, the ethical implications of generative AI, the impact on communication as it gets easier to combine different types of AI-generated media (e.g. text + audio + video), and the responsibility put into the public’s hands.

While it might be possible to spot AI-generated images and videos, it’s safe to assume that won’t be the case for long. Technical mechanisms of identifying AI-generated content are few and it is a game of cat-and-mouse for such tools to keep up. Regulation is gearing up, but it might also not be the answer to each and every problem that generative AI tools in the wrong hands and minds can create.

So, what’s the solution to all this? Right now, the best we can do as humans is to stay vigilant, to activate our critical thinking, to be skeptical about polarizing information, and to keep raising awareness among those who are not as close to the technology as we are.

This is the final post in a 4-part series. I’m sharing my thoughts and approach to using generative AI to create a video of a fabricated news story. Thank you for following along from creating the idea to generating the script and imagery, the voices and videos, to this discussion of the final results.

How will you know what’s real anymore? How can you tell?