Dr Leon Furze

What is Digital Plastic?

Last year, I wrote an article introducing the concept of digital plastic, coining the term to describe AI-generated synthetic media, which, like its real-world counterpart, can be both helpful and harmful. Lately, I’ve been revisiting the term amidst ongoing discussions around quickly developing technologies like video generation. In this post, I’m going to update and refine my earlier definition and offer up a few more examples of why I think AI-generated media online should be treated as digital plastic.

(Re)Defining Digital Plastic

In much the same way that physical plastic is a synthetic material produced by various methods and chemical combinations, digital plastic is synthetic media that can be created through a variety of machine learning techniques. I use the term digital plastic not just because of the synthetic nature of AI-generated content, but also as a way of thinking through the implications of that media.

Plastics were invented to solve real problems, such as:

• A lack of natural materials
• The difficulties in working with materials such as wood and metal
• The desire to create materials with significantly different properties to natural elements, such as flexibility, strength, and durability

Plastic can bend without breaking, can last for years without rusting or decaying, and can be used in positive ways, from prosthetics to sterile packaging and more.

Of course, plastics have also come to be one of the most problematic inventions of all time. Alongside its other petroleum-based cousins, plastics are a major contributor to pollution, damaging to both wildlife and human life and often carcinogenic. One of their biggest strengths – durability – is also one of the biggest environmental problems of the material.

I use the term digital plastic to draw similar comparisons with generative AI. While synthetic media has positive uses, it will also have an impact on the digital environment. Like plastics in the real world, I’m not sure we’re fully prepared for the implications of an online ecosystem filled with digital plastic.

Content on Demand, or a Tsunami of Slop?

In the last few months, we’ve seen a surge in research and product releases related to generative artificial intelligence, moving beyond text and image generation and into the lesser-established areas of audio and video.

• In February 2024, OpenAI teased Sora, a powerful-looking video model which still has yet to be released, but is promised to be offered soon as a plugin for Adobe Premiere Pro.
• Two new music generation platforms, Udio and Suno, arrived almost simultaneously, offering free, high-quality audio generation that outstripped any other platform on the market.
• Hot on the heels of Sora’s video generation, Runway and Luma have both released new products which can create short video clips of a much higher quality than what was available in 2023.
• In voice generation, OpenAI also signaled updates to its speech models, but have been delayed partly due to controversies over how they trained those voices.
• ElevenLabs, a commercially available platform which can make deepfake voice clones, has lately been releasing product after product, including voice editing tools for creators, audiobook and PDF voice readers, and their expanding repertoire of synthetic voices.

If you take these platforms in combination, it’s already fairly straightforward to create multimodal content on demand. I used a combination of those platforms (Runway Gen 3, Suno, GPT-4, and 11 Labs) to create two short videos in under 10 minutes each, including this one:

The ability to create engaging content quickly is obviously attractive, given we live in a society where video content platforms such as YouTube and TikTok have risen to be the most popular parts of the internet. But what about the ethical issues behind these technologies?

The aforementioned music generation platforms, Udio and Suno, are both currently being sued by Universal Music Group and Sony for their probable use of unlicensed copyrighted music in their training data.

OpenAI were accused of deliberately imitating Scarlett Johansson’s voice in the Sora voice model that they demonstrated in spring.

Runway, Luma, and other video generation platforms are largely built using open-source Stable Diffusion image generation models, models which have proven to be tainted not only by copyrighted images in the dataset, but by CSAM and non-consensual explicit images scraped from sources online.

You can see the possibilities, but also the dilemma. And aside from how these products are created, there are already significant concerns over how they will be used.

The term “slop” is currently gaining traction for AI-generated content that’s fully automated, soulless, and rapidly filling the internet. Websites are using automated chatbots with language models that are churning out useless and often factually inaccurate materials at a scale we have never seen before.

What happens when slop becomes multimodal? As soon as it’s possible, easy, and cheap to run text, audio, and video generation in concert, then you can all but guarantee the internet will fill up with hours and hours of generated, synthetic media that has had little to no human intervention. Just as real plastic clogs the arteries of our streams, rivers, and oceans, digital plastic will fill the online ecosystem in a short few years.

Protecting the Digital Ecosystem

I think the digital plastic analogy is helpful because it also suggests a way forward. Regulation in manufacturing and environmental law has been slow to catch up with real-world plastic. In fact, despite the best efforts of environmental organisations, businesses, researchers, and political groups worldwide, we are still drowning in plastic.

It doesn’t need to be this way with digital plastic. Recently, video platform Vimeo announced that they will not allow generative AI developers to train models on their users’ content, nor will they train models themselves. They will, however, incorporate some artificial intelligence technologies into their platform.

This is an acknowledgement that some forms of digital plastic are more useful than others. Audio-to-text transcription, for example, is a form of artificial intelligence that can be used to help creators make their videos more accessible and more searchable online. It’s also an acknowledgement that some forms of digital plastic can be harmful to creators. Training models without consent on creators’ intellectual property for the purpose of creating an application which could theoretically compete with them seems like one such obvious harmful use.

Just as our understanding of different manufacturing methods has evolved over time, and there are some plastics which are better than others – biodegradable, less resource-intense, recycled, or made using more sustainable or environmentally conscious plant-based materials – there are already organisations helping to make the “manufacture” of digital plastics more ethical.

Ed Newton-Rex’s Fairly Trained, for example, offers certification to organisations that can demonstrate they have licensed their training materials and appropriately compensated the creators. There are already over a dozen platforms, predominantly music generation, who have achieved the certification. You can use music generated in a platform like Beatoven.ai to create commercially viable, ethically sound audio tracks for use in social media, commercial applications, education, or personal use.

This stands at a distinct angle from the approaches of companies like Suno, whose business model seems to be to scrape knowingly and indiscriminately, appropriating as much data as possible, build and deploy a powerful model for free, attract users and investors, and use venture capital money to fight off lawsuits.

In fact, the business model of companies like Suno, Udio, and even OpenAI is nothing new. It’s the business model of cigarette companies, casinos, and firearms manufacturers.

And so we have some decisions to make. How can we protect the digital ecosystem in ways we failed to with the physical one? What measures can be put in place to stem the tsunami of slop while still allowing for the good uses of digital plastic?

• LinkedIn
• Bluesky
• Facebook
• ShareCopied to clipboard
• Mastodon
• Threads

Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch: