Do algorithms dream of electric sheep?

If this title caught your attention, then I expect you – like me – to be a fan of Ridley Scott’s epic 1982 film Blade Runner. Scott left audiences guessing whether the film’s protagonist, Rick Deckard, was human or a replicant with implanted memories, just like those he was tasked with tracking down and eliminating.

Forty years later, in 2022, it’s fair to say that we live in a world in which machines are replicating what humans can do in increasingly sophisticated ways. Pre-trained natural language processing transformer models, such as OpenAI’s GPT-3 and Microsoft/NVIDIA’s MT-NLG, can produce text that looks like it was written by a human. Potential artists can play around with publicly available models such as Midjourney or Stable Diffusion to render images that very few of us could produce with our own hands and eyes (for example, the image accompanying this post was rendered by Stable Diffusion following a prompt based on the title).

These templates do heavy lifting and have tremendous business utility – writing texts, creating blogs, etc. However, they are not (yet) Rick Deckard. Why?

  • They are highly dependent on the large amount and quality of data needed to train them.
  • They also depend on significant human input, whether in selecting the training data, defining the goals to be achieved, defining the parameters, and selecting the prompts needed to guide the model toward achieving those goals, and any subsequent refinement of the rendered result. .
  • Small human refinements to prompts (instructions such as: “/imagine prompt: elephant, on bicycle, in the evening sky, Photorealism style”) can yield different results. The more refined and detailed your prompts are, the closer the result will be to what the human imagined.

The human contribution always counts a lot. This is important because copyright law is still generally rooted in the notion that a human author and intellectual creativity are necessary for copyright to arise. And copyright is important when it comes to determining who has the ability to control and exploit the outputs rendered by these models. It is therefore foreseeable that as the adoption of these models grows, more and more disputes will emerge about which humans are entitled to a copyright in the products they render. Perhaps most interestingly, the UK’s ‘outlier’ provision for computer-generated works that have no human authorship (S9(3) CDPA 1988) will get a new airing in the courts.

Disclosure: I got lost for hours using the Midjourney beta typing random prompts on Angry Judges and Space Llamas.

Pre-trained natural language processing transformer models, such as OpenAI’s GPT-3 and Microsoft/NVIDIA’s MT-NLG, can produce text that looks like it was written by a human.

www.bristows.com/…

Sharon D. Cole