One might be convinced that our robot overlords have finally arrived, with all the noise in the news and social media about the new generation of generative AI tools. Tools such as GPT-3 & GPT-4, Midjourney, and Stable Diffusion, have resulted in a wave of creativity as we experiment with them, discovering what they can do, the new opportunities they represent, how to trick them, and where they fail. It’s now possible to turn a rough drawing into a functioning web site, create a recipe from a picture of potential ingredients, or develop a Seinfield-spoof streaming show. Conversations with these tools have even led some users to believe that the technology is conscious.
It’s not surprising, given this, that some pundits are arguing that we stand at the precipice of a new AI-powered age where everything will be different. Our jobs, businesses, and even our daily lives will be disrupted due to the pervasive nature of the technology. Other pundits claim that generative AI is no more than a stochastic parrot—a “system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning.” Generative AI tools are prone to ‘hallucinations’, such as confidently inventing references for an essay. Many of the answers these tools generate might be little more than a hallucination. This point of view sees generative AI as a (potentially fatally) flawed and so dangerous tool, one to be controlled or even banned.
It’s important to note that generative AI is a useful tool, despite its flaws. With a few words or a rough sketch one can quickly generate a prototype of an image, poem, web site, or essay, photo, or lesson plan. We can combine capabilities to create more complex works, such as crafting a prompt, which generates a script for a video, which is used in turn to generate an actual video. Sometimes the prototype is convincing enough on its own, such as the deep fake that fooled voice recognition technology used by Centrelink and the ATO. Or we might use it to represent ourselves on a dating web site and so more easily screen potential dating partners. More often though, the prototype will need some editing before it’s acceptable.
This quick and easy generation of prototypes is incredibly useful. It’s also something that we were already using tools like internet search engines to do. One might, for example, want to write a job description. In the past this would likely involve searching the internet for descriptions for similar position. We’d collect the best examples and fuse them into a prototype job description which we then edit to create our final text. Now, with generative AI, we can provide an AI tool with a description of what we want and have it generate the prototype more-or-less instantly before editing the prototype into a final job description (which is especially important given AI’s tendency to hallucinate). A similar approach could be taken with an essay or report. Or when we want stock art, ‘searching’ a generative AI for a suitable image rather than an art library (Shutterstock, a stock art tool, has launched a generative AI feature that creates images from the service’s pool of stock art. And so on. The difference between the web-search based and generative AI is that the generative approach is more convenient, it makes us to be more productive.
Generative AI also allows us to work across media—converting a textual scene description into a prototype image, or a rough drawing of a web site into working code—without the need for deep expertise in the target media. This is similar how the invention of the camera empowered anyone with a camera to create a “window onto the world” without having the technical facility required to paint. Or when an artist focuses on the idea, and not the production of the work. We can do for ourselves what, previously, we had to rely on others for. It’s important to be mindful when using generative AI in this way though, as it’s possible that you won’t have the expertise required to correct the prototype. Your newly generated web site, for example, might misprice items, for example. Or you might find it hard to generate a compelling photo as you cannot articulate (to the AI) what would make the photo compelling.
Where society finally goes with generative AI will, as always, be somewhere between competing visions for the future, between ‘everything will be different’ and ‘stochastic parrot’. New technology creates new opportunities, but it’s up to us to determine how the technology will be used. Technology might best be seen not as a thing, but as a form of human action. Indeed, disruption is typically not the consequence of the invention of some disruptive technology, but rather the result of a renegotiation of the relationships within society. Workers, departments, or entire firms and industries, easily absorb technology once the benefits are obvious. Blockbuster, for example, had developed both DVD by mail and video streaming services prior to going bankrupt. The firm was disrupted when its relationships, and relationships of those around it, changed in a way that made Blockbuster redundant. For Blockbuster this was the need to unpick a network of store-based franchises in a world where video was no longer distributed on physical media.
Predictions of the impact of generative AI focus on the technology, rather than the relationships. Take the recent prediction that “approximately 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of GPTs, while around 19% of workers may see at least 50% of their tasks impacted.” The assumption behind this prediction is that the capabilities provided by generative AI will replace some or all of the capabilities of the human workers, forcing them to upgrade their skills or become redundant. Skill upgrading does occur, such as when the automation of telephone exchanges gradually made telephone operators redundant, starting with simple local exchanges then working to more complex regional, national, and international exchanges. However, we assume that this process ubiquitous when it is actually quite rare. The early 2000s, for example, saw a near unanimous assumption that self-service web sites would eliminate jobs like travel agents, replaced by self-service web sites, but this never came to pass. What is more common, is that workers absorb a new technology and do their job in a different way. Travel agents use the new web sites so that they can focus on their conversation with clients, rather than the challenges of using a cryptic booking tool. Similarly, the work of sign writers changed radically with the introduction of low-cost desktop publishing and vinyl printers, only to see the profession expand into new customers (by lowering costs) and new domains (wrapping cars is hot at the moment) without major disruption.
Predictions of disruption are unlikely to play out. We’ve been asked, for example, if generative AI will disrupt consulting. We see this as unlikely, though these tools may change how consulting is done. Using a generative AI tool to rapidly create a prototype report is likely a good thing as it makes the consultant more productive, enabling them to develop a deeper understanding of the problem and focus on the details of the solution. Report prototypes are document templates on steroids. Similarly, generative AI is unlikely to disrupt education. These tools do compound existing problems with assessment, but these are not new problems. Students were already using online services to contract out their assignments. Generative AI just makes this cheaper and more convenient.
We, humans, tend to assume intelligence wherever we see complex behaviour. Bacteria, for example, are considered intelligent when they follow a concentration gradient and (collectively) solve a puzzle. Or that a horse is capable to maths, when it’s really just watching its trainer for cues. AI is no different. We see a piece of AI generated art which we consider creative, and attribute it to the AI, rather than the human who drove the AI to generate the image. The AI hasn’t automated the creative act, but it has made the creative individual more productive.
There’s a growing realisation that while impressive, and even surprising to many, generative AI is primarily a productivity tool. The new generative tools might, for example, make it easy to compose and orchestrate music. “Create a jingle for toothpaste in the style of reggae.” In some instances, this will enable individuals to create their own music directly, if they can articulate what they want, and the prototype generated by the AI is good enough. The next step up might be to commission a composer to generate the music for you, to combine their expertise with generative AI. And there will always be instances where an experienced composer is required, where the individual commissioning the work needs the (creative) help of a composer to determine what it is they want. Like signwriting, composition will expand into new customers and new domains.Most importantly, if we do find ourselves in an AI dystopia where everything is different, then the blame lies firmly withus, the humans, and not generative AI. It will be our choices and actions that create the dystopia. Not AI.