Is AI “Just A Tool”?
People often say that AI is just a tool. But it’s not. That’s a fundamental mistake and likely to be wrong by several orders of magnitude.
Why is it wrong? Don't we have it in our hands to ensure that AI will always remain a tool? Probably not. At a facile level, when was the last time your power drill wrote a poem? How often has your screwdriver recommended a list of books about the European Enlightenment? When was the last time your text editor wrote a Javascript module that's unintelligible and yet works better than anything a human could have devised, and did it in six seconds?
However much we might want AI to remain a tool, that's a bit like wanting to cross the Atlantic by flapping your arms.
That's not to say you can't use generative AI as a tool. You can make it fix distorted or degraded images and repair broken audio. It's becoming possible to zoom into a picture without losing quality (if not authenticity). Whether or not you can validly describe AI as a tool ultimately depends on the context.
But why are we even having this discussion? Does it matter what we call it? Does anyone care?
Yes, it matters, and yes, we should care.
Broadcasters live in two separate domains: technical and creative. Both will be drastically affected by AI, and the degree to which they're affected depends largely on how we take a stance on AI—and a large part of that depends on whether we call it a tool or something else. This is about more than how we name something. It's about how accurately we recognize what it is.
One thing that should make us hesitate to call AI "just a tool" is that it is unpredictable. It's not the unpredictability of a 25-year-old car trying to start on a cold winter morning; it's the tendency of AI to take off in unexpected directions as new and unexpected capabilities emerge. These so-called "emergent properties" might be trivial and insignificant. Or they might be less trivial, like "I'm now sentient".
Seriously? We at least have to consider it, if only to reassure ourselves that we haven't reached that stage yet.
Some experts are skeptical about whether emergent properties exist at all, and that's fair because these novel capabilities tend to take us by surprise, and their causes are not always clear. Let's look at a few examples where it's unclear whether AI is a tool.
Quality control has been automatable for some time. It's relatively straightforward to devise a set of tests and criteria and then apply those to an automated workflow. (Of course, it's not simple at all - some of these tests are incredibly sophisticated - but there is nothing that is, in principle, difficult). At present, it is typically done algorithmically, but it seems likely that AI could do a better job. Before you ask whether AI is a tool, you have to ask whether an algorithm is a tool, which it is arguably not. An algorithm is a predetermined method designed to produce a desired outcome. Ultimately, it all comes down to how you define a tool.
AI seems eminently suitable for watching content to spot anomalies and contraventions in much the same way it is currently - and increasingly - being used to spot anomalous skin markings to detect cancer. AI could be excellent for compliance monitoring. Should it act on its judgements? That's a huge question. Driverless cars will act on their judgements in a life-or-death manner. Whether or not you find that concerning depends on your confidence in the technology - and the same applies to almost any AI-automated task. Perhaps only time will tell, and indeed, time will tell us whether, statistically, AI tools give better results. When they do, it arguably only then depends on the severity of the downsides when it makes mistakes.
Although not strictly a broadcast topic, color grading has been given a productivity boost by products like Colourlab AI [https://colourlab.ai/], an AI-based system for - amongst other things - taking the repetitive work out of color grading. One of its capabilities is to process all the incoming media and grade it so that all the shots - even from different types of camera - match in a basic way. It removes the inevitable differences to present the (human) color-grading experts with a level playing field. You can imagine a version of this would be helpful in complex live broadcasts. It has been trained on a vast range of material - film and video productions across an entire spectrum of genres. In post-production, the aim is to automatically match all the footage in a project to a base level, at which point the human color grading expert can begin to work their magic. In this context, we can perhaps feel relaxed about calling AI a tool, albeit a very capable one.
Once again, it is a question of judgment. You can say that any technology that has the capacity to intervene - even a fire sprinkler system makes judgements about whether to trigger the fire extinguishers, but that would be stretching the definition as far as it can go without breaking. More reasonably, you might argue that something that can "judge" at least has an element of intelligence about it. It's one thing for the system to match the color look of multiple types of cameras on a production, but it's quite another for it to carry out the "artistic" color grading itself.
And yet, AI can write poetry and paint pictures. When it does that, is it being "artistic"? It would be hard to deny if a controlled experiment showed people couldn't distinguish between a human-generated painting and an AI-generated one. It would be even harder if the experimental subjects consistently preferred the AI pictures.
Just because an AI is capable of making artistic judgements doesn't mean it always has to. It's easy to imagine where AI could help create efficiencies in highly technical situations. As cloud production ramps towards a default way of working, the complex data paths, data rates, and interstitial routing add up to a hugely complex optimization task. With its ability to spot patterns, make inferences and suggest solutions, AI could be an ideal assistant or co-worker. It would still have to demonstrate, statistically, that it can perform better, more economically, and with smaller downsides than "traditional" human operators.
It is always going to be a difficult conversation to have when AI models replace human operators. It almost never seems right to call a human being a tool: you would need an extraordinarily perverse type of abstraction to do that. And yet, there were times when, before computers, human beings were treated exactly as tools. Before the widespread use of computers, "computer" referred to a person who performed mathematical calculations by hand. These individuals were often employed in scientific and engineering fields to analyze data, solve equations, and create tables.
It seems likely that helpful AI models will gradually cease to be called tools, and become "assistants": willing, able and flexible helpers that can do repetitive work, optimize workflows and watch out for technical failures. Before long, AI will leave the constraints of living in a cloud server, grow arms and legs, and have senses like vision, touch, and hearing. It will sit in the same chairs that we do and press the same buttons. At the same time, broadcast technology will become increasingly virtual, with AI "living" in the technology. All of which makes the word "tool" seem inadequate.
As a postscript, it is worth mentioning that whether we treat AI as a tool or as an assistant, we need to find a way to incorporate these new capabilities into a consistent and robust way of working. One way to do this would be to devise a technology stack that would be analogous to existing methods (like the layers in a network) but which would include a "cognitive" layer, whose job would be to interpret either the artistic or commercial intent of the production and assist in making it happen. It would be open to new developments and designed with interfaces or APIs to allow existing technology (and existing technology layers) to participate in the new cognitive workflow.
Some of this may sound far-fetched, but we should probably have been planning this some time ago and ready to implement it today. In mitigation, that's hard to do when the AI landscape changes so quickly and constantly.
You might also like...
HDR & WCG For Broadcast: Part 3 - Achieving Simultaneous HDR-SDR Workflows
Welcome to Part 3 of ‘HDR & WCG For Broadcast’ - a major 10 article exploration of the science and practical applications of all aspects of High Dynamic Range and Wide Color Gamut for broadcast production. Part 3 discusses the creative challenges of HDR…
IP Security For Broadcasters: Part 4 - MACsec Explained
IPsec and VPN provide much improved security over untrusted networks such as the internet. However, security may need to improve within a local area network, and to achieve this we have MACsec in our arsenal of security solutions.
IP Security For Broadcasters: Part 3 - IPsec Explained
One of the great advantages of the internet is that it relies on open standards that promote routing of IP packets between multiple networks. But this provides many challenges when considering security. The good news is that we have solutions…
The Resolution Revolution
We can now capture video in much higher resolutions than we can transmit, distribute and display. But should we?
Microphones: Part 3 - Human Auditory System
To get the best out of a microphone it is important to understand how it differs from the human ear.