Until around half a decade ago, when I had to conduct an interview for a writing task, I was confronted with the arduous task of transcribing the hours-long discussions that had been documented. This was precisely the type of monotonous paperwork I assumed I’d sidestepped when I chose a writing career over financial stability and conventional respectability. The transcription procedure was so endlessly grinding that I would usually contemplate throwing in the towel on my writing endeavours and pursuing a mainstream job, particularly after four to five hours of continuous transcription.
However, things took a beneficial turn when I began using an application that exploited machine learning to transcribe verbal exchanges into written text. To my surprise, the app was quite adept and accurate, and I found myself appreciating its prowess in automating a task that I found incredibly dreary and exhaustive. While it wasn’t flawless in handling thick accents, inaudible words, and would occasionally make illogical mistakes, I too was guilty of the same, albeit at a slower pace.
Following the completion of my latest book’s research nearly two years ago, I shifted my focus on different writing styles and no longer needed the transcription software. But recently, I embarked on an extensive magazine article that required interview transcripts, and I resumed using the software. Quite truthfully, I was astounded how far the technology had advanced since I last used it. The software’s “artificial intelligence” element, which seemed vague to me initially, was now evidently visible and highly efficient. Accuracy of the transcriptions had significantly improved and now included detailed, pinpointed summaries of dialogues, categorised under relevant themes. This service was as efficient as the one I might have received if I had employed an exceptionally competent individual to undertake the repetitive, yet crucial tasks that my job necessitates but which I absolutely loathe and am notably incompetent at.
As a means of reducing labour-intensiveness and allowing more time for the creative aspects of my profession, AI has proven to be an unarguably powerful tool.
There’s a certain degree of uncertainty hovering around this topic: although I was not going to hire someone to perform these tasks, I’m cognizant that it is a type of work for which normally people receive remuneration, or at least did until recently, and now, this technology is gradually taking their place.
For a long time, I’ve been quite hesitant towards numerous assertions made about machine learning and its potential capabilities. But this recent advancement in transcription software strongly underscores the potential of the technology. As an instrument, it is immensely potent in reducing labour and freeing up time for more creative aspects of my work.
However, the utility of this software suggests some of the ways in which there is over-promotion of the Large Language Model (LLM) technology that it is based on. As a tool for converting dialogue to text, it does a humble yet highly specific task extremely well. It creates an initial text based on spoken input, followed by secondary texts— a summarised outline and elongated summary—built on automated evaluation of the initial output. In this sense, it’s an encapsulation of more generic LLM technologies such as ChatGPT, which manufacture secondary texts not from a sole source, but from the vast scope of the internet.
This often leads to considerable complications and missteps.
Take, for example, the comical launch of Overviews, Google’s fresh AI search tool, that leverages LLM technology to condense a broad field of search outcomes for a typed input into a compact text summary. Recently, social media platforms have been bombarded with screenshots of increasingly bizarre responses to even the most benign search terms. A typical example, in response to “how to swiftly pass kidney stones”, advised drinking at least 2 quarts (2 litres) of urine every day while ensuring the urine remains a light shade. (In order to prevent a flurry of rebukes, I must state here that, of course, the appropriate volume of urine for consumption in a day is between 500-750ml, with a preference for unleavened.)
If I had an extremely efficient individual to carry out all the mundane yet necessary tasks that my profession demands, that would be something I’d anticipate. An abundant error from Overviews suggested that to make cheese cling to a pizza, one might add approximately 1/8 cup of harmless adhesive to the sauce to enhance its stickiness. This blatantly ridiculous advice seemed to originate from a decade-old jest on Reddit, concocted by someone with the username “f**ksmith”. Overviews, akin to all Language Learning Models (LLMs), extracts any information it deems related from its database. However, being an Artificial Neural Network, not a human, it’s underequipped to discern valuable outcomes from the worthlessones – misinformation, Reddit posts from individuals like “f**ksmith”, and downright erroneous information.
The tech industry has a gigantic task of building faith in AI, as heard in the committee. In a conversation with the Financial Times the previous year, acclaimed American science-fiction author Ted Chiang gave a somewhat tongue-in-cheek but thoroughly practical definition of AI. He referred to a tweet describing it as “an ill-chosen term from 1954”, suggesting that if the computer scientists who innovated the technology post-war had opted for a different label, a substantial amount of confusion might have been bypassed. “Applied statistics” was Chiang’s proposed moniker, less captivating for sure, but more precise in depicting what these networks truly do, and less prone to fuel imaginings of machines gaining consciousness or thinking ability.
This tech unequivocally possesses its uses and is progressively becoming potent as a tool to condense labour. However, when it comes to “intelligence” – it is still figuratively ingesting nonsense and edible adhesive.