A half-decade ago, every time I conducted an interview for a writing assignment, I had to endure the painstaking process of transcribing my recorded discussions, which often spanned several hours. I had become a writer to avoid such numbingly dull secretarial duties; duties that I thought I had escaped when I sacrificed steady income and societal acceptance. Once I reached the verge of giving up writing, vowing to find a more routine job around the fourth or fifth hour of transcription.
However, things took a turn when I started utilising an app that leverages machine learning to transform verbal communication into written words. Surprisingly quick and precise, the app left me astonished with its ability to automate one of the most tedious tasks in my writing profession. While it had its shortcomings like misunderstanding heavy accents, unclear words or making absurd mistakes, I found it wasn’t significantly different from me, just more efficient.
After putting the finishing touches on my most recent book approximately two years prior, I got engaged in different varieties of writing and refrained from using the transcription software. Recently, however, I picked it back up for use in a lengthy magazine article, involving extensive interviews. I was genuinely astounded by the marked advancement in the technology during my absence. The enigmatic concept of “artificial intelligence” that seemed obscure to me earlier was now crystal clear. It wasn’t just the improved accuracy of the transcripts; it was the ability of this software to provide a comprehensive breakdown of the dialogues under various subject headings along with an impressively precise summary. It felt akin to having a hyper-efficient assistant taking care of all the gruelling, monotonous, necessary aspects of my job that I detest and am not particularly good at.
Undeniably, as an instrument for reducing labour and creating more time to engage in the creative parts of my profession, artificial intelligence proved to be a potent tool.
Here’s a conundrum for you: while I’d personally refrain from hiring someone to complete certain errands, I’m acutely conscious that such tasks often provide a source of income for many people. However, it appears that such labour is now being replaced by burgeoning technological advancements.
Traditionally, I’ve been uninclined to believe the ambitious predictions about machine learning capabilities, but the recent advances in transcription software have partially swayed my perspective, illustrating the substantial potential of this tech. Its effectiveness as a tool to reduce labour-intensive tasks, thereby freeing time for the more imaginative elements of my work, is beyond question.
Yet, despite its apparent usefulness, the software highlights the exaggerated hype surrounding Large Language Model (LLM) technology on which it is based. As a transcription tool that converts speech to text, it performs a specific, restricted function admirably. It uses the recorded speech to produce a primary text, and based on an automated examination of this text, manages to generate two forms of secondary text, namely a bullet-point summary and a more detailed one. On a smaller scale, it mirrors more extensive LLM technologies such as ChatGPT, which produce secondary texts from various sources across the internet. This is where problems often arise.
Consider, for example, the somewhat comedic introduction of Overviews, Google’s novel AI search tool. This tool incorporates LLM technology to condense extensive search results into a brief summary. Recently, social media has been awash with examples of its bizarre responses to otherwise harmless search terms. As an example, a user searching for a quick remedy to pass kidney stones received advice to consume at least two litres of urine daily, which should be light-coloured. I feel obliged to correct this erroneous information here to avoid an inundation of complaints – the recommended daily intake of urine is indeed between 500 and 750ml, preferably without any added sweeteners.
In spite of expecting efficiency from a dedicated assistant to carry out the tedious yet necessary work that my profession demands, one might be shocked to find misleading suggestions. A prominent instance of error gaining popularity from Overview alleged answering the question of getting cheese to stick to pizza by advising to incorporate roughly 1/8 cup of harmless adhesive to the sauce for enhancing its stickiness. The ridiculous recommendation traces back to a joke a decade old on Reddit by a user aptly titled “f**ksmith”. Overview, akin to all LLMs, pools whatever information it deems suitable from its database and, given its status as a neural network rather than a human, cannot discern valuable results from futile ones – including false information, posts by Reddit users like “f**ksmith” and outright errors.
In a conversation with the Financial Times the previous year, the fantastic American science fiction author, Ted Chiang humourously, yet helpfully, defined AI. He referred to a tweet that described AI as ‘an unfortunate terminology from 1954’ – implying that if the computer scientists from the post-war era who invented this technology had opted for another name, we could have sidestepped a lot of misunderstandings. In his opinion, a more suitable phrase would’ve been ‘implemented statistics’. Even though it’s less evocative, it’s more indicative of the actual operations of these networks, and wouldn’t have led to misconceptions about machines ‘cognising’ or developing self-awareness.
This technology undeniably has its uses and exhibits growing potency as a device saving time and effort. However, regarding ‘intelligence’ – well, it’s still consuming glue and urine, metaphorically.