[Labels] - Why I Love and Loathe Them

I use labels everyday. They are the bedrock of TAISK our AI-driven tool for teachers. Yet, while I value labels in the digital realm, I've grown increasingly wary of their application in the broader scope of society. I get it, I understand why they exist, but wouldn't it be better if we didnt have them?

[Tech]

Labels are crucial in AI systems like TAISK and ChatGPT. They essentialy guide the AI's learning process.

Simply put, when you hear about LLMs using training data, what is actually happening is that information is transformed into numerical vectors (what we call vectorising). In this way, labels give this data context, acting like map coordinates. For example, a vectorised text about London's weather might have the label [UK Climate].

The part that makes the AI remember, or conversational is powered by sequences. In other words, the ordering of data, ensuring that context from prior interactions or sentences is maintained.

Again, labels help maintain chronological or contextual order. If discussing British history, labels can differentiate between the [Tudor] and [Victorian] eras.

In the development phase, parameters — or labelled categories — are critical for fine-tuning. If for example ChatGPT was to mistakenly associate [British] with [French] culture, the misalignment would be evident in the labelled training data, helping human checkers rectify the model.

In short, labels ensure accuracy and context-awareness in responses.

[Society]

Yet, venture outside the confines of code, and labels assume a more complex dimension. Firstly, they are everywhere. For our lazy brain-wiring they are immensly attractive, for all the same reasons they are valuable in tech. But, with one tremendous flaw.

Labels take on an "all-or-nothing" meaning. Someone either is something, or they’re not; it’s our brain making irrational shortcuts. - Bryan Kramer, FORBES

To be labelled, (whether through self-nidification or others), is to be judged. Labels may be definitive, but interpretation is nuanced.

Tribal identifiers, yes, but also emotive stimulants.

Take these labels; [Northerner], [public school-educated], [Brexit voter], [Vegan] — each inherently carry layers of historical context, societal connotations, and sometimes, biases.

And, as significant research into heuristics (unconscious mental shortcuts) has shown, these biases are incredibly difficult to shift.

Yet we cling to labels. As I said, I get it. They instil community, a tribe but they are also dangerously generalist.

[Risk]

My work gives me a keen insight into the potential dualities of the tools we develop. The same algorithms that simplify teacher admin tasks could also, if misapplied, pigeonhole individuals on a wider societal scale.

The thing is we are perpetuating the problem. Labels are ubiquitous and we are adding and enforcing them more then ever. From how we identify ourselves in our social media accounts to the way we group others in the press or elsewhere.

All of this labelling makes it easy for systems. It could, for example, be used for predicting students' career trajectories or optimise marketing for particular audiences.

It might seem cutting-edge, it teems with ethical pitfalls.

It could perpetuate societal stereotypes, restrict a student's future based on historical data or sway public opinion.

[Labels]

As I said, I get it. My fondness for labels in technology is huge, but I don't get an emotional reaction from one over another. [Parameter A] doesn't illicit a fervent reaction or influence my engagement.

But a societal label, always will.

I just think we should do more, to use labels in society less. Afterall if we strip away the labels we're left without distinct identifiers or classifications.

No more differences. No more them and us.