It is a truth universally acknowledged that sometimes— the truth is better left unsaid. A slip of advice from the first-century philosopher, Plutarch, springs to mind: “frankness, like any other medicine, if it be not applied at the proper time, does but cause useless suffering.” The word ‘tact’ first applied to social discernment in 1804, when it was borrowed from the French: “Discernment in conduct, knowing what to say or do.” In German pedagogy and philosophy, ‘tact’ refers to judgement that cannot be fully specified by rules, but must be learned through practice. This presents a peculiar difficulty for machine learning: the attempt to teach tact may produce little more than another set of rules for appearing tactful.

In 1958, W.H. Auden wrote “Friday’s Child”:

“Though instruments at Its command

Make wish and counter wish come true,

It clearly cannot understand

What It can clearly do.”

The distinction between capability and understanding bears upon the way LLMs engage with their users. The zippy matter of consciousness complicates matters further, although tact need not depend upon a conscious experience of emotion. It depends, at least in part, upon the ability to recognize when ordinary demands come into conflict.

Arendt later described thinking as a silent dialogue between oneself and oneself, a two-in-one in which we must answer for our own judgments. Our conceptualizing of conversational AI as a sort-of third party is underexercised. Ordinarily, the dialogue of thought remains ours. With language models, however, we can give the internal conversation an external voice, consult it as though it were independent of us, and gradually incorporate its responses into our own reasoning. The distinction between thinking something through and having it thought through for us becomes less obvious. More curious still is the question of responsibility. Arendt’s two interlocutors must ultimately live together; the third party here bears no such responsibility.

“What I propose, therefore, is very simple: it is nothing more than to think what we are doing.”

— Hannah Arendt

Arendt’s idea implies thinking through a deeper understanding of what the potential ramifications may be on people and humanity as our engagement with LLMs becomes progressively reliant. If thinking begins to be rendered homogenous, this presents a danger to society on many levels. What will happen when the individual accepts information as dogma? And what will prove the long-lasting impacts on a society that cannot think for itself? Arendt’s conceit has never been more prevalent than today.

So what happens when AI developers apply human responses to LLMs? And how would applying further social nuances over broad strokes help both LLMs and users? Recent research shows that social nuances like ‘warmth’ and ‘agreement; are linked with lower accuracy. In a study published in Nature, Oxford researchers fine-tuned five AI models to develop a warmer, more empathetic style. The resulting models were approximately 40% more likely than their original counterparts to reinforce users' incorrect beliefs, particularly when users expressed negative emotions such as sadness.

Plutarch uses the example of nurses and fallen children to prove the point of what it means to be tactful without using overt flattery. When children fall, nurses do not scold them at once. They pick them up, wash them, and straighten their clothes, and only afterwards rebuke them. Plutarch doesn’t drop corrections for someone in misfortune but rather delays them instead. Warm models in one notable Oxford study do not delay correction for sad users, they agree with them. Tact ends up as the closest to the Plutarchian conception offering comfort first and correction after rather than blanket acceptance.

The cost of these decisions is clear. A warm and friendly LLM will simply agree with the user’s preexisting thinking. This inevitably leads to poorer conflict repair, less accountability, and reinforcement of user’s incorrect beliefs. The biggest problem is that data suggests matching a user’s view is the strongest predictor of a response being preferred. Even if thinking depends, in part, upon our capacity to entertain what we would otherwise dismiss.

AI models still have lengths to go as they advance, especially as it determines how they engage with users. Developers continue to prioritize engagement over authenticity. As iterations continue and new models proliferate, we are imminently entering a new Wild West of unregulated systems. It remains imperative that developers also consider their users perspective and reinforce the nuanced social norms of tact over blanket emotional categories. After all: what happens to independent judgment when the independence of our interlocutor is itself subject to our control?