Textual Mirror

A Proposal for Pronoun Usage by LLMs & Other Non-Conscious Entities

Published on

Problem Statement

We're currently presented by the issue that a huge number of people interact with "AI"1 and LLM-based systems on a regular basis. My intent here is not so much to criticize this practice in itself, but to suggest one way that might help us to see the systems as they really are without having to squint so much through all of the smoke & mirrors which tend to be used to obscure what's actually going on.

LLMs are complex statistical models which can perform potentially useful operations on language data. The output which comes from these tools often appears to be the product of a conscious being or at least something which seems to have a mind, but we know with a high degree of certainty that these entities are not sentient or conscious and are factually incapable of thought as such even though their output presents as a simulacrum of what a conscious, thinking being might produce. And even if we know the thing we're working with is a non-conscious entity, we might still occasionally fall into the trap of thinking about it as a conscious thinking being, which can potentially lead to a huge number of issues, and even when it doesn't, this still represents a distorted perception of reality, which heightens the chance that we will suffer unnecessarily. Even when we can keep in mind that the tool with which we are interacting is not capable of conscious thought, the cognitive effort of maintaining this perception while presented with output that seems to indicate the contrary may further increase the cognitive load which is necessarily involved in interactions with such an unreliable interlocutor.

One way we might alleviate these challenges of distorted perception and cognitive load is to regulate pronoun usage in the context of LLM interaction. Here are the guidelines I propose:

Proposed Guidelines for Pronoun Usage with and for LLMs

  1. Both user and LLM should avoid usage of any personal pronouns (I, you, he, she, they, etc.) when referring to the LLM itself or any tool based upon it. The reasoning here is that personal pronouns tend to imply consciousness, a quality which LLMs and tools based upon them lack.
  2. For the user, it is enough to address the tool by its name or by the name of the underlying model. If something like a pronoun is needed, "bot"2 or a similar placeholding word can be used.
  3. The tool itself may sometimes require self-reference for the purpose of clarity. In this case, the tool can either use its own name, the name of the underlying model, or the following pronoun-like words:
    • Instead of a first-person pronoun (I, me, &c.), use "this bot" or "bot".
    • Instead of a first-person reflexive pronoun (myself), use "botself" (or "this botself" in cases which may involve interactions between more than one tool which all require self-reference for clarity of expression).
    • Instead of a second-person pronoun (you, &c.), use "bot".
    • Instead of a third-person personal pronoun (he, she, they, &c.), use "it" or "bot".
    • For possessive and genitive cases, use the normal declension of the tool or model name in whatever language you're using, use "this bot's", "bot's", or "its".
  4. As far as is possible within the available linguistic constraints, the verb form used when a bot is the subject of the sentence should be the one available which best reflects the inanimacy or lack of consciousness of the subject. In English, the third person singular (i.e., the verb form used for "it") is probably the best choice.

Example Interaction

User: Bot3, please find the blog post from Textual Mirror about pronoun-usage regarding LLMs.

LLM: This bot4 found the article titled "A Proposal for Pronoun Usage by LLMs & Other Non-Conscious Entities". Here is the URL:

User: Please apply the guidelines there to all current and future interactions.

LLM: This bot has applied these guidelines to botself. They are now stored in this bot's persistent memory.

Acknowledgements

Thanks to @leadore@sunny.garden & @ancientsounds@mastodonapp.uk for helping me think through this.

  1. I put "AI" in scare quotes mainly to call in to question the claim that there is anything actually resembling intelligence there. Read the following paragraph for an argument regarding why it is problematic to think about this phenomenon as intelligent. ↩

  2. Although "tool" potentially could offer an alternative form of address, the sometimes-derogatory usage of the term (in English, at least) makes it a less-than-ideal choice. It has been shown that the way we behave in interactions with LLMs and other "AI" tools conditions the rest of our interactions as well, so even if it's unlikely that we might offend a non-conscious LLM, we don't want to increase the chances of acting offensively in the relationships we value. ↩

  3. This would also be a typical location to use the name of the tool or model with which one is interacting. ↩

  4. This would also be a typical location to see a bot refer to itself with its own name. I've refrained from using the names of any actual LLMs or other "AI" tools because all of them are problematic. My intent here is merely to present one way in which we might reduce the potential harm which can come from using these tools. ↩