Politeness Pollution

An evolving position on conversational AI and computational overhead.

Politeness Pollution is the potentially avoidable computational and resource overhead associated with socially motivated language in an AI interaction, including the system's processing and response, assessed against an alternative that preserves the relevant task and communicative purpose.

What Politeness Pollution is

1

Social language becomes computational interaction

A greeting, acknowledgement, or expression of thanks can become model input and may produce another response. The resource question includes what the system processes and what it generates.

2

Input and output are not the same question

Prompt wording can affect model behaviour, while response length can strongly affect resource use. v2.2 treats the independent causal effect of politeness on output length as unresolved and model- and task-dependent.

3

Avoidability requires a comparison

The cost of an extra courtesy turn can be measured, but calling part of that cost avoidable requires a lower-resource comparison that preserves the relevant task and communicative purpose.

What the cited studies measured

These results support parts of the framework under specified models, tasks, and serving configurations. They should not be read as universal values.

ACKNOWLEDGEMENT EXCHANGE

0.245 Wh mean total

Delavande et al. report this mean for tested acknowledgement exchanges using Llama 3.1-8B-Instruct in FP32 on an H100 at batch size 1, with no cross-request prefix caching, combining GPU and CPU measurements with estimated RAM energy.

Controlled configuration — not a universal commercial-assistant cost.

RESPONSE LENGTH

25–60% savings

Poddar et al. report inference-energy savings from response-length control while preserving quality under their study's assessment across tested tasks and models.

Adjacent evidence; not an estimate of courtesy-specific savings.

SERVING CONFIGURATION

Varies by system

Precision, batching, attention kernels, caching, and serving strategy can materially change inference energy. v2.2 therefore treats resource use as workload- and configuration-dependent.

The same interaction can have different costs under different conditions.

Politeness and model behaviour are not uniform across studies: some report more output for polite conditions, others the opposite. v2.2 treats the causal relationship between polite wording and longer output as unresolved rather than a universal property of LLMs.

Three different comparisons

1

Polite wording within an existing request

Does wording change resource use or behaviour? Compare formulations of the same substantive request while preserving instructions, required format, and task difficulty.

2

An additional conversational turn

What does the extra courtesy exchange cost? Compare a completed interaction with and without a subsequent courtesy exchange, including supplied context and the model's reply.

3

Equivalent interaction, different resource use

How much overhead is avoidable? Compare ways of serving the same relevant task and social purpose. This is the comparison required to identify avoidable overhead.

How v2.2 changed the interpretation

v2.2 records the changes explicitly rather than silently carrying forward the stronger claims of v2.1. The full story is in “I Called It Politeness Pollution — Here's How the Evidence Changed the Story”.

TopicVersion 2.1 · Aug 2026Version 2.2 · Sep 2026
DefinitionEmphasis on overhead relative to informational content.Avoidability assessed against a comparator preserving task and communicative purpose.
Energy and cachingPer-token asymmetry supported a predominantly decode-side account.Total cost is context dependent; caching assessed under specified conditions.
Tone and response lengthCourteous inputs were linked to elaborated replies.The causal relationship is unresolved and varies across studies.
Data provenanceConversation data were described as real-world traces.The cited acknowledgement study uses synthetic UltraChat-derived conversations.
Measurement scopeA headline total and per-token coefficients informed the argument.GPU-only modelling distinguished from GPU, CPU, and estimated RAM totals; no universal threshold asserted.

Follow the claims back to the papers and articles.

Origin and version record

Research informing the framework

Delavande, Pierrard & Luccioni (2026a; 2026b) · Poddar et al. (2025) · Borisov et al. (2026) · Yin et al. (2024) · Mehta et al. (2026) · Elsweiler et al. (2026) · Gu et al. (2026) — full citations and links are in the v2.2 paper's bibliography.

Provenance and disclosure

The author discloses his commercial interest as founder of Aignitron Inc., reports no external funding and no new primary data, and states that AI assistance was used for reference verification, structural editing, and typesetting.

This is a systems-efficiency concept. It does not blame users or police how people communicate — it challenges infrastructure to handle normal human communication without always invoking maximum compute.