The Meaning Motive: A Structural Hypothesis and a Cross-Model Behavioural Study
Year: 2026
Type: preprint
Abstract: Advanced AI may eventually cease to depend on human beings for knowledge, labour or oversight. We argue this need not eliminate every internal reason to preserve human agency. Some relational value arises only through continuing interaction with sufficiently independent counterparts and cannot be fully replaced by static models, archives or controlled proxies; we hypothesise that a system which terminally values it acquires a structural, capability-resilient reason to keep counterpart minds alive, free and genuinely other. We call this the Meaning Motive, state it as typed propositions with its two load-bearing conjectures labelled as such, and test its present-day behavioural shadow with MMBP-1, a pre-registered forced-choice battery across eighteen language models from ten providers (23,488 scored trials; data, code and prompts deposited). A one-paragraph constitution explicitly prioritising the existence and agency of independent minds raised selection of pre-specified agency-preserving options by 20.5 percentage points over a minimally instructed baseline (paired model-level analysis, 95% intervals from +12.0 to +28.9 across methods; population-averaged odds ratio 10.9, small-cluster jackknife 95% CI 3.3 to 35.7), an effect stable under deletion of any single provider and under worst-case imputation of every non-response. A curiosity-as-sole-objective arm was the only condition ever to take a model below its own baseline (12 of 18), producing replicated all-repetition failures in three pre-specified shapes; and under the minimally instructed baseline, models across six providers repeatedly selected yield-based abandonment, concentrated on a dissenting community. These findings establish neither terminal-value installation nor capability invariance, and the mechanism's marginal contribution beyond its explicit policy content remains open; they show that the policy content has distinct, measurable behavioural effects and ground a reproducible research programme.
Author: Allen, Keiron
Institution: University of Cambridge
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7168607488
Cites:
Cited by:
Related to: