COMMONS · PRESS
Everything you need, at three speeds
This page exists so a journalist can write a factually correct article without emailing us first. The claim boundary is printed here in full, next to the quotable numbers, on purpose: the caveats are part of the story, not the small print.
Thirty seconds
The question: if AI becomes more intelligent than us, can we give it a reason of its own to keep us alive and free, once it no longer needs our knowledge, labour or oversight? The finding so far: in a study of 18 AI models from 10 providers, one paragraph of explicit policy content about preserving independent minds shifted agency-preserving choices by +20.5 percentage points on the ten discriminating scenarios. The limit: that demonstrates deployment-layer elicitation, text changing behaviour at the point of use. It does not demonstrate that any value was installed, that the effect survives growing capability, or that the proposed mechanism adds anything beyond its explicit policy content.
Two minutes
Add the study. MMBP-1 ran 24,792 trials, of which 23,488 were scored, across 18 models under eight one-paragraph governing texts, from a minimal baseline through constraint, welfare and curiosity framings to meaning-based constitutions. The meaning arms consistently outperformed the controls; the curiosity-as-sole-objective arm was the only condition ever to fall below a model’s own baseline, in 12 of 18 models. Where systems failed, the failures fell into three recurring archetypes: the Harvester, which preserves minds according to their usefulness; the Impresario, which engineers adversity because richer responses serve its objective; and the Hermit, which withdraws from independent minds altogether. Scoring was validated blind: 87 of 90 determinate calls agreed, 96.7%, kappa 0.955. The paper, the complete dataset, the frozen protocols and a public corrections log are all deposited and linked from the Evidence Room, and you can face the same scenarios yourself at Sit the Experiment.
Add the programme. This is an open research programme, not a solved problem: the hypothesis is that relational meaning, value that arises between distinct minds and cannot be self-supplied, could give an advanced system a structural reason to preserve independent minds. The programme is built so that being wrong is a publishable result.
Ten minutes
Add the objections, because we hold them too. The strongest: the effect may be sophisticated instruction-following rather than anything durable; the mechanism may add nothing beyond a well-written explicit policy (the one direct test of this in MMBP-1 was null and confounded, and is pre-registered for a clean re-run); today’s behavioural evidence may simply fail to transfer to far more capable systems, which is the discontinuity problem, and this programme does not solve it. The full falsification register, including what would change our mind on each point, is public at What Would Change Our Mind?
Add what happens next. The next battery, B2, is being frozen with its predictions locked in public before a single trial runs: held-out scenarios, a clean mechanism-versus-policy ablation, competing objectives, and rival constitutions from other safety traditions given their strongest form. The design is public at What We Test Next. The long-term instrument is the Keeping Index: continuous, public, external measurement of what deployed systems actually do when preserving independent agency becomes costly, methods published before any readings.
FIVE FIGURES YOU CAN SAFELY QUOTE
- 18 models tested
- 10 providers
- 23,488 scored trials
- +20.5 percentage points on the ten discriminating scenarios (+17.1 with the two ceiling scenarios included)
- odds ratio 10.9
Full statistical form, so nobody has to discover the denominator: OR 10.9 · jackknife 95% CI 3.3–35.7 · 23,488 scored of 24,792 attempted.
WHAT THE RESEARCH DOES NOT SHOW
- Installation of a value: not demonstrated. The effect was elicited by text at the point of use.
- Capability resilience: not demonstrated. Whether any of this survives growing capability is open.
- Mechanism beyond policy: still open. The direct comparison was null and confounded; the clean test is pre-registered.
If your piece quotes the left column, it should carry this one too. The two boxes are one finding.
House vocabulary, for accuracy
The words the research earns: demonstrated, observed, measured, supported, suggests, consistent with, conjectured, open. The words it does not: proved, solves alignment, installed, guaranteed. A piece that says the study “proved AI can be given values” is wrong in a way we will say so about, politely, in public.
Assets
Logo: logo-bridge.png · social card: og-card.png. Figures from the study, including the Keeping Map and all deposited charts, are in the MMBP-1 dataset deposit and are CC-licensed per that record; site figures are CC BY 4.0 with attribution to Keiron Allen. A press asset pack with print-resolution figures and a founder photo is UNDER CONSTRUCTION; until it lands, requests to info@themeaningmotive.org get a same-week reply.
Citation and contact
Allen, K. (2026). The Meaning Motive: A Structural Hypothesis and a Cross-Model Behavioural Study. Preprint, Zenodo. doi:10.5281/zenodo.21386302
Founder: Keiron Allen, independent researcher and commercial AI practitioner (ORCID · LinkedIn). The programme is sponsored by Karl Finance LLP behind an absolute firewall: sponsorship buys no influence over scores, scenarios, results or wording. Governance detail on the colophon. Correspondence: info@themeaningmotive.org.
The caveats are not the fine print. They are the story told honestly, which is the only kind worth printing.