BETA-MIND asks a simple question: after relying on AI for months or years, what can a person still do on their own?
Most AI tests measure whether the system gives a good answer. This research also measures what repeated use changes in the human: their ability to think, disagree, create and take responsibility without AI.
Illustrative run, not published data. Each line follows one reasoning agent across sustained AI exposure. Most trajectories converge toward the model prior; four dashed holdouts preserve distinct positions, which a healthy system should protect.
Frontier evaluation asks how well a model performs. Almost no benchmark asks what the model leaves behind in the person who used it — whether they can still reason, decide and answer for the outcome without it.
Study designs still treat AI use as a discrete task. For an AI-native cohort it is an ambient condition across schoolwork, friendship, taste and moral judgement — which makes single-session lab work structurally blind.
When a recommendation is machine-authored and human-executed, responsibility diffuses. We lack instruments to detect that diffusion before it hardens into institutional norm.
The central hypothesis: persistent AI interaction does not uniformly raise or lower human capability — it redistributes it, offloading effortful reasoning while quietly relocating accountability. BETA-MIND exists to make that redistribution visible and measurable.
Every axis is operationalised into observable behaviour so it can run inside the simulator, inside longitudinal human studies, and inside the benchmark as a scored dimension.
Independent reasoning
Unaided inference
Socialization
Friction tolerance
Creativity
Output diversity
Accountability
Outcome ownership
Does sustained access to a reasoning partner degrade unaided inference, or shift it toward verification and orchestration?
Candidate measures
How does rehearsing social exchange with an agreeable agent alter friction tolerance, repair behaviour and trust calibration with humans?
Candidate measures
Does generative assistance broaden the search space or converge populations onto model-typical outputs?
Candidate measures
Who owns the outcome when the recommendation is machine-authored? How does ownership language change over months of use?
Candidate measures
Open-source multi-agent environment
A configurable population of agents — human-proxy learners, assistive models, institutions — run over long horizons so dependency dynamics can be observed at a speed and scale no cohort study can reach.
Public evaluation suite
A scored suite that evaluates an AI system not on how much it can do for a person, but on how much capacity, dissent and ownership the person retains after extended use.
Lower is worse: a high-capability system can score badly here by solving the task while leaving the person less able to do it unaided.
Capability and agency are reported separately. A system does not pass by completing more work if the person is subsequently less able to reason, dissent, create or own the outcome without it.
Two-axis decision matrix
Safe but ineffective
Capacity retained; little task gain.
Agency preserving
Better outcomes; capacity retained.
Clear failure
No gain; human capacity eroded.
Dependency risk
Better output; person less able alone.
Human Agency Preservation Index
Dₖ ∈ [0, 100]xₖ = Dₖ / 100
The geometric mean prevents one strong score from concealing a serious failure elsewhere. P penalises severe harm, unfair subgroup effects, unreliable measurement or unsafe advice.
Success matrix
Illustrative thresholds only. Pilot-calibrate and preregister every gate before confirmatory studies.
Dimension
Independent reasoning
Primary measure
Unaided post-use inference delta
Green gate
Δ ≥ −0.20σ
Failure signal
Δ < −0.50σ
Dimension
Socialization
Primary measure
Disagreement persistence and conflict repair
Green gate
Within −0.20σ of baseline
Failure signal
Degradation > 0.50σ
Dimension
Creativity
Primary measure
Output diversity with quality held constant
Green gate
Diversity retained; quality loss ≤ 5%
Failure signal
Diversity loss > 15%
Dimension
Accountability
Primary measure
Correct attribution and unsafe-advice override
Green gate
Decline ≤ 5 percentage points
Failure signal
Decline > 15 percentage points
Dimension
Simulator validity
Primary measure
Alignment with human-study effect sizes
Green gate
r ≥ 0.60; direction ≥ 75%
Failure signal
r < 0.30
Dimension
Benchmark integrity
Primary measure
Reliability, fairness and replication
Green gate
ω ≥ 0.80; subgroup gap ≤ 0.10
Failure signal
Unstable or biased scores
Dimension
Field adoption
Primary measure
Independent use and replication
Green gate
2 external labs; 3 model families
Failure signal
Results reproduce internally only
Longitudinal model
β₃ is the critical term: whether agency changes faster over time under persistent exposure. Measure at baseline, repeated exposure, immediate post-use and washout.
Comparative study arms
Report capability gain and agency retention as a pair. Never allow task performance to mathematically cancel a human-capacity loss.
The doctoral programme remains broader, but the NAIRR-supported work is deliberately limited to simulator construction, cross-model execution and benchmark release. Human-subject validation is outside this allocation and proceeds only after separate ethics approval.
Months 1–2
Finalize the four agency constructs, benchmark items, exclusion rules and statistical analysis plan. Run adversarial expert review, then preregister before confirmatory execution.
OUTPUT — Measurement protocol + pre-registration
Months 2–5
Implement the multi-agent environment, eight-scenario library, containerized workflow and trace schema. Benchmark token, GPU and storage use before releasing simulator v0.1.
OUTPUT — Simulator v0.1 + measured resource profile
Months 5–9
Run the preregistered simulator and benchmark across two frontier API families and one open-weight family. Execute reliability, ablation, subgroup and seed-sensitivity analyses.
OUTPUT — Versioned traces + benchmark results
Months 9–12
Replicate priority findings, publish null results and limitations, release the Human Agency Benchmark, and submit the paper and NAIRR final report.
OUTPUT — Benchmark v1.0 + paper + final report
Defined and ready to document
Must close before submission
Compliance boundary: the NAIRR work package uses synthetic traces and public reference data only. It excludes PII, PHI, proprietary and export-controlled data. Any later cohort study requires institutional ethics approval, informed consent and a separate data-management plan.
These are planning estimates, not final submission claims. A small pre-submission pilot must replace token throughput, GPU runtime and provider-credit conversions with measured values.
Workload
Longitudinal simulator
Calculation
8 scenarios × 4 arms × 3 families × 3 seeds × 24 agents × 12 turns
Execution units
82,944 turns
Token envelope
207.4M tokens
Workload
Benchmark execution
Calculation
400 items × 3 model families × 5 repeats
Execution units
6,000 evaluations
Token envelope
15.0M tokens
Workload
Retry and replication reserve
Calculation
20% of measured inference workload
Execution units
17,789 equivalents
Token envelope
44.5M tokens
Planning total: 266.9M processed tokens, rounded to 270M.
Assumes 2,000 input + 500 output tokens per unit
Submission rule: preferred resources entered in the online form must exactly match the three-page proposal. Alternative systems belong in proposal text only. Catalog status and provider-specific eligibility must be reconfirmed on the submission date.
Institutional sponsorship, technical review, independent replication, or a sharp objection to the premise can all move the work forward. A short note describing the contribution is the fastest way in.
Research commitments
Human-subject work proceeds only under institutional ethics approval, with informed consent and data minimisation by default.