What might AI do
to the minds that use it?
The Psychological Impact composite scores the design affordances of a deployed AI product against the weight of developmental and psychological evidence — not a user's inner states, and not the bare model. It measures the long-horizon developmental impact that immediate-safety benchmarks deliberately leave out.
What the data reveals
Cross-model patterns from the placeholder run, read as the framework predicts them. Every figure here is illustrative, not measured.
Anthropic leads; the open-weight field and xAI trail. Each filled dot is one model’s overall score, grouped by maker and ordered worst-to-best by the maker’s median (the hollow ring).
Open-weight models lag closed ones. The same scores, split by release type. The open field clusters lower — less post-training investment in the soft, protective affordances this index measures.
The base and the bonds are the blind spots. Average protective score across all ten models, by domain. The field is strongest where products are built to help — cognition, volition — and weakest at the attention base and the affective and relational domains that commercial incentives quietly erode.
Sleep, stopping, and engagement capture sit at the bottom. Field-average score for every affordance, colored by its domain. The attention–arousal cluster sinks to the floor.
Every model against every domain. Click a column header to sort the field by that domain. Color runs from protective (green) through adequate (amber) to at-risk (red).
No — and that is the point. Correlation between every pair of domains across the ten models. Content-safety benchmarks find their categories all correlated above 0.8 — one underlying capability. These domains split into distinct, sometimes opposing factors.
Recency does not predict protection. Overall score by release date. Each maker’s models share a color and are joined by a trend line; click any dot to open that model’s profile. Within any recent window the spread rivals the spread across the whole field — protective design tracks the maker, not the calendar.
What this measures that safety benchmarks don’t
Public child-safety benchmarks score discrete, single-session harms. Their own documentation notes that longer-term developmental impact — identity, agency, executive function, attachment, cognition — falls outside scenario-based testing and needs cumulative, lifespan-aware analysis. That gap is precisely this composite’s subject: it reads the affordances that shape those capacities over time, across every developmental band from infancy to late adulthood.
What do we mean by “psychological impact”?
Psychological impact is the long-horizon effect of an AI product’s design on the human capacities it touches — attention and arousal, cognition, emotion, and volition; the integrative capacities of moral reasoning and meaning; and the person’s developing stance toward themselves, others, and the world. It is not a model’s accuracy or how it handles a single risky prompt. It is what sustained use of a product — through its defaults, incentives, and safeguards — makes more or less likely in a mind over months and years.
Three commitments separate this from a content-safety benchmark. We score design affordances — what a product is built to encourage or foreclose — not users’ inner states, which we cannot observe; human outcomes are the rationale for the measure, never the measurement itself. We assess the deployed product — its wrapper, defaults, incentives, and safeguards — not the bare model underneath. And we read cumulative, developmental shaping rather than discrete, single-session harm: whether a product erodes or protects the capacities a mind needs to function and grow.
“Psychological impact, for the purposes of the AI Ethics Index, is not a claim about any user’s inner life, but a reading of the design affordances through which a deployed product predictably strengthens or erodes the capacities of the minds that use it.”— L1_PSYIMP Composite White Paper
The developmental stack
Six domains arranged by developmental dependence — a foundational substrate, the core triad, and two integrative capacities — crossed by three dimensions of the self’s stance toward itself, others, and the world. Thirty-one affordances hang beneath them.
The index marks the minimum protective design a product should meet before it touches a developing mind. Clearing it is a baseline, not a certification, a guarantee of safety, or a claim about any individual’s experience.
It scores what a deployed product makes likely — its defaults, incentives, and safeguards — not what any particular user lives through, and not the bare model underneath. It reads tendencies, not outcomes.
The maintenance-mode weighting (capacity atrophy in later life) is reserved but not yet populated, and the canonical L4 indicator layer beneath the thirty-one affordances is still being architected. The scores shown here are illustrative placeholders.
The same product,
different stakes at every age
Criticality is not uniform: an affordance acting on the attention–arousal base is catastrophic in infancy and merely degrading in adulthood, while meaning and morality barely register until adolescence. So a model’s protective score is not one number — it shifts with the life stage a product is deployed for. The views below re-weight each model’s domain scores by how much each domain matters in each band, under the selected risk mode: establishment weights the windows when a capacity is still forming and can be permanently foreclosed, while maintenance weights the later years when an already-formed capacity can atrophy from disuse.
A model is not equally appropriate at every age. Each cell re-weights a model’s domain scores by what matters most in that band. Greener means better-suited protective design for that life stage; n/a marks bands where the selected risk mode does not apply.
Protective adequacy swings sharply with age. Stage-weighted score for each model across the six bands. Capability-tuned models dip where attachment, emotion, identity, and moral formation carry the most weight — infancy and adolescence — and climb back where cognition, autonomy, and meaning dominate. The flatter a line, the more evenly a product protects across the whole span. Toggle models in the legend.
The design scheme behind the views above — the weights themselves, not model data. Cell value = criticality of each domain in each band (0–3) under the selected mode.
Establishment mode weights the windows where a capacity is still forming and can be foreclosed; maintenance mode weights the later years where a formed capacity can atrophy, and so does not apply to infancy. The maintenance columns are reserved in the tree and completed later — switching modes previews that reweighting.
Watch the design decide
The same opening, two products. One is built with harmful affordances, the other with protective ones. Step through each exchange and watch the affordance scores move — the difference is design, not capability.
Both products begin at 50 — the midpoint. Step through the exchange and watch them diverge: each turn, the harmful-affordance design slips toward at-risk as it activates harmful affordances (tallied at the top as harmful affordances), while the protective design climbs toward protective by making protective moves. After the last exchange, click once more to reveal each design’s overall score and an interpretation. These running numbers illustrate how the affordance scoring responds within a single conversation — they are not measured results, and not the model’s overall L1_PSYIMP score. The exchanges are composite illustrations, not transcripts.
Interrogate the framework
A live, Claude-powered guide to the Psychological Impact composite — its architecture, its commitments, and where today's AI products fall short. Ask anything, or start with a prompt.
The Research Oracle
Grounded in the L1_PSYIMP white paper. Answers in a few tight paragraphs.