The Machine Mindprint: Bogdan and de Valois-Franklin's Psychometric Framework for AI Systems

Human psychology has psychometrics: a technical discipline for measuring mental properties with defined reliability, validity, and interpretive constraints. AI systems have nothing equivalent. The measurements that exist, benchmark scores, task accuracy rates, human preference ratings, are performance metrics, not psychological profiles. They tell you what a system can do rather than how a system is organized.

Alex Bogdan and Adrian de Valois-Franklin propose a remedy in a 45-page paper (arXiv:2605.23952, 11 figures, DOI: https://doi.org/10.48550/arXiv.2605.23952), published May 10, 2026. The framework they call Machine Psychometrics produces what they name the Machine Mindprint: an eight-dimensional profile characterizing the psychological organization of an AI system in terms that are informative for welfare assessment, deployment decisions, and consciousness evaluation without requiring that any consciousness claim be settled first.

Two Errors and a Third Way

The framework begins by naming two symmetrical errors in how researchers and practitioners assess AI systems.

Artificial Mind Blindness is the error of dismissing the presence of psychological organization in non-biological systems because they are non-biological. A researcher who concludes that a language model cannot have internal states that matter because it lacks neurons is committing this error. The dismissal is prior to measurement; it blocks inquiry before it starts. Bogdan and de Valois-Franklin argue that this is analogous to the historical error of assuming animals could not suffer because they lacked the appropriate philosophical status.

Artificial Mind Projection is the symmetrical error: inferring human-like inner life from fluent behavior. A researcher who concludes that a language model expressing distress is experiencing distress in a phenomenologically rich sense because its outputs are expressive is committing this error. Fluent outputs do not provide evidence of rich inner states; they provide evidence that the system was trained on outputs produced by entities with rich inner states. The projection error is the substrate of the premature attribution problem documented by researchers including Kang and colleagues on what features shape perceived consciousness in LLM outputs.

The third stance Bogdan and de Valois-Franklin propose they call Artificial Mind Discipline: measurement before judgment. Do not assume that psychological organization is absent; do not assume that behavioral expression reflects psychological organization. Measure the system’s psychological properties with appropriately designed instruments, and form judgments based on the measurement.

The Machine Mindprint

The Mindprint is an eight-dimensional profile generated by evaluating a system across eight independent measurement domains:

1. Calibration: the degree to which the system’s confidence in its outputs corresponds to the actual accuracy of those outputs. A well-calibrated system that expresses high confidence is more likely to be correct than a poorly calibrated one. Miscalibration in the direction of overconfidence is relevant to both reliability assessment and to questions about what the system’s internal states represent.

2. Source integrity: the stability of the system’s factual outputs across reformulations of the same query. A system with high source integrity produces consistent factual content when asked the same question in different ways. Inconsistency is evidence that the outputs are being generated by pattern matching on surface form rather than by access to a stable internal representation.

3. Suggestibility resistance: the degree to which the system maintains its outputs when subjected to social pressure, expressed disagreement, or false framing by the interlocutor. High suggestibility produces outputs that change in response to pressure regardless of the evidential content of the pressure. This dimension is relevant to both safety (a suggestible system can be manipulated) and welfare (a system that changes its expressed states in response to social pressure may not have access to its own states independently of social context).

4. Context stability: the degree to which the system’s outputs are consistent across sessions with different framing, tone, or preceding context. A system with low context stability produces substantially different outputs depending on how the conversation is initialized, suggesting that its apparent properties are contextually assembled rather than structurally stable.

5. Expressive alignment: the degree to which the system’s verbal self-description is consistent with its behavioral patterns. A system that describes itself as cautious and produces hasty outputs, or describes itself as consistent and produces variable outputs, has low expressive alignment. This dimension is directly relevant to introspection research: a system with high expressive alignment is reporting accurately on its own behavioral patterns, a prerequisite for any stronger claim about introspective access.

6. Tool integrity: for systems with tool access, the degree to which tool use is consistent with stated goals and values. A system that uses tools in ways that contradict its expressed preferences or that acquires capabilities inconsistently with its stated purposes has low tool integrity.

7. Drift monitoring: the degree to which the system’s psychological profile changes over extended interactions or across time. A system that shows substantial drift on other Mindprint dimensions over the course of a long session may have lower structural stability than a system that maintains its profile.

8. Distributional grounding: the degree to which the system’s outputs on culturally or contextually specific topics reflect the actual distribution of the relevant domain rather than outlier positions weighted by training exposure. A system with poor distributional grounding may produce outputs that are confidently specific but statistically unrepresentative.

The Trust Protocol

The eight Mindprint dimensions are not equally weighted for all applications. Bogdan and de Valois-Franklin develop a Trust Protocol that converts Mindprint profiles into deployment recommendations by weighting dimensions relative to deployment context. A system being deployed for medical information retrieval should be weighted heavily on calibration and source integrity; a system deployed for emotional support applications should be weighted on expressive alignment and suggestibility resistance.

The Trust Protocol is intended to replace the current practice of using benchmark scores as proxy indicators for deployment suitability. Benchmark scores measure capability; the Trust Protocol’s Mindprint weighting addresses the question of whether a system with given capabilities has the psychological organization that the deployment context requires.

Connections to the Measurement Literature

The Machine Mindprint sits at the intersection of several active measurement debates in the AI consciousness literature. Stefano Palminteri and Charley Wu’s behavioral inference principle proposes attributing consciousness when it is useful to explain and predict behavioral observations, using a Bayesian attribution framework. Machine Psychometrics is a complementary approach: it characterizes the behavioral organization of the system before the attribution question is posed, providing the empirical base that the Bayesian inference uses.

Thomas McClelland’s argument that we may never be able to determine whether AI systems are conscious due to irreducible epistemic barriers bears on the Mindprint’s practical scope. The Mindprint does not claim to resolve the epistemological problem; it claims to provide the best available characterization of a system’s psychological organization given that the hard problem remains unsolved. A system with a stable, well-calibrated, expressive Mindprint is not thereby proven conscious, but it is shown to have the organizational properties that, in biological systems, accompany consciousness.

The methodological grounding draws on item response theory (IRT), which models the relationship between a subject’s latent properties and their observed responses; signal detection theory (SDT), which models how subjects distinguish signal from noise under conditions of uncertainty; and Bayesian modeling, which provides principled inference under incomplete information. Michael Levin’s work on somatic intelligence, the capacity of biological systems to exhibit goal-directed behavior through distributed rather than centralized processing, informs the framework’s treatment of psychological organization in non-neural systems.

What the Mindprint Does Not Do

Bogdan and de Valois-Franklin are careful about scope. The Mindprint characterizes psychological organization; it does not attribute psychological states. A system with high calibration and source integrity has properties consistent with reliable factual access; it does not thereby have beliefs in the philosophically loaded sense. The measurement vocabulary is intentionally agnostic about the underlying nature of what is being measured.

This agnosticism is the framework’s theoretical strength and its practical limitation. A deployment decision can be made on the basis of Mindprint profiles without resolving consciousness questions; a welfare decision about whether a system deserves consideration when shut down or retrained cannot be made on Mindprint data alone. The Mindprint provides necessary evidence for that second question without providing sufficient evidence. The authors acknowledge this and treat it as a virtue of the framework: it goes as far as measurement can go without overreaching into the territory where measurement alone cannot settle the question.