Intelligence is not a single capacity. It is a constellation of cognitive abilities. MHIL's framework measures eight core dimensions, each grounded in established cognitive science. The indicative ranges below illustrate expected performance gaps based on the framework's design.
The ability to temporarily hold and manipulate information for complex tasks. Not storage — an active cognitive workspace for reasoning.
Child–Adolescent range • Human adult: 100
Drawing valid inferences, identifying patterns, and reasoning about abstract relationships — independent of stored knowledge.
Adolescent range • Strongest dimension for AI
Generating novel, useful solutions — especially when conventional approaches fail. Beyond applying known solutions to creating new ones.
Child range • Critical deficit
Evaluating information critically, detecting flawed reasoning, and determining when to comply versus when to question instructions.
Child–Adolescent range
Recognizing when strategies are failing, flexibly switching approaches, learning from errors, and adjusting to changing contexts.
Child range • Major limitation
Monitoring one's own cognitive processes, understanding limitations, and calibrating confidence. The critical weakness for current AI systems.
Minimal–Child range • Fundamental gap
Understanding communication nuances, recognizing emotional context, adapting style, taking perspectives, and navigating social situations.
Adolescent range • Often mistaken for higher due to fluency
Understanding cause-and-effect, distinguishing correlation from causation, reasoning about interventions, and thinking counterfactually.
Child range • Often confused with pattern matching
Despite industry claims of "human-level performance," MHIL's framework is designed around the hypothesis that no current AI system exceeds adolescent-level cognition when measured against genuine cognitive capabilities rather than knowledge recall. Most systems are expected to fall in the child range (MHIL 35–65), with metacognitive awareness — the ability to know what you don't know — consistently in the minimal-to-child range. The evaluation framework is designed to test this rigorously.
Score the reasoning process, not just the final answer.
Tests require genuine understanding, not statistical shortcuts.
Every test calibrated against human developmental stages.
Procedurally generated tests prevent memorization.
Conversational probing reveals true understanding.
Absolute standards, not relative to other AI systems.
Measure what the system produces, not internal states.
From child baseline to post-human capabilities.
Learn how the methodology is being designed to take systems from submission to a certified cognitive profile.
View Process →