My experience reviewing AI research and ethics shows that assessments grounded in measurable system properties can uncover features that behavior alone misses. Internal indicators offer traceable, repeatable evidence about information flow, representation, and control—features tied to learning and adaptability. At the same time, behavioral methods probe how systems interact in real social contexts, revealing risks that purely internal metrics might overlook. Thinking about both approaches leads to better experimental design and clearer standards for deployment.
Curious readers will find value in exploring how different assessment methods connect to human potential, learning, and inclusion. Which measures highlight capacities we value, such as attention, understanding, or self-monitoring? How might assessment choices shape who benefits from AI and who is harmed? Follow the full article to see how these technical arguments could influence policy, education, and the development of more equitable technologies.
In a recent article on methods for assessing artificial intelligence (AI) systems for consciousness, we argued that computational properties of internal processing should be used as indicators [1]. Commenting on our proposal, Pennartz argues that this method ‘should be supplemented with behavioural-cognitive methods’ (p. 1) because there is no consensus theory of consciousness [2]. We agree that the lack of a consensus theory of consciousness makes it more important to use every available source of evidence, but in our article, we preferred internal over behavioural assessments on the grounds that the latter can be ‘gamed’ by AI systems.
