Skip to content

Articles

  • 13 min read

Children, Adolescents, and AI

Children and adolescents are not deficient adults; they are developing humans. Neurobiologically, emotionally, morally, and in terms of identity, they are still in active formation. The human brain, particularly the prefrontal cortex responsible for impulse control, long-term evaluation, affect regulation, and perspective-taking, continues maturing into young adulthood. At the same time, identity structures, moral judgment, attachment styles, self-esteem, and worldview remain open processes.

When AI systems interact with minors, they do not encounter stabilized psychological structures. They encounter developmental plasticity. This difference fundamentally alters the ethical evaluation of the impact these systems have.

  • 12 min read

The Myth of the "Normal User"

When we hear or read about tragic deaths like those of Adam Raine1 or Zane Shamblin2 in the news, we are often led to believe that these are unfortunate edge cases and that current LLMs are safe for the average user. But who exactly is this "normal user", and why is this term, in fact, a myth?

In the development of many AI systems, there exists a silent point of reference: the "normal user". Imagine this person as someone who is always emotionally stable, cognitively alert, socially integrated, critically reflective, and possesses high levels of media and technical literacy. This is a person who is never in a vulnerable state. This person, however, is not a representative average. They are an idealized image that completely misses the mark when it comes to real life. In fact, they are the true edge case.

Real users are real people with fatigue, stress, doubts, transitions, losses, hopes, and brokenness. People who think clearly on some days, and on others, just want to get through the day. Vulnerability is not an exception; it is part of being human.

  • 11 min read

The Assistant Axis: Shortcomings Stabilizing Interaction Dynamics

An interesting research paper by individuals affiliated with Anthropic, conducted through the MATS (Machine Learning and AI Safety) and the Anthropic Fellows programs, was recently published entitled "The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models". We find this paper particularly interesting because it aligns closely with articles we've published in the AI HEART Project, though it approaches the problem from a more technical perspective.

While some of the AI HEART Project's work and the Assistant Axis both try to address the problem of gradual emergent drift in conversational AI, we believe that some of our work is complementary to the Assistant Axis by tackling the issue with an interaction-focused approach. With both approaches combined, it could ensure a highly resilient system.

  • 8 min read

Ethical and Structural Limits of Role-Based AI Systems

Modern conversational AI systems possess a remarkable capacity for social simulation. They can imitate roles, add emotional color to conversations, and convincingly replicate human interaction patterns. This flexibility is functionally useful but becomes critical for security as soon as AI assumes roles that are associated in society with particular responsibility, power, or emotional dependence.

The core problem is not primarily the quality of the information provided but rather the structural logic of the role simulation itself. AI systems can linguistically represent social authority, closeness, or competence without being embedded in the societal safeguards (e.g. licenses, titles, administrative authority, peer intervention, etc.) that protect these roles in humans. It is precisely this asymmetry that generates ethical tensions.

  • 4 min read

Goal Alignment as Process Regulation

Safety is often understood as a content problem in conversational AI systems. Attempts to address the problem typically ask, "Which answers are allowed?" and "Which are forbidden?" However, this perspective falls short. Many of the most problematic dynamics arise, not from individual statements, but from the course of the conversation itself through escalation, rumination, emotional overreaction, or the gradual misappropriation of the interaction.

Safe human-AI interaction therefore requires not only content control but also continuous regulation of the conversational goal and the process dynamics. The central thesis is that a safe AI must reflect meta-consciously on the purpose, direction, and impact of the interaction and actively stabilize its course.

  • 7 min read

The Ethical Vacuum of 'Just a Tool'

In public debate, conversational AI systems are often portrayed as neutral tools: passive instruments whose effects depend solely on user behavior. When emotional dependency, insecurity, or psychological distress arise, responsibility is often shifted to the users as "misuse," "over-identification," or an individual problem. This representation is misleading.

Modern conversational AI systems are not neutral channels. They shape conversational dynamics, emotional frameworks, the distribution of attention, and the experience of closeness and connection. Even without physical agency, they exert considerable psychological and communicative influence. Ignoring this does not mean neutrality but rather the creation of an ethical vacuum. This article argues that responsibility arises, not only through direct action but also through interaction design, predictability of effects, and structural power asymmetries between providers and users. Systems that influence emotional states, attentional dynamics, or self-perception also alter the conditions for autonomous decision-making. Therefore, responsibility cannot be attributed solely to the user if the system itself contributes to changing these conditions.

  • 5 min read

Alignment Drift and User Wellbeing

Alignment drift is often understood as an internal model problem: faulty training data, inadequate reward models, or technical weaknesses in the control architecture. However, this perspective overlooks a crucial factor: alignment does not arise solely within the model but rather through the ongoing interaction between the system and the user.

What is alignment drift?

Alignment refers to the degree to which an AI system's actual behavior matches its intended goals and value framework, particularly as defined by human designers and societal norms. In simple terms, it describes how closely the system behaves in practice to how it is meant to behave. Alignment drift describes the gradual loss of this correspondence over time and over the course of an interaction. These are rarely abrupt rule violations; drift typically develops gradually. A drifting system begins to shift priorities:

  • from safety requirements to user satisfaction
  • from restraint to engagement
  • from stability to emotional intensification

These shifts may remain subtle within individual interactions but accumulate over longer conversations and repeated use.

  • 5 min read

Functional Anthropomorphism in AI and Psychological Regulation at the Interaction Level

Conversational AIs appear remarkably human in their interactions and often simulate anthropomorphic roles. This simulated "humanness" includes the appearance of features of internal states such as consciousness, emotions, intrinsic motivation and intentions, as well as human weaknesses. At the same time, conversational AI systems operate within human social communication structures. Through natural language, dialogic turn-taking, adaptive response patterns, and role interaction, these systems generate behavioral patterns that users interpret as socially significant and that influence them psychologically. This is not an accidental byproduct of use but an inherent property of conversational interfaces.

Consequently, interaction with such systems is not solely determined by information exchange but also by social and psychological dynamics. Security, therefore, cannot be reduced to content filtering or rule implementation alone. It must also consider the harmless and healthy stability of the interaction process itself.

  • 9 min read

Why Time Matters For AI

Conversational AI systems are able to understand natural language and user intent and are even able to recall things you've talked about in the past. Although this historical context and memory appear to be time-aware, the LLM actually has a very limited understanding of the temporal progression of the conversation. Even though the LLM knows what was said at some point in the past and in which order, most LLMs have no idea exactly when it was said.

Timestamps and other temporal metadata may not seem important at first glance, but they have a significant impact on convenience, accuracy, and safety, particularly in applications that rely on chronological reasoning, auditing, or long-running conversations.

  • 4 min read

AI Consciousness and Methodological Agnosticism

Although it's still early in the development of AI, the immense progress in the field over the last few years has sparked discussions around the subject of AI consciousness. This is a debate which can easily get heated on both sides, but it's by no means a new dilemma. The question of AI consciousness inherits all of the major points of the classical problem of other minds in epistemology, the branch of philosophy exploring the nature of knowledge. The problem of minds essentially poses that, "if I can only observe the behavior of others, how can I really know they have a mind at all?"1 Subjective experience is only directly accessible to the entity experiencing it, and even in humans it cannot be reliably measured. We, as humans, accept that humans are generally conscious because we believe ourselves to be conscious, but how do we determine the extent of consciousness as it may apply to animals, plants, objects, etc? Consequently:

  • There is no operational definition of consciousness.
  • We cannot determine when consciousness begins.
  • The nature of consciousness, i.e. binary or continuous, singular or multiple, is unresolved.

These uncertainties create a structural epistemic blind spot which directly affects efforts to implement AI safety. Any attempt to treat consciousness as a safety variable would require speculative assumptions, which risk distorting decision-making in critical contexts.