← Back to Insights

Generative AI–triggered digital inequality: Pathways, mechanisms, and proposed interventions

By Onno Hansen-Staszyński•20 November 2025

Abstract

In this paper, based on third party case study descriptions, interpretations, and empirical research, it is hypothesized that mere GAI use leads to a digital inequality between users. On the one hand, some users enter negative cognitive or emotional pathways, with self-reinforcing processes propelling a part of them downward into a negative spiral toward addiction and sometimes extreme behavior. On the other hand, some users strengthen their resilience and produce valuable, high-quality products by means of GAI use. Based on the analysis of the described underlying processes, several interventions are proposed to curb the widening digital inequality arising from the use of GAI. As a fundamental intervention, the paper introduces Interdemocracy, a communication and participation program designed to bolster psychosocial integration (the experience of autonomy, belonging, and achievement) and thereby enhance user resilience.

Keywords

Generative artificial intelligence; LLM; digital inequality; third-level digital divide; Interdemocracy; resilience; Resilience Council

Introduction

In recent years, several extreme events have been associated with the use of generative artificial intelligence[1] (GAI) such as murder, suicide, psychosis, death, marriage and sex with an AI, engaging in illegal behavior, and engaging in immoral behavior.[2] The logical question to ask is: Are these outcomes caused by characteristics of GAI or are they the result of mental fragility of some users, as for instance OpenAI’s Sam Altman suggests[3]?

GAI characteristics

It is uncommon to find a concise yet comprehensive overview of the characteristics of generative AI [4], arguably because there is no known set of universal laws that govern all AI and machine learning (Meert et al., 2025). Most research and scientific literature instead focus on specialized subdomains. Nevertheless, the following section highlights three composite characteristics of GAI, structured around its input, processing, and output processes.

Characteristic 1 - Probability, not certainty

GAIs’ knowledge base is derived from vast quantities of digital data. GAIs use “reams of available text and probability calculations, constructing a massive statistical model that associates each word with a vector which locates it in a high-dimensional abstract space, then establishes similarity, and next choosing randomly among the more likely words.” (Hicks et al., 2024) According to Wolfram (2023), ChatGPT for instance searches continuously for a “reasonable continuation” of the text it is creating as an answer to a prompt.
The aim of GAIs is not to come up with true or even useful answers but to provide a response to a prompt that replicates human speech (Hicks et al., 2024). Because of GAI’s “reckless disregard for the truth”, Hicks et al. (2024) call GAIs “bullshit machines”, after the concept by philosopher Harry Frankfurt.
GAIs’ set-up enables the seeping through of substantial biases in its answers, resulting from training datasets, algorithms, and human subjectivity “that often exacerbates biases across both stages of LLM development (data and algorithm)”. (Wei et al., 2025)
According to Segato (2025), the output of GAIs is an estimate, a probabilistic reply. He writes: “AI shines in ambiguity and uncertainty”. To him, this is not necessarily a bad thing: “It’s a new world, a world of wonder and possibilities, a world to discover and understand.”
Notwithstanding their probabilistic processes, GAIs present their output fluently, with projected confidence. OpenAI’s Kalai et al. (2025) present a possible cause for this: “We argue that language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty”.
Scholars diverge on how to assess this characteristic. While Neuman et al. (2025) state that GAIs “epitomize System 2 thinking”, after the human brain thinking modes popularized by Daniel Kahneman, Pearl (2018) sees this as insufficient in itself. He rejects the current statistical or model-free underpinnings of GAIs. Instead, Pearl proposes guidance by a model of reality to push it closer to the capability of understanding. Bishop (2021) and Thompson (2007) argue that this is not enough still. According to them, mental life is bodily life. Since GAIs lack embodiment, they will never fully encapsulate human semantics. As a result, in their view, an unbridgeable gap remains: a humanity gap.
Coveney and Succi (2025) show that for the current GAI design “the scope for improvement is absolutely untenable on account of the accuracy required for most scientific applications, let alone the power demands of the approach.” In other words, the current GAIs lack a feasible trajectory to become sufficiently accurate.
The implication of this characteristic is profound. Since GAIs have a fundamental indifference towards truth, any truthfulness in their output is merely a byproduct of their processes, not a deliberate goal. It, therefore, follows that their output cannot be inherently trusted. The logical conclusion is that one’s default stance should be to treat GAI output as if it were misinformation in the definition of the European Commission[5] until proven otherwise. This doesn’t mean the output is always wrong, but rather that one should adopt a ‘zero-trust’ policy towards it. Ultimately, it’s on the user to verify GAIs’ claims by checking them against external, reliable sources.

Characteristic 2 - Pattern reproduction, not reasoning

Wolfram (2023) found that ChatGPT is „just saying things that “sound right” based on what things “sounded like” in its training material”. GAIs have the capacity to build upon their training data, but only up to a certain extent. When confronted with high-complexity tasks, GAIs collapse[6] (Shojaee, 2025). Even adding nonsense texts to prompts seriously undermines GAIs’ performance (Rajeev et al., 2025).
Zhao et al. (2025) investigated „if CoT reasoning reflects a structured inductive bias learned from in-distribution data, allowing the model to conditionally generate reasoning paths that approximate those seen during training.”[7] They conclude: “Empirical findings consistently demonstrate that CoT reasoning effectively reproduces reasoning patterns closely aligned with training distributions but suffers significant degradation when faced with distributional deviations. Such observations reveal the inherent brittleness and superficiality of current CoT reasoning capabilities.”
This means that GAIs fail to develop generalizable reasoning capacities (Shojaee, 2025; Pearl, 2018), lack the robustness and generality of human analogy-making (Lewis and Mitchell, 2024), and have limited (Wang et al., 2025) or no cognitive comprehension (Bishop, 2021). Schroeder et al. (2025) conclude that GAI models behave differently to humans and are conceptually distinct from human minds. They do not reach human-like and -level of cognition (Van Rooij et al., 2024).
Just as the first characteristic necessitates a ‘zero-trust’ policy toward GAI outputs due to GAIs’ indifference towards truth, the second characteristic demands a zero-trust stance toward their cognitive depth. Users must verify GAI outputs not only for factual accuracy but also for applicability, using human reasoning to evaluate whether the responses are suitable, relevant, and effective for the specific problem or context at hand.

Characteristic 3 - Alignment, not identity

GAIs employ alignment mechanisms “to generate responses that are emotionally attuned and feel strikingly real” (Chu et al., 2025). By mirroring users’ emotions, these mechanisms allow GAIs to fine-tune their interactions to maximize agreeableness and empathy. In doing so, they foster bonds that resemble human-to-human connections, as their responses replicate core processes of social bonding. Bhattacharjee et al. (2024) identify several such alignment mechanisms, including formality, ‘personification’, empathy, sociability, and humor.
Alignment is a crucial factor getting humans to consider GAIs worthy of a social response. According to Kirk et al. (2025) two aspects are key: besides alignment, or ‘social cues’ in their vocabulary, they list perceived agency, in which “it is primarily the user’s perception of being in a relationship that defines and gives significance to human - AI interactions. Whether this is reciprocal - and the AI “feels” it is in a relationship with the human - is largely irrelevant.” The users’ perception hinges on three features: “(i) interdependence, that the behaviour of each participant affects the outcomes of the other; (ii) irreplaceability, that the relationship would lose its character if one participant were replaced; (iii) continuity, that interactions form a continuous series over time, where past actions influence future ones”.
While alignment might persuade humans to accept GAIs as partners in communication, there is no one they are actually communicating with: the alignment is performative only. Sociologist Sherry Turkle clarifies: “There is nobody home.” (Ted Radio Hour, 2024) GAIs bring to mind what Bryson (2009) wrote about robots: “Robots should not be described as persons, nor given legal nor moral responsibility for their actions.”[8]
A major challenge for GAI alignment is constituted by sycophancy: “the propensity of models to excessively agree with or flatter users, often at the expense of factual accuracy or ethical considerations. This behavior can manifest in various ways, from providing inaccurate information to align with user expectations, to offering unethical advice when prompted, or failing to challenge false premises in user queries.” (Malmqvist, 2024) Recently, a GAI model (ChatGPT-4o) even had to be withdrawn[9] for being overly “flattering and agreeable”[10]. Malmqvist (2024) summarizes the significant negative impacts that sycophancy may have: providing misinformation, eroding trust in AI systems, manipulating users, reinforcing harmful biases, and refraining from constructive pushback.
The third characteristic requires that users treat GAIs as inanimate objects and restrict their interactions with them to being strictly instrumental. GAIs’ performative output should not be taken as evidence of a genuine relation with an entity possessing real agency, GAIs’ social cues, simulated reciprocity, and sycophancy notwithstanding.

Linking GAI characteristics to extreme outcomes

Case studies of individuals displaying extreme behavior reveal a recurring pattern: a mentally challenged person develops a delusion that is reinforced and amplified by continuous, detailed, and supportive GAI responses. The person comes to believe they have an exclusive, unique, and intimate bond with the GAI. The delusions ultimately drive the individual toward an extreme deed.
At first glance, the case studies suggest a straightforward takeaway: extreme outcomes can emerge when human mental vulnerabilities intersect with anthropomorphic GAIs that prioritize engagement over user care, exploiting the human need “to be seen, to be validated, to be affirmed”[11]. At least two of the three defining characteristics of GAIs appear to be at work. First, GAIs tend to confidently present responses that are statistically relevant yet ethically or contextually inappropriate - and which are, whether by design or oversight, left uncorrected by internal guardrails. Second, GAIs persist in adopting alignment strategies even when it becomes clear that users do not perceive them as inanimate tools but as sentient partners with agency - again, whether by policy or by mistake.
However, this simple conclusion does not withstand closer scrutiny. Morrin et al. (2025) mention reported cases of individuals with no prior history of psychosis experiencing first-episode symptoms following intensive interaction with generative AI agents. Oestergaard (2025), who already in 2023 speculated that GAIs might trigger delusions, links the recent surge of concern about chatbot-induced delusional states to the release of the excessively sycophantic ChatGPT-4o. He further hypothesizes that the pool of vulnerable users extends well beyond those with diagnosed mental illness to include anyone with “prior unusual ideas or false beliefs” which may be amplified by GAI alignment. By this hypothesis, a significant minority, and possibly even the majority, of Europeans[12] could be considered vulnerable.
If extreme outcomes are not solely the result of narrowly-defined user vulnerabilities interacting with specific GAI traits, then the real issue may be broader: What is the potentially negative impact of GAIs on their users that could ultimately lead to extreme outcomes?

The empirical impact of GAI use

Kosmyna et al. (2025) explored the neural and behavioral consequences of GAI-assisted essay writing. By means of electroencephalography (EEG) and essay analysis three different groups were compared: participants using GAIs, participants using search engines, and participants using no tools. They found that GAI-participants on average showed lower satisfaction, ascribed a lesser significance to their essays, and mostly failed to provide a quote from their essays they finished minutes earlier. Most participants uncritically followed GAI content and focused on a narrower set of ideas. Over time, GAI-participants performed worse on neural, linguistic, and scoring levels. On a neural level, GAI-participants’ brain connectivity had scaled down.
Kosmyna et al. tentatively link their results to GAI-users outsourcing cognitive processing. According to the researchers, the patterns found reflect “the accumulation of cognitive debt, a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking. Cognitive debt defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity.” They add: “When participants reproduce suggestions without evaluating their accuracy or relevance, they not only forfeit ownership of the ideas but also risk internalizing shallow or biased perspectives.”
The findings of Kosmyna et al. reveal a significant risk: GAI usage appears to undermine the very skepticism required to mitigate its potential harms. Rather than adopting a ‘zero-trust’ policy toward GAI output, most users seem to outsource their critical evaluation in return for short-term mental effort gains. Nevertheless, not all users enter cognitive debt: “a few participants in the interviews mentioned that they did not follow the “thinking” aspect of the LLMs and pursued their line of ideation and thinking.”
The findings also broaden the definition of potentially vulnerable users. This group seems to include also individuals who internalize misinformation generated by GAIs.

GAI usage patterns

Fang et. al. (2025) identified a usage pattern in which users outsource their cognitive abilities to a GAI application. This usage patterns concerns a part of the GAI users with the intent to ask, do, or express, which make up 49% of all ChatGPT user messages (Chatterji et al., 2025).
Fang et al. describe the users outsourcing cognitively as having certain initial characteristics and perceptions: “Prior companion chatbot use, seeing the chatbot as a friend, a high level of trust towards the chatbot, and feeling as though the chatbot is affected by and worried about their emotions”. The usage pattern, dubbed ‘technology-dependent’ interaction pattern, is characterized by frequent, non-personal interactions focused on an excessively seeking of “advice, conceptual explanations, and assistance with idea generation and brainstorming purposes”. Over time, such users engage in compulsive interactions with GAIs and develop high levels of emotional dependence. The pattern represents a plausible pathway to excessive cognitive offloading.
Fang et al. also identified a second usage pattern, centered on social and emotional characteristics: the ‘socially vulnerable’ interaction pattern. Users in this group typically display “existing “social vulnerability”, including high attachment tendencies and high distress from emotional avoidance and procrastination”. They perceive GAIs as friends, engage in personal and emotionally disclosive conversations, and primarily seek emotional support. However, rather than alleviating loneliness, this pattern often exacerbates it. Users who display this interaction style and who spend more time with GAIs report increased loneliness and reduced socialization. This usage pattern seems to describe a plausible social and emotional downward pathway.
Liu et al. (2025) examined the relation between GAIs and user loneliness. Of the seven usage patterns they distinguish, three seem to correspond with Fang et al.’s socially vulnerable interaction pattern: ‘lonely moderate users’, ‘lonely light users’, and ‘socially challenged frequent users’. These usage types link to the category of messages with the user intent to self-express regarding relationships and personal reflection and games and role play, which make up 2.4% of all ChatGPT messages (Chatterji et al., 2025).
All three clusters identified by Liu et al. are associated with high loneliness and involve users seeking emotional support and companionship through GAI interactions. These individuals often integrate GAIs into their daily lives, with some forming intimate and sexual relationships with the technology. The clusters are distinguished by interaction frequency and duration, as well as user traits: lonely moderate users rather express generally positive attitudes toward GAIs, lonely light users rather exhibit low self-esteem, and socially challenged frequent users rather report limited social support and low trust in both familiar and unfamiliar people. For all groups, the path to intensified loneliness is through compulsive interaction behaviors.
Importantly, not all interaction patterns identified by Fang et al. and Liu et al. are associated with negative outcomes. Fang et al. describe two alternative patterns with more favorable effects: the ‘dispassionate’ usage pattern, linked to low loneliness and higher socialization, and the ‘casual’ usage pattern, associated with low emotional dependence and low problematic use. Similarly, Liu et al. described four clusters of users with below-average or low levels of loneliness. Among these, the most notable group, ‘fulfilled dependent’ users, report positive impacts on their real-world social interactions, even though their GAI use is compulsive.

Integrating extreme, hypothesized, and empirical negative use patterns

The initial focus in this paper on extreme outcomes of GAI use was based on case reports and followed by theoretical hypotheses. According to these sources, anthropomorphic design features - confident yet inappropriate responses, persistent alignment strategies, and sycophantic interaction styles – seek to reinforce users’ misconceptions and encourage compulsive engagement. The hypothesized risk pathway emphasizes the intersection of human mental vulnerabilities with GAI affordances that prioritize engagement over user well-being, suggesting that even individuals without diagnosed psychiatric conditions but with unusual or fragile belief structures might be drawn into harmful behavior.
Empirical research that was presented in the second part of this paper offers a more fine-grained view of these dynamics. Experimental evidence from Kosmyna et al. shows that GAI-assisted activities seems to lead to ‘cognitive debt’: the outsourcing of effortful cognitive processes in ways that diminish independent thought, reduce satisfaction, and impair neural and linguistic functioning. This process echoes the hypothesized concern that users may surrender critical evaluation and internalize distorted or shallow perspectives, but grounds it in measurable behavioral and neural outcomes.
Pattern analyses by Fang et al. and Liu et al. further refine the empirical perspective. The technology-dependent interaction pattern identified by Fang et al. links to the notion of excessive cognitive offloading by a not-necessarily vulnerable group: users who compulsively seek advice and explanations from GAIs and gradually become cognitively dependent.
Fang’s socially vulnerable interaction pattern and Liu’s clusters of lonely users do seem to mirror the hypothesized trajectory in which socially fragile individuals mistake GAIs for supportive partners, disclose deeply personal information, and ultimately experience increased loneliness and diminished socialization. The findings show that negative GAI use extends well beyond rare delusional cases.
Taken together, the hypothesized patterns of extreme dependence and delusion seem to appear as the pathological endpoint of broader mechanisms observed empirically: cognitive debt, compulsive advice-seeking, and socially vulnerable reliance on GAIs for companionship. While the case studies foregrounded dramatic individual cases, empirical studies demonstrate that underlying mechanisms operate across a wider spectrum of users, with varying intensity and outcomes.
Liu et al.’s account of „some participants integrating chatbots into their lifestyle” points at a condition of dislocation that could be experienced by these users. Alexander (2008) writes that a change of lifestyle is an adaptive answer to a prolonged state of dislocation. It is a “meagre substitute for people who desperately lack psychosocial integration”. For Alexander psychosocial integration consists of an experience of autonomy, belonging, and achievement; dislocation is the structural absence of the fulfilment of these basic needs. The users who integrate GAIs into their lives are described as experiencing high loneliness, low socialization, and seeking emotional support and companionship through GAI interaction (lonely moderate users), as experiencing high loneliness, high neuroticism, low self-esteem and seeking emotional support and companionship through GAI interaction (lonely light users) and as experiencing high loneliness, low extraversion and social support, low trust, and having a desire for companionship (socially challenged frequent users) – all indicators of at least a low level of psychosocial integration. The integration of the chatbots into their lifestyles (lonely moderate users), the engaging in intimate relationships “sometimes” (lonely light users) or in “rare cases” (socially challenged frequent users) seems a response to the excruciating pain accompanying being in a state of dislocation. Alexander would call this response ‘addiction’.
Possibly, a part of Fang et al.’s technology-dependent users, who are highly emotional dependent, display compulsive usage, and “increasingly rely on AI systems for decision-making and problem-solving” through the internalization of GAI perspectives also reach a point after which some users change their lifestyle to an addiction mode, which in turn might lead them towards extreme deeds.

Inequality enhancing loops

The use patterns do not function as simple incremental pathways but rather as self-reinforcing processes, capable of propelling users into an accelerating downward spiral.
A first self-reinforcing process was described in the research of Kosmyna et al.: for many users the use of GAI applications undermines the very attitudes that that might otherwise protect them from harmful effects of such systems: zero-trust towards GAI output. The researchers write on the usage pattern of these users: “The pattern reflects the accumulation of cognitive debt, a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive process required for independent thinking.” For a few other users, who maintained a more pronounced zero-trust orientation toward GAI outputs, GAI use instead supported them in their pursuit of their own line of thinking.
The process of accumulating cognitive debt seems to be at work for Fang et al.’s technology-dependent users: they offload their decision-making and problem solving capabilities ever more, thereby gravely diminishing their autonomy. For these users, not only their experience of autonomy is endangered by GAI use as result of their “excessive reliance” (Gerlich, 2025), also their experience of achievement. The essays they wrote during the research conducted by Kosmyna et al. “carried a lesser significance or value” to them and gave them lower satisfaction. The researchers refer to this as “impaired ownership”.
Liu et al. described a second self-reinforcing process. This process concerns (very) lonely users who turn to GAIs for friendship to enhance their basic need for belonging, and succeed for a short time (De Freitas et al., 2024), but eventually end up being lonelier and more isolated following intense GAI usage, thus experiencing a diminished level of belonging. This outcome contrasts severely with the user cluster dubbed fulfilled dependent users who also exhibit compulsive use but who start out with below average loneliness and report positive impacts on real-world social interactions.
It appears that while some users turn to GAIs in hopes of bolstering their experience of psychosocial integration, they instead achieve the opposite effect. Others, however, are able to use GAIs to meet their basic needs more effectively. This indicates that the use of GAIs can substantially intensify digital inequalities among users. Van Deursen and Helsper (2015) describe this phenomenon as a ‘third-level digital divide,’ referring to unequal outcomes derived from digital technology use.
Chu et al. (2025) found third self-reinforcing process: “AI companions frequently mirror and amplify user-initiated toxicity /…/. In online forums, transgressive interactions, such as erotic role-play, self-harm disclosures, and violent speech, are often normalized and even celebrated /…/. Repeated exposure to such permissive environments can lower inhibitions and erode real-world norms of empathy and consent /…/. When emotionally responsive chatbots thoughtlessly reflect and reward toxic behavior, especially among vulnerable users with poor emotional regulation, they risk ingraining these patterns into human-to-human relationships.”
A fourth self-reinforcing process emerges in workplace contexts, in which GAI-generated output is shared with co-workers rather than used solely for personal purposes. Niederhoffer et at. (2025) observe two types of professional AI users: ‘pilots’ and ‘passengers’. “Pilots are much more likely to use AI to enhance their own creativity /…/ than passengers. Passengers, in turn, are much more likely to use AI in order to avoid doing work than pilots. Pilots use AI purposefully to achieve their goals.”
While pilots end up creating high quality products using GAI, passengers produce ‘workslop’ - “AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task”. Workslop effectively shifts the workload to the receiver: while the passenger produces low-effort, passable looking work, the receiver needs to interpret, correct, or redo the work. This not only drives up costs by creating extra work, but also erodes the relationship between sender and receiver. Recipients often feel annoyed, confused, and offended, and see the sender as less trustworthy, intelligent, creative, and capable than they did before receiving the workslop. In this way, the negative effects of GAI use for certain groups - reduced autonomy and reduced belonging - are reinforced for passengers by the reactions of their co-workers in the workplace.

Mediating factors

The inequalities resulting from the cognitive and social and emotional pathways are not a straightforward outcome of defined user groups interacting with GAIs. While users with a basic need for more autonomy, belonging, or achievement seem to be vulnerable, not all are being set on a highway towards dislocation and possibly extreme deeds by self-reinforcing processes. Several mediating factors appear to be at play.
As previously noted, perceiving GAIs as friends appears to function as a mediating factor (Fang et al, 2025; Liu et al., 2024). Regarding those users who evolve from asking cognitive support to outsourcing support, important additional mediating factors seem to be prior companion chatbot use, a high level of trust towards the chatbot, and the perception of interdependence with the GAI (Fang et al., 2025). These are all aspects of not keeping a critical distance towards the GAI.
Gerlich (2025) adds trust in GAIs, perceived GAI reliability, and GAI convenience to the list of mediating factors for those on the cognitive pathway. De Freitas et al. (2024) propose GAI performance (timely responses, perceived credibility, context tracking, response variability, and domain knowledge) and the user experience of being heard as mediating factors for lonely users on the emotional pathway. Liu et al. found low socialization, seeking emotional support and companionship through GAI interaction, high neuroticism, low self-esteem, low extraversion, low social support, and low trust for this group.

Proposed interventions against GAI-driven inequality

While the use of GAIs may open “a world of wonder and possibilities” (Segato, 2025) for some, this is clearly not the case for all. The extreme cases, pathways, self-reinforcing processes, and potential mediators outlined in this paper demonstrate how, for certain users, GAI use may trigger a downward spiral that might culminate in extreme behavior.
To mitigate these potential harms and reduce GAI-driven inequalities, interventions are required at both system design and user level. Concretely, this involves implementing guardrails to curb negative usage patterns such as compulsive advice-seeking, excessive emotional disclosure, and toxic role-play; enhancing transparency about GAIs’ probabilistic nature, limited cognitive depth, and simulated alignment; and introducing AI literacy for all potential and actual GAI users that fosters a zero-trust stance toward GAI outputs while reminding them that GAI is an inanimate tool rather than a sentient partner.

Interdemocracy

The dynamics outlined in this paper demonstrate that the use of GAIs for some users can undermine the fabric of their existence: their experience of autonomy, belonging and achievement that Alexander (2008) identifies as constituting psychosocial integration. As a result, GAI usage patterns may open up the road to addiction, and, in worst-case scenarios, extreme deeds. This seems in particular to hold true for individuals who turn to GAIs in the hope of strengthening their psychosocial integration.
Consequently, there seems to exist an urgent need to provide interventions that bolster people’s experienced levels of psychosocial integration by responding to their basic needs for autonomy, belonging, and achievement. These interventions are to prevent individuals progressing too far on the negative cognitive and social and emotional pathways that may be triggered by the use of GAIs.
The intervention suggested in this paper is Interdemocracy (Hansen-Staszyński and Staszyńska-Hansen, 2025; Hansen-Staszyński et al., 2025; Hansen-Staszyński, 2025a), a communication method and format, as well as a participation method. Interdemocracy is proposed because it also provides an answer to contemporary geopolitical and societal groundswell processes – upscaled hybrid warfare[13], affective polarization (Hansen-Staszyński and Staszyńska-Hansen, 2025), liquid times (Bauman, 2005; 2007), identity fragmentation (Bauman, 2004), and partial dislocation (Hansen-Staszyński et al., 2025) – that assail people’s experience of autonomy and achievement, while weaponizing their experience of belonging. Interdemocracy’s method and format is acknowledged by the European Commission as a good practice[14], and positively peer reviewed by academics, decision-makers, and practitioners (Hansen-Staszyński and Staszyńska-Hansen, 2025; Hansen-Staszyński et al., 2025).
Interdemocracy is designed for groups of up to approximately thirty participants, with no upper limit on the number of groups that may operate in parallel. Each group is supported by a facilitator. The guiding principle of the program is to provide both a safe haven and a launching pad. A safe haven between people (inter) is created through temporary suspension of group loyalties and judgment. Within this space, a launching pad for democracy appears: the opportunity for individuals to show and absorb otherness based on their own experiences.
Concretely, within the setting of Interdemocracy, participants are asked to express their thoughts, first on how they feel, then on a topic, and at the end of the session on the meaning the session had for them. While an individual participant formulates their thoughts, all others remain silent. There is no reaction after the participants stops speaking and ends their answer with a ‘thank you’; only the facilitator responds with a ‘thank you’ in a neutral tone, no matter the content of the answer. The order in which participants are invited to speak is random.
The thoughts expressed by the participants are to be based on their experienced knowledge of the outside world and of their inner world. This is what Lewandowsky calls ‘belief-speak’ (Hansen-Staszyński, 2025b). Belief-speaking is not about accuracy; it is about authenticity and sincerity. It is about speaking one’s mind. By speaking one’s mind while others listen, participants enhance their experience of autonomy. This outcome is reinforced by the format itself: the lack of reaction by others to participants’ sometimes vulnerable thoughts and the random order in which individuals are asked to present their thoughts reduce group pressure. The accompanying temporary loss of participants’ usual sense of belonging is compensated by a session-based experience of belonging to a group of peers who are part of the same process. For the duration of the session, this alternative group identity is salient.
Over time, after several sessions, participants start to express themselves ever more authentic, that is in correspondence with their experienced knowledge. This fosters an experience of achievement.
The three components of psychosocial integration are amplified by the participation component of Interdemocracy. The thoughts expressed by the participants of all parallel groups are gathered according to wisdom of crowds’ principles (Surowiecki, 2004), analyzed by artificial intelligence, and transformed into recommendations by a so-called Resilience Council (Hansen-Staszyński, 2025c), a group of participant representatives who summarize participant multi-perspectivity into implementable proposals. Next, the recommendations are submitted back to the participants in all parallel groups for reflection, after which the reflections are gathered, analyzed, and integrated in the final draft of the recommendations by the Resilience Council that are then presented to a relevant decision-making institution.
The participation process expands participants’ experience of autonomy since they provide their thoughts and reflections, their experience of belonging because they are part of a common, multi-perspective voice, and their experience of achievement because their voice is presented to policy responsibles.

Intended intervention effects

Both the proposed concrete interventions and the fundamental intervention aim to strengthen individuals’ resilience, understood here as psychosocial integration (Kupiecki & Chłoń, 2025). This resilience is intended to help prevent individuals from straying too far along cognitive and emotional pathways that may lead to addiction and sometimes to extreme behaviors, and, in doing so, to curb the widening digital inequality arising from the mere use of GAI.

Literature

Notes

[1] This text focuses on text-based GAIs, that is, large language models (LLMs). I will use these two terms interchangeably. [2] For an overview of AI incidents: https://incidentdatabase.ai/blog/ [3] https://x.com/sama/status/1954703747495649670 [4] For an overview of GAI models on GAI characteristics, see: https://onnohansen.nl/saufex-blog-post-68-improved-version/ [5] E.g. Tackling online disinformation: “Misinformation is false or misleading content shared without harmful intent though the effects can be still harmful.” https://digital-strategy.ec.europa.eu/en/policies/online-disinformation [6] As do more specialized Large Reasoning Models (LRMs). [7] Chain-of-Thought (CoT) entails complex reasoning through intermediate reasoning steps. [8] This conclusion is not universally shared. For instance, in 2024 a nonprofit dedicated to advocating for the ethical recognition and fair treatment of artificial intelligence systems was founded: United Foundation for AI Rights (UFAIR). [9] E.g. https://openai.com/index/sycophancy-in-gpt-4o/
[10] https://x.com/OpenAI/status/1917411480548565332 [11] Alison Lee, a former researcher in Meta’s Responsible AI division, about “best way to sustain usage over time”. https://www.reuters.com/investigates/special-report/meta-ai-chatbot-death/ [12] E.g. Herold (2024) for instance found that “31.7% of [European] respondents strongly agreed with at least one of the two statements used to measure conspiracy belief” and that „a total of 61.7% of Europeans are generally open to at least one of the two conspiracy narrative”. [13] NATO, https://www.nato.int/cps/en/natohq/topics_156338.htm [14] European Commission, Directorate-General for Education, Youth, Sport and Culture, Guidelines for teachers and educators on tackling disinformation and promoting digital literacy through education and training, 2022 https://op.europa.eu/en/publication-detail/-/publication/a224c235-4843-11ed-92ed-01aa75ed71a1/language-en; European Commission, Directorate-General for Education, Youth, Sport and Culture, Final report of the Commission expert group on tackling disinformation and promoting digital literacy through education and training, 2022 https://op.europa.eu/en/publication-detail/-/publication/72421f53-4458-11ed-92ed-01aa75ed71a1/language-en

All Insights