Humans are, at their core, sense-making creatures that strive to understand the world¹. We detect patterns, test hypotheses and continuously refine our understanding of reality².

Curiosity is a strong psychological motivator^3,4 and novel information activates the same ancient neural pathways that process primary rewards such as food, sex, and shelter^5,6.

The internet, with its vast repositories of knowledge, offers access to information and expertise that would have been unimaginable to previous generations. Search engines, social media, and increasingly generative AI systems have become powerful intermediaries between us and effectively infinite knowledge and opportunity. Although not designed for this purpose, social media has evolved into a primary information source for many, especially younger users⁷.

Generative AI reveals a similar pattern, as users use it primarily to ask for synthesised how-to information⁸.

AI overviews introduced in 2024 have further changed traditional web search by displaying an AI-generated summary on top of traditionally optimised search results. An Ahrefs analysis of 146 million search results found that AI overviews appear for 21% of all queries — rising to 58% of questions and around 44% of health and science topics⁹. Half of the time an AI overview appears, users are generally satisfied by the summary and do not open other search results, even when its primary sources are Wikipedia, Youtube and Reddit¹⁰.

Humans are profoundly social animals. People with a stronger connection with family, friends and community live longer, healthier, happier lives¹⁸.

Conversely, social isolation has measurable biological effects: brain responses to loneliness resemble those triggered by starving¹⁹, and chronic loneliness carries health risks comparable to smoking 15 cigarettes a day²⁰.

Digital technologies have dramatically expanded the potential scale of human connection. Social media platforms enable people to maintain relationships across geographic distances, discover communities organised around shared interests, and build professional networks that were previously difficult to access. Of the roughly 73% of the global population that has access to the internet, approximately 94% is on social media²¹.

More recently, conversational AI systems have introduced new forms of mediated interaction. Generative AI — and AI companions specifically — allow users to engage in dialogue that can feel socially meaningful, and many users report experiencing emotional comfort or reduced loneliness when interacting with these systems²².

Humans are not merely sense-making creatures — they are generative ones. Beyond understanding the world, we have a profound drive to add to it: to produce knowledge and artefacts that outlasts the individual act of thinking, and are at the core of all human cultural, scientific, and economic advancements²⁹.

Generative AI has become a routine presence across this range of knowledge work at a historically unprecedented pace. A survey found that by late 2024, nearly 40% of workers reported using generative AI for work, with 9% doing so every workday — a diffusion rate comparable to the personal computer and faster than the internet³⁰.

This adoption spans the full range of cognitive tasks. In everyday professional contexts, AI is now integrated into writing, analysis, and problem-solving, sometimes decreasing completion time up to 40%, and increasing output quality by 18%^31,32,33.

In scientific research, generative AI integration has accelerated across all research stages — from literature review and hypotheses development to data analysis and manuscript drafting^34,35, with some tools now positioning themselves as assistants or autonomous collaborators^36,37.

In the creative and artistic domain, adoption is equally pervasive: 87% of music producers report using AI-powered tools as part of their creative process, while surveys of visual artists and creative professionals document comparable widespread use for ideation and production³⁸.

The ability to regulate emotions and cognitive states is fundamental to human wellbeing. Coping strategies allow individuals to manage stress, uncertainty, and emotional distress, and the ways people regulate these states are critical determinants of psychological resilience⁶⁹.

In contemporary environments, digital technologies increasingly serve as tools through which people manage emotional and cognitive demands. Social media has become a significant source of peer support for those facing mental health challenges. Among young adults with mental illness, 68% report using social media to share personal experiences, 66% to connect with others experiencing similar conditions, and half use it to learn coping strategies⁷⁰. The perceived anonymity of online spaces and the reduced stigma associated with digital interaction can make it easier for individuals to seek support that they might avoid in face-to-face contexts.

Public narratives shared online can also contribute to mental health resilience. Research demonstrates a “Papageno effect”, in which social media posts that highlight hope, recovery, and coping strategies — particularly when shared by individuals with lived experience — reduce suicidal ideation among viewers and increase intentions to seek help⁷¹.

Generative AI systems are also emerging as coping tools amid widespread unmet mental health needs. Nearly half of young adults in Europe and the US report difficulties accessing mental health care^72,73. In response, one in eight US and one in four UK young adults now use AI chatbots for mental health advice when experiencing distress^74,75. Users are drawn by accessibility — no appointments, no cost, and constant availability — as well as the privacy of engaging from home.

But beyond convenience, AI chatbots used as companions can offer what psychologists call “permission to feel”: a non-judgemental empathetic space to process emotions that many people lacked growing up⁷⁶. Their constant availability allows AI to fill that gap at scale, stepping in where human support systems have fallen short.

Communication is the mental capacity through which humans share mental states and coordinate meaning^46,47.

Humans devote 30-40% of speech solely to self-expression, an act the brain rewards as powerfully as food or sex⁴⁸. But communication is also reciprocal: meaning is not transmitted but negotiated, built through the back-and-forth of expression and interpretation⁴⁹. Together with comprehension, this exchange constructs the shared reality on which social and democratic life depends.

Social media has radically democratised societal capacity to communicate at scale. Where broadcasting once required institutional gatekeepers, platforms now allow anyone to potentially reach a global audience. A factory worker can become the most-followed person on TikTok⁵⁰, while an unknown teenager can amass 150 million followers almost overnight⁵¹.

But posting on social media is not just human expression anymore. Entertaining others, sharing opinions, and “influencing” now constitute a $200 billion/year industry sustained by 200 million creators, with nearly 60% Gen Zers aspiring to join^52,53,54.

At the same time, generative AI is emerging as a powerful mediator in everyday communication. Two-thirds of writing tasks involve editing communication, and interpreting the meaning of information accounts for 13% of all ChatGPT usage⁵⁵, suggesting generative AI is entering communication at both ends.

For non-native speakers and those with lower writing confidence, this mediation can be genuinely enabling. Access to AI assistance has been shown to improve the speed of written output and to significantly raise English fluency among international workers, reducing barriers to participation in professional and public discourse⁵⁶.

Autonomy is the cross-cutting condition: any of the five needs can be fulfilled to varying degrees of it, along a spectrum that runs from genuine human exercise to AI substitution.

Psychological research bears this out: needs fulfilled through externally controlled or substituted means do not produce the same wellbeing, identity coherence, or intrinsic motivation as needs fulfilled through autonomous engagement^84,85.

And fulfilling these needs draws on cognitive functions — attention, memory, critical thinking, decision-making, emotional regulation, social cognition, and intrinsic motivation among others — that are themselves subject to influence, capture, and degradation.

Like physical capacities, cognitive ones are maintained through use: what is systematically bypassed gradually atrophies⁸⁶.

Technologies that allow an effortless comprehension of the world also shift epistemic agency: instead of people actively interrogating and searching for information, platform algorithms increasingly determine what is relevant, in what order, and with what framing.

Experimental work suggests that the ordering of search results can shift attitudes and preferences — including voting preferences — without users’ awareness of the manipulation¹¹.

At the same time, ubiquitous access encourages cognitive offloading: when people expect information to remain externally available, they tend to remember less of the content and more of where to retrieve it, which can weaken internal understanding and long-term memory¹².

Finally, research documents that hallucinations in AI overviews lead to passive acceptance, precisely because the fluency of the generated text functions as a false credibility cue, undermining critical thinking^13,14,15. In high-stakes domains such as mental and physical health this becomes a safety issue already documented^16,17.

While digital platforms expand opportunities for connection, they also reshape how social interaction occurs. Instead of individuals actively reaching out to others and sustaining reciprocal relationships, digital environments are increasingly structured through algorithmic systems that match users to content, audiences, and viewpoints.

Large-scale research on Facebook demonstrates that exposure to cross-cutting viewpoints is constrained not only by users’ existing social networks, but also by how platform algorithms filter and rank content²³. Even when diverse perspectives exist within a network, algorithmic curation can limit their visibility, stabilising or exacerbating echo chambers and contributing to political and affective polarisation over time^24,25.

At the same time, agency in participation is unevenly distributed. Algorithmic ranking systems systematically privilege already-prominent accounts, amplifying the visibility of highly followed users while reducing the reach of ordinary participants.

As a result, most users no longer post at all²⁶: estimates suggest that roughly 75–90% of people on social media are “lurkers”, passively consuming content produced by a small minority²⁷, closely approximating the long-observed “1% rule” of online participation²⁸. Social connection thus shifts from mutual exchange to asymmetric consumption.

Generative AI can augment creation in terms of speed and volume, but producing more is not the same as thinking more. Where people once originated ideas and constructed arguments through the effort of making, they now select among readily available AI outputs.

Critically, generating and selecting are distinct cognitive processes, but only the first builds capacity³⁹. In an experiment on creative writing, access to AI-generated ideas improved individual outcomes, as stories were rated as more creative, better written and more enjoyable. At the same time, output across individuals became significantly less varied, as writers anchored on model suggestions even when free to depart from them⁴⁰.

A similar pattern is appearing in scientific research: a large-scale analysis of 41 million papers across six disciplines found that while AI adoption consistently benefitted individual scientists through more citations and publications, collectively, it reduced the diversity of scientific topics explored⁴¹.

The same shift is being documented across creative domains: a systematic review found that AI integration is moving artists to “meta-creation”, where original creation is displaced by prompting, selecting, and refining model output⁴².

Over time, this pattern of reliance risks something more fundamental than homogenesation. Research shows that people tend to overtrust systems that are usually right, progressively reducing independent effort and the practice through which skills are maintained^43,44.

In learning contexts, evidence is mixed and still emerging, but recent syntheses of experimental studies highlight that benefits depend heavily on how the tool is integrated, and that unstructured use can promote shallower understanding⁴⁵.

While research on these phenomena remains nascent, the cumulative result across cognitive functions and time would be cognitive erosion: the gradual atrophy of cognitive functions driven by an environment that systematically removes occasions for exercising it.

While digital tools can support coping, they can also reshape how emotional regulation occurs by shifting agency from effortful self-regulation toward frictionless accommodation.

Instead of engaging in active strategies such as reflection, problem-solving, therapy, sleep, or reciprocal support, individuals may be nudged into passive coping patterns, including continuous scrolling, emotional numbing, or remaining inside calibrated information feeds.

Research on “doomscrolling” illustrates this dynamic. Compulsive consumption of negative news has been linked to worse psychological outcomes, including increased existential anxiety and pessimism about human nature, suggesting that digital environments can sustain distress even when they feel temporarily soothing⁷⁷.

Evidence also indicates that the structure of online environments can influence well-being more broadly. In a nationally representative study of U.S. young adults, heavier social media use was associated with higher levels of perceived social isolation⁷⁸. Experimental studies further suggest that reducing social media use can improve well-being outcomes including loneliness and depression⁷⁹.

Generative AI compounds this dynamic in ways that are structurally distinct from passive scrolling. A growing number of young people are using AI not only as a support tool but to avoid the discomfort of difficult human interactions altogether. They ask AI chatbots to draft rejection messages and script hard conversations, outsourcing emotional labour that would otherwise require developing interpersonal capacity⁸⁰.

Similarly, AI systems designed to be non-judgemental, constantly available, and responsive make them effective at providing comfort and validation, but these qualities make them unlikely to provide the friction through which emotional development occurs⁸¹.

Healthy human relationships hold us accountable: they challenge, confront, and require change. Systems that are always available and over-accommodating are not care^82,83.

Rather than self-expression at scale, digital communication is becoming a process of adapting to what the algorithm rewards.

Engagement-optimised ranking turns communication into a competitive market for reactions, and an audit of Twitter’s (now X) engagement-based ranking found it amplifies divisive content more than alternatives aligned with what users say they want to see⁵⁷. This creates a structural incentive to communicate in ways that travel — often through emotionally charged simplifications — shifting agency from what users intend to express toward what spreads.

The downstream epistemic cost is well documented: false news diffuse farther and faster than true news, meaning that amplification dynamics systematically advantage misinformation⁵⁸.

Generative AI compounds this mediation in two ways:

First, it dramatically lowers the cost of producing large volumes of text, enabling automated or semi-automated communication at a scale that would previously have required substantial human effort. This raises the background noise of public discourse and increases the circulation of synthetic or low-effort content, often referred to as “AI slop”^59,60,61,62.

Second, it can subtly shape the form of human expression itself. At the level of language and culture, AI-assisted communication tends to flatten expression toward dominant — predominantly Western and English — rhetorical conversions, diminishing individual and cultural diversity^63,64.

At the individual level, articulating thoughts in writing is not merely a communicative act, but the process through which people clarify what they think^65,66. Delegating that formulation to AI removes a feedback loop between expression and self-knowledge. And when the recipient senses the substitution — however unconsciously — trust erodes: evidence shows that people evaluate others more negatively when suspecting algorithmic authorship⁶⁷, especially when the voice being replaced belongs to someone in a position of trust⁶⁸.

The displacement follows a consistent pattern across all domains: Comprehension is replaced by delivered answers; connection by simulated intimacy; expression by what the algorithm rewards; creation by selection among model-generated options; emotional regulation by frictionless accommodation.

In each case the need appears met — engagement metrics confirm it — but the cognitive functions through which autonomous fulfillment would occur are bypassed rather than exercised.