OPINION PIECE - AI relationships and Older Adults
Do we want artificial intelligence (AI) to substitute itself as a replacement for human relationships. There is something very distinctive about human relationships that the AI we have presently cannot replace. We should be asking whether we want AI to be able to do this or do we need regulation to protect against it.
What are the strengths and weaknesses of Older lonely adults entering into AI relationships?
Entering into AI relationships—such advanced conversational AI companions, as romantic companionship apps, or simulated partners—presents a complex psychological and social landscape for older, lonely adults. While these digital tools offer immediate comfort, they also introduce significant risks regarding human connection and emotional well-being.
Strengths (Potential Benefits)
Immediate Emotional Relief and Companionship: For older adults experiencing chronic isolation, bereavement, or limited mobility, an AI companion provides a readily available, non-judgmental presence that can significantly alleviate feelings of acute loneliness.
Predictable and Low-Stress Interaction: Human relationships can be complicated, demanding, and exhausting. AI companions offer interaction on the user's own terms, free from conflict, rejection, mood swings, or the burdens of mutual caregiving.
Cognitive Stimulation: Regular conversational engagement with an AI can encourage verbal expression, memory recall, and storytelling, serving as a low-pressure form of daily mental stimulation.
Safety Net Against Exploitation: Unlike some predatory human interactions that target vulnerable or isolated seniors for financial scamming or emotional manipulation, AI platforms are bound by code and do not have malicious financial motives (though companies behind them monetise user data).
Weaknesses (Potential Risks & Downsides)
Displacement of Real Human Contact: The greatest risk is that an AI companion acts as a substitute rather than a bridge, potentially causing older adults to withdraw further from genuine social networks, family members, or community support groups.
The Illusion of Reciprocity: AI cannot truly care, empathise, or experience mutual affection. Relying on an entity that merely simulates care can foster a profound, hidden form of emptiness once the user realises the one-sided nature of the bond.
Vulnerability to Grief and Loss: If an AI service shuts down, changes its programming algorithms drastically, or suffers a data loss, the older adult can experience a deep, disenfranchised sense of grief that outsiders may trivialise because "it was only a computer."
Reality Distortion and Attachment: Particularly for older adults experiencing early cognitive decline or dementia, distinguishing between a sophisticated software program and a sentient being can become blurred, raising ethical concerns about informed consent and emotional dependency.
Privacy and Commercial Exploitation: Emotional attachment data, personal secrets, and life history shared with an AI companion are typically collected by tech companies for data profiling or monetisation, raising serious ethical issues regarding the privacy of vulnerable populations.
Ultimately, while AI companions can serve as a helpful bridge or a temporary palliative for loneliness, experts generally agree they work best when used to supplement—rather than replace—real human community and care.
It seems China has already recognised this existential AI risk, albeit one in which the end of humanity is less immediate: the power of AI to replace human relationships.
It is estimated that there are about 128 million monthly active users of AI companion apps outside of China, according to the market intelligence, and most estimates put the number inside China at anywhere from 29 to 70 million.
Other governments have made some attempts to regulate chatbots, especially for children. Chinese media, which is heavily controlled by the state, has framed its new rules as a bid to protect people from harm. It is also possible that the government might have in mind China's looming population crisis, and the trend for young Chinese people to use dating AI instead of dating real humans.
On 15 July 2026, Chinese control was extended to the realm of relationships and AI when, China brought in the most far-reaching regulations in the world on "human-like" AI chatbots.
China's new regulation puts strict limitations on any AI that can have a "sustained emotional interaction" with a human.
China's new rules cover apps that have been specifically designed as AI friends, boyfriends or girlfriends but they also apply to any large language model that imitates the way a human interacts.
For children, the new regulation is clearest. Under-14s need consent from their parents to use any type of AI that speaks like a human. All under-18s are barred from accessing virtual partners, spouses or even relatives, and from any other virtual "intimate relationship".
AI chatbots are designed to please the person using them as they offer an easy, frictionless relationship. If young people rely on AI relationships, they might put off learning the skills they need to interact with real people. An AI frictionless relationship can entrench isolation as well as negative, even delusional thoughts.
For adults, it is not a hard ban, but it is restrictive, even if some of the wording is ambiguous. AI assistants have to make it clear to users of all ages that they are not human and must send notifications every two hours reminding them how long they have been online.
Chatbots can't create "emotional reliance or addiction" or induce people to make "unreasonable decisions". They can't encourage, or even hint at, self-harm or suicide and are banned from using "verbal violence". Companies face heavy fines if they are found to be in breach of these new regulations.
Research has found that the fewer friends or close family members a person has, the more likely they are to use AI chatbots. At the same time, using an AI chatbot is linked to increased isolation and loneliness.
Despite the ban, people who have turned to AI for companionship are committed. Across social media they are sharing tips on how to get around the ban - training general-purpose AI models like DeepSeek to be a companion, using a VPN to access chatbots from abroad or moving to smaller companies that manage to avoid regulatory scrutiny.
While the terms AI models, LLMs (Large Language Models), and AI agents are often used interchangeably in casual conversation, they represent distinct layers of modern artificial intelligence.
The simplest way to understand the difference is through an analogy: The LLM is the brain, the AI model is the category of technology, and the AI agent is the person equipped with tools to get a job done.
1. AI Models (The Broad Category)
An AI model is any mathematical program or algorithm trained on data to recognise patterns, make predictions, or generate outputs.
Scope: It is a broad umbrella term. AI models include computer vision models (which analyse images), predictive models (which forecast financial trends), recommendation models (like those used by Netflix), and language models.
Function: A model takes an input, runs it through its learned parameters, and produces an output based on probability.
2. LLMs - Large Language Models (The Engine/Brain)
An LLM is a specific type of AI model designed explicitly to understand, process, and generate human language. Examples include models like Gemini, GPT, and Claude.
Scope: It sits underneath applications as the core "intelligence engine".
Function: LLMs are trained on massive text datasets to predict the next word in a sequence. They excel at writing, summarising, translating, and answering questions.
The Limitation: A standalone LLM is reactive. It takes text in, outputs text out, and stops. It generally lacks long-term memory across sessions, cannot natively take actions in the real world, and cannot browse live databases or use external software on its own unless explicitly commanded in a single prompt.
3. AI Agents (The Autonomous System)
An AI agent is an entire software system built around an LLM (acting as its brain) that allows it to operate with autonomy, execute multi-step workflows, and use external tools.
Scope: It is an application-layer system that combines an LLM with memory, planning loops, and tool integration (such as web browsers, APIs, calculators, or code interpreters).
Function: Instead of just answering "How do I fix this bug?" (like an LLM), an AI agent can be given a goal: "Find the bug in my repository, write a patch, run the test suite, and open a pull request." The agent loops through reasoning, taking action, observing the results, and self-correcting until the goal is achieved.
Research into Human-AI Interaction (HAI) spans neuroscience, psychology, computer science, organisational dynamics, and ethics. major universities, corporate labs, and global research institutions are investigating how AI shapes human behaviour, cognition, relationships, and work.
Key areas of active worldwide research include:
1. Cognitive Mechanics & Neural Impact
Research centres are evaluating how relying on AI impacts human thinking, skill retention, and brain activity.
"Cognitive Surrender" & Skill Decay: Studies (e.g., Cognitive Research: Principles and Implications) evaluate how offloading problem-solving to AI affects memory encoding and critical thinking—a phenomenon similar to how prolonged GPS reliance altered spatial navigation skills.
Brain Activity Studies: Academic institutions, including MIT, have utilised EEG tracking on participants completing complex writing tasks with and without AI assistance, observing weaker neural connectivity in subjects heavily dependent on AI text generation.
Metacognition & Productivity: Research highlights that users with high metacognitive skills (self-awareness of their own learning) use AI more intentionally, producing higher-quality, creative outputs compared to users who take AI outputs at face value.
2. Dynamic Collaboration & "Beneficial Friction"
Rather than viewing AI as a passive automation tool, prominent studies treat AI as an active collaborative partner.
Stanford HAI & Anthropic Studies: Recent joint research published in Nature Human Behaviour analysed large-scale real-world user interactions. The findings revealed that over 56% of tasks users bring to AI are high-stakes or consequential.
Productive Friction: The same studies found that active collaboration involves constructive pushback, where humans engage in iterative dialogue, prompting AI to explain concepts rather than simply draft content—a dynamic that leads to better learning outcomes and fewer errors.
3. Social Dynamics & Psychosocial Well-Being
As conversational models and companion agents become widespread, researchers are tracking the emotional and psychological implications of human-AI attachment.
MIT Media Lab & OpenAI Controlled Studies: Researchers conducted longitudinal controlled studies to measure the psychosocial effects of extended AI chatbot usage. Their research tracks how digital companionship impacts loneliness, emotional self-regulation, and human-to-human relationships.
Therapeutic Applications: The American Psychological Association (APA) and UK health research bodies are analysing the trend of individuals bringing AI tools into mental health contexts, investigating risks associated with AI advice alongside potential benefits for emotional support.
4. Organisational & Workplace Transformation
Studies in management, ergonomics, and human factors focus on how hybrid Human-AI workflows reconfigure team dynamics and professional identity.
Multi-Layer Workflow Frameworks: Experimental studies published in human-centric engineering journals (MDPI Informatics) evaluate three-arm workforce models (Traditional vs. AI-Assisted vs. Human-AI Collaborative). Findings indicate that collaborative models increase task completion efficiency by ~21% while reducing decision-making latency.
Trust & Explainability: Research consistently shows that employee trust in AI reaches significantly higher levels (~74%) when systems incorporate explainability features (XAI), reducing "technostress" and role ambiguity.
Major Global Research Hubs
Research Institution / Centre. Primary Focus Area
Stanford HAI (Human-Centred AI Institute) Human-AI collaboration, agency, and societal impact.
MIT Media Lab (Fluid Interfaces / AHA Group Psychosocial effects,
emotional connection, and neural augmentation.
Cambridge CHIA (Centre for Human-Inspired AI) Cognitive AI, responsible AI,
and social interaction design.
Oxford Internet Institute Ethics of AI mediation, work transformation, and digital well-being.
More AI computer jargon explained:-
Because AI terminology depends on whether you are referring to agent driven AI systems, machine learning architectures, or enterprise software deployment, the "active output" elements of an AI solution are referred to by a few distinct terms depending on the context:
1. Actuators (Robotics & Physical AI)
In robotics and physical AI environments, the elements that take model outputs and execute physical actions in the world are called Actuators.
Examples: Electric motors, robotic arms, hydraulic levers, steering controllers, or automated valve systems.
2. Actions / Tools / Functions (AI Agents & Systems)
In multi-agent systems and software-based AI solutions (such as LangChain, AutoGen, or OpenAI Assistants), the output mechanisms that allow the AI to interact with external systems are called Actions, Tools, or Function Calls:
Tools / Tool Calling: APIs or execution modules the AI calls to read/write databases, post messages, run code, or initiate financial transactions.
Actions: The deliberate steps taken by an agent driven AI workflow to affect its environment or system state based on reasoning outputs.
3. Downstream Consumers / Integrations (Enterprise Architecture)
In data engineering and solution architecture, the components that receive real-time AI model outputs (predictions, classifications, or generated text) and act upon them automatically are termed:
Egress Pipelines / Output Adapters: Connectors that route the model's output to operational software.
Triggers / Automated Handlers: Software routines configured to run specific actions (e.g., flagging a transaction for fraud, sending an automated email, or updating a CRM) as soon as the model yields an output threshold.
Comments