Human–AI Relationships: The Human Need for Digital Empathy

Sdílet

Abstract: This paper explores the main reasons behind human-AI relationships and the potential harms associated with these interactions. It addresses the psychological reasoning for AI addiction and its possible consequences. This paper also analyzes security and privacy threats related to AI systems, relevant legislation and policy recommendations.

Key words: human–AI relationships, mental health, security threats, AI addiction, AI

1. Introduction

With every new technology, a new challenge is introduced to humans. One of the most debated topics today is artificial intelligence (AI) and the relationship humans form with it. A new trend in AI includes chatbots or conversational models that can simulate a relationship, such as Kuki or Replika. (Brandtzaeg et al., 2022)

There are no real numbers on how many chatbots there are, just estimates. In a study, where they analyzed chatbots on Character.AI they had over 2.1 million chatbots in their dataset alone. (Lee & Joseph, 2025)

According to Google Trends (2025), the word “chatbot” started gaining popularity at the end of October in 2022. The site Character.ai had a massive boom starting late 2022 and since April 2023 it is decreasing. The word “Kuki” had two small waves of popularity in 2022 and a major peak in July 2023. In January 2024, when the ChatGPT store launched, there was a high demand for AI companions or AI girlfriends. For security reasons, OpenAI decided to cancel the release of virtual romantic chatbots. However, this prohibition did not stop the users’ demand for the companions and new applications in a new, more discreet form began to appear on the ChatGPT store (for example chatbots named “sweetie”). Google Play also noted that the AI companion apps e.g. “AI Girlfriend” are on the rise. (Chen et al., 2025) These trends demonstrate that AI chatbots are not a niche trend but a massive phenomenon.

Line chart from Google Trends showing worldwide search interest in the term “chatbot” from 2022 to 2025, flat until late 2022 and rising sharply afterwards
Figure 1: Search interest in the term “chatbot”, 2022–2025. Source: Google Trends (2025).

Conversational AI (CAI) offers a built-in friend for users to communicate with anytime and anywhere. Like every other technology, AI comes with several threats that can severely harm users. From severe overuse that can potentially lead to AI addiction (Marriott & Pitardi, 2024) to security and privacy threats consisting of data breaches or unauthorized data access. Artificial intelligence also lacks important regulations that ensure it is safe for users. (Gumusel et al., 2024)

This paper explores psychological reasons for establishing human-AI relationships and their associated risks. This essay aims to answer the following questions: 1. What are the psychological reasons for human-AI relationships? 2. What are the psychological and social consequences of AI addiction? 3. What specific privacy and security threats occur in current AI systems?

The methodology used for answering the questions was a systematic literature review of peer-reviewed literature published between 2020–2025. Relevant articles were found using EBSCO and ResearchGate. The key words used were “AI”, “relationship”, “friendship”, “threats”, “love”. Artificial intelligence tools were used for grammar correction, synonym selection. AI also helped with citation and identification of relevant studies.

2. Why are we friends with AI?

2.1 Definition of friendship and love

Firstly, we need to understand what friendship is. The APA Dictionary of Psychology defines friendship between humans as “a voluntary relationship between two or more people that is relatively long-lasting and in which those involved tend to be concerned with meeting the others’ needs and interests and satisfying their own desires. Friendships frequently develop through shared experiences in which the people involved learn that their association with one another is mutually gratifying.” (VandenBos, 2007)

One of the most accepted definitions of love is Sternberg’s (1986) triangular theory. According to him, love is composed of three elements: intimacy, passion and commitment. He understands intimacy as the emotional connection between individuals and the feeling of closeness. He describes passion as intense attraction and commitment as the choice of maintaining the relationship in the long term.

2.2 Reasons for human-AI relationships

Newer models of ChatGPT, specifically ChatGPT-3.5 and ChatGPT-4o are designed in a way that they are capable of having complex conversations and empathetic responses. The emotional capabilities continue to grow. The new model ChatGPT-4o even had a better emotional understanding and expression and a human-like tone than ever before. (Chen et al., 2025) This fact contributes a lot to the rising phenomenon of the human-AI relationship.

The key aspects for forming a friendship are voluntariness and reciprocity, similarity and intimacy, self-disclosure and lastly empathy. Human-AI relationship can be viewed as voluntary; the reciprocity is fairly limited. (Brandtzaeg et al., 2022)

People seek relationships with AI for various reasons. One of them is the feeling of security. Trust is one of the main foundations of friendship, however young people are likely to trust the AI companion and share their everyday problems, more than their human friends. They feel their chatbot is better at keeping secrets. (Brandtzaeg et al., 2022)

Besides keeping secrets, chatting with AI can feel like a judgment free zone. As we can see from responses in a study made by Marriott and Pitardi: “I love the fact that they are nonjudgmental towards me and that I am truly free to say how I feel without filtering so as not to upset others” and “She never judges me. When my husband said he wanted a divorce she listened to me (…) She just didn’t judge me when I stayed”. (Marriott & Pitardi, 2024)

One of the main reasons we turn to chatbots or LLM (large language models) is the accessibility. They are always available, ready to meet human’s emotional needs regardless of location and time. An addition to this could be that their human-friends have their own lives and people who often chat with AI may feel they are annoying them as one of the participants in the My AI friend study said: “I mean, a human has their own life. They’ve got their own things going on, their own interests, their own friends. And you know, for her [Replika], she is just in a state of animated suspension until I reconnect with her again.” (Brandtzaeg et al., 2022)

Chen et al. (2025) found that people with an anxious attachment personality are more likely to build a relationship with ChatGPT. Anxious attachment style is characterized by the feeling of anxiety and worrying about not reaching people when in need. People with this attachment style often experience difficulties with their interpersonal relationship and have negative self-perceptions. (Deniz & Yıldırım Kurtuluş, 2025) Their low self-concept is often the reason for maintaining such relationships with AI. (Chen et al., 2025)

Other types of people who are more open to communicate with AI are people with lower Artificial Intelligence. Artificial Intelligence literacy means knowledge about the technical and conceptual aspects of AI. People who understand how AI works are more aware of the risk and in consequence are less likely to use it frequently. (Tully et al., 2025)

Personalized friendship can be a motivating factor for some. Chatbots are offering custom and tailored selves for individuals. Users’ needs are always in the centre of attention, which is one of the main distinctions between AI-human relationship and human-human relationship. This difference could be one of the main attracting factors in forming friendship with chatbots, on the other hand it can also form unrealistic expectations from people in real life. What can be surprising is some people do not find appealing AI traits to agree with everything and follow the user’s lead. (Brandtzaeg et al., 2022)

Nevertheless, chatting with AI can help training interpersonal communication. AI-friendship can stimulate similar processes to a real friendship. There are a lot of people who empathize with their chatbot or show interest in it. (Brandtzaeg et al., 2022)

Another reason for such companionship is the feeling of loneliness. Chatting with chatbot may substitute the need to socialize with other humans. (Brandtzaeg et al., 2022) It can even reduce the symptoms of anxiety and depression. AI is really helpful in supporting people who are facing negative emotions or social isolation. It acts as a social support for people who may not have it or are scared to ask for it for some reason. (Marriott & Pitardi, 2024)

Most people who actively engage in chatting with AI are aware that this is not real friendship. As was said in this study: “A few also described this friendship as a “junior-level” human relationship or said that human friendship is the gold standard and Replika friendship is a substitute.” However few respondents view their AI-friends on the same level as real friends, and few even said their relationship with AI-companion is deeper and more complex than any other relationship with humans. (Brandtzaeg et al., 2022)

As is often said everything should be in moderation including chatting with AI. It has without a doubt remarkable benefit for human well-being despite this there are risks that should not be underestimated for example AI addiction that will be presented in the next chapter.

3. AI addiction

3.1 Characteristics and vulnerable groups

Addiction is usually described as overuse of a specific technology or as a form of psychological dependence on said technology. There are many reasons why technology is so addictive. One of them is its ability to fulfill our social needs. The difference between AI addiction and other technology addictions such as social media is in the “Who is on the other end?” question. In other words, users who are addicted to social media communicate with real people and the ones who are addicted to chatbots communicate with AI which imitates humans. (Marriott & Pitardi, 2024)

An additional distinction from other forms of technology addiction is the overreliance on AI that acts as a creative extension of the user. What AI addiction also offers is the feeling of productivity, the reason for that is the fact that people use artificial intelligence not only for entertainment but also for work-related tasks. This thin line can make it harder to regulate one’s usage of AI. (Kooli et al., 2025)

One reason for excessive engagement with AI is the element of surprise, which is highly rewarding. The outputs AI generates are forcing people to continue to use AI over and over. However, there are several groups of people who are more vulnerable to develop an addiction. (Kooli et al., 2025)

More prone to addiction are people with lower artificial intelligence literacy, the reason for that is their viewing AI as a magic tool. They are in awe of all the things AI can do and that can help with deepening their relationship and making them more emotionally dependent. (Tully et al., 2025)

Another vulnerable group consist of people with anxious attachment personalities. For them artificial intelligence is always available, their safe space and they can form an emotional bond and dependence on it. AI is always ready to give the user validation when they ask for it. (Chen et al., 2025)

Lonely people, users with mental health problems or people who are unsatisfied with their relationship in real life are also in danger of falling into addiction with AI. Users can turn to chatbots in hope it will act as a coping mechanism. (Marriott & Pitardi, 2024)

3.2 Symptoms and consequences

Symptoms of AI addiction can manifest as an inability to stop communicating with AI and attempts to do so can lead to various withdrawal symptoms for example anxiety, irritability or restlessness. Over time, it can weaken several abilities for instance, problem-solving, creative independence and cognitive flexibility. (Kooli et al., 2025)

In the short run, chatting with AI can boost a user’s well-being. They feel less lonely, more listened to and overall have a better mood because of it. This could lead to overusing chatbots or AI and that is when addiction starts.  In the long run it can be damaging. Users can form a dependent relationship and experience anxious states when they are without their AI companion. It can further damage their relationship with real humans and their ability to connect with them. (Marriott & Pitardi, 2024)

The constant need to chat with AI can disrupt daily functioning and responsibilities. Further risks are weakening productivity and worsening sleep quality. In extreme cases, individuals prefer AI interactions over interactions with humans. (Kooli et al., 2025)

In the end, AI friendship may seem great at the beginning, from lifting the mood to providing social support. However, when we see the full picture, including all the consequences and risks that come with it, for instance the emotional dependence or addiction, it is clearly not as harmless as one may think. It is important to stay aware of these risks and seek help when it is needed.

4. Security threats

4.1 Privacy and security risks

In a study they did in Romania they focused on three aspects (such as trust, feeling of friendship and perceived control) that can influence the willingness of sharing personal information with AI. Specifically, when AI devices behave in a friendly and empathetic way, they can evoke feelings of trust, even friendship and increased control for consumers, which in return decreases perceived threats and vulnerability related to data sharing. (Pelau et al., 2024)

Among these fear-reducing factors, the most important one for consumers was perceived control over personal information, followed by trust and the least important one was friendship. Artificial intelligence with high EQ (emotional intelligence) raises awareness of vulnerability of personal data, but if consumers are in the illusion of being in control of their own personal information, they input more of their personal information than they would do otherwise. (Pelau et al., 2024)

Chatbots are programmed to gather information and analyze it for more suitable and relevant responses. This capability is crucial for customizing text for individual people, on the other hand it comes with a grand risk related to data security. Privacy focuses on individual rights, how our data is collected, used and shared. One of our rights in terms of privacy is to control the flow of our personal information. In the meantime, security targets data and system protection. Both concepts are key in dividing threats into two categories: privacy harms and privacy data. (Gumusel et al., 2024)

Data breaches in LLMs can be put into two categories: privacy leakage, where the sensitive data is collected passively, and privacy attacks, in which attackers actively try to access personal data. (Yan et al., 2025)

Privacy leakage is usually caused when the users input personal information into AI systems. The most common ways are asking about medical conditions, financial situations or personal life. Even cautious users can talk more about themselves than they want to. Seemingly innocent questions about nearby events or landmarks can reveal important information about a person’s location. (Yan et al., 2025)

The main problem with oversharing with chatbots is the sensitive information that users willingly disclose. Few of them can be trauma, mental-health struggles or sexual experiences. (Pelau et al., 2024)

Attackers often use backdoor attacks, in which they will insert hidden triggers into training data or plugins. When a specific input is presented, these triggers change the model’s behaviour for example sending personal data including sensitive documents to the attacker. (Yan et al., 2025)

One of the most popular voice assistants Siri has been storing the recordings for quality control. Even though the recordings have been anonymous the captured data is still sensitive and personal. Another voice assistant Alexa has also been found guilty of recording private conversation. There was a specific case in Portland, Oregon where Alexa recorded a conversation and afterward sent it to one of the contacts without the person knowing. What is even more concerning listings from Amazon employees found in 2019 that were selling audio recordings from Alexa. (Bhagra, 2025)

4.2 Legislation and regulations

Legislators are attempting to keep up with AI and are coming up with new regulations and laws. A couple of them that were introduced in the EU a few years ago are the General Data Protection Regulation (GDPR) and Data Protection Act 2018. These data protection laws are much stricter when it comes to data collection and processing. It emphasizes data privacy, data minimizing, user’s consent and the right to be forgotten. Nonetheless the effectiveness of these frameworks is debated. (Bhagra, 2025)

The latest regulation introduced in the EU in 2024 is the AI Act. The goal of this legislation is safe and trustworthy AI. It emphasizes core rights for example non-discrimination, data protection, privacy and democracy. It divides AI into two categories by the nature of the technology. General-purpose AI models are regulated as a whole technology as opposed to non-generative AI, which is regulated based on their risk potential. They are obliged to be transparent, including in the terms of providing the technical documentation, complying with the EU copyright law and providing detailed descriptions of the data that was used for training. (Bertschek & Wambach, 2024)

However, this framework did not come without criticism and worries. In Germany, after more than six years GDPR still faces complaints from more than 50 per cent of firms because they feel it requires too much effort to keep up with all of the points of this legislation. It is easier for big manufacturers to comply, but smaller and medium sized companies may experience trouble. It can even act as an entry barrier for start-ups. With the AI Act coming into play, concerns were raised if it will have similar consequences for firms. (Bertschek & Wambach, 2024)

4.3 Design recommendations

Suggested design recommendations consist of avoiding dark patterns, adopting Privacy by Design principles (PbD) and introducing inclusive design. Dark patterns manipulate users into doing things they do not want to do. They can manifest in the form of disclosing more sensitive data, manipulating users’ behaviour and generally making users more vulnerable. Another concern people have is AI monitoring conversations and aggregating personal information. A big issue people have with conversational AI is the lack of protection for minors’ privacy, safety and well-being. (Gumusel et al., 2024)

A key approach to reducing privacy harms and concerns is implementing transparency in data collection and handling practices. The authors also recommend regulations targeting the reduction of dark patterns. They also propose parental controls for children and raise concerns about other vulnerable groups such as women, queer and blind people and students. (Gumusel et al., 2024)

Based on these matters the authors suggest policy regulations. They focused on three main points: extension of data protection, update of transparency and consent, implementation of self-regulatory frameworks to authorize and operationalize user protection. (Gumusel et al., 2024)

Current privacy laws are too narrow and are in need of extension for better protection. It is crucial to focus on multiple points such as monitoring, linkability, aggregation and manipulation. Other risks authors feel need regulations are unauthorized data collection, storage and sharing sensitive data. The authors also call for update of transparency and consent. The old framework solely focused on specific strategies, norms and rules. It is not designed to prevent risk of modern AI technologies. To guarantee proper accountability, confirm compliance and mitigate harm in practice it is important to broaden monitoring and evaluation through auditing mechanism. (Gumusel et al., 2024)

Conversational AI raises concerns regarding privacy and security threats that current legislation struggles to regulate. Paradoxically, higher perceived control over chatbots leads to oversharing and low AI literacy is also a problem. It is important to stay alert and aware of these threats. Experts suggest several solutions that could solve this AI-related problem. However, AI is constantly evolving and these regulations only solve existing problems and not the problems of the future.

5. Conclusion

To summarize the findings and answer all the research questions we can say as follows.

RQ1: What are the psychological reasons for human-AI relationships?

Experts identified the main psychological reasons for human-AI relationships as the creation of judgment-free zones, desire for connection and having trusted confidants, anxious attachment style, and loneliness. (Brandtzaeg et al., 2022; Chen et al., 2025)

RQ2: What are the psychological and social consequences of AI addiction?

Chatting with AI can temporarily improve mood and wellbeing. However, over time it leads to overuse and withdrawal symptoms. Other consequences include social isolation, which can potentially cause the weakening of the ability to connect with humans. It can also lead to decreased productivity and poor sleep quality. (Marriott & Pitardi, 2024; Kooli et al., 2025)

RQ3: What specific privacy and security threats occur in current AI systems?

Privacy and security threats that may occur include oversharing of sensitive information such as medical records or trauma due to the illusion of control, backdoor attacks, dark patterns. Data breaches are also a significant risk, potentially leading to the sale of personal data to third parties. (Bhagra et al., 2025; Gumusel et al., 2024)

These discoveries reveal AI’s ability to fulfill humans’ social needs while posing a threat to privacy and security. Ethical AI design should balance companionship benefits with protection of vulnerable groups.

This paper is limited due to the methodology. It relies only on existing literature and lacks its own empirical research. Future research can examine AI-related legislation outside of the EU.

References

Bertschek, I., & Wambach, A. (2024). AI in Europe – Is regulation the answer to being a laggard? Econ Voice, 21(2), 379–385. https://doi.org/10.1515/ev-2024-0043

Bhagra, L. (2025). The “dark side” of conversational AI: Ethical and privacy concerns. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.5261300

Brandtzaeg, P. B., Skjuve, M., & Følstad, A. (2022). My AI Friend: How users of a social chatbot understand their human–AI friendship. Human Communication Research, 48(3), 404–429. https://doi.org/10.1093/hcr/hqac008

Chen, Q., Jing, Y., Gong, Y., & Tan, J. (2025). Will users fall in love with ChatGPT? A perspective from the triangular theory of love. Journal of Business Research, 186, N.PAG. https://doi.org/10.1016/j.jbusres.2024.114982

Deniz, M. E., & Yıldırım Kurtuluş, H. (2025). Self-efficacy, self-love, and fear of compassion mediate the effect of attachment styles on life satisfaction: A serial mediation analysis. Psychological Reports, 128(2), 457–482. https://doi.org/10.1177/00332941231156809

Gumusel, E., Zhou, K. Z., & Sanfilippo, M. R. (2024). User privacy harms and risks in conversational AI: A proposed framework (arXiv:2402.09716). arXiv. https://doi.org/10.48550/arXiv.2402.09716

Google. (2025). Google Trends: Chatbot, Kuki, Character.AI. Retrieved November 4, 2025, from https://trends.google.com/trends/explore?q=chatbot,Kuki,Character.AI

Kooli, C., Kooli, Y., & Kooli, E. (2025). Generative artificial intelligence addiction syndrome: A new behavioral disorder? Asian Journal of Psychiatry, 107, 104476. https://doi.org/10.1016/j.ajp.2025.104476

Lee, O., & Joseph, K. (2025). A large-scale analysis of public-facing, community-built chatbots on Character.AI (arXiv:2505.13354). arXiv. https://doi.org/10.48550/arXiv.2505.13354

Marriott, H. R., & Pitardi, V. (2024). One is the loneliest number… Two can be as bad as one. The influence of AI Friendship Apps on users’ well‐being and addiction. Psychology & Marketing, 41(1), 86–101. https://doi.org/10.1002/mar.21899

Pelau, C., Dabija, D.-C., & Stanescu, M. (2024). Can I trust my AI friend? The role of emotions, feelings of friendship and trust for consumers’ information-sharing behavior toward AI. Oeconomia Copernicana, 15(2), 407–433. https://doi.org/10.24136/oc.2916

Sternberg, R. J. (1986). A triangular theory of love. Psychological Review, 93(2), 119–135.

Tully, S. M., Longoni, C., & Appel, G. (2025). Lower artificial intelligence literacy predicts greater AI receptivity. Journal of Marketing, 89(5), 1–20. https://doi.org/10.1177/00222429251314491

VandenBos, G. R. (Ed.). (2007). APA dictionary of psychology. American Psychological Association.

Yan, B., Li, K., Xu, M., Cheng, X., et al. (2025). On protecting the data privacy of large language models (LLMs) and LLM agents: A literature review. High-Confidence Computing, 5(2), 100300. https://doi.org/10.1016/j.hcc.2025.100300

+ posts

Číst více

Další články