When Love Runs Out of Words: The Promise and Limits of Emotional AI

Illustrative fictional story. John and Ann

John had brought the wrong charger.

Of all the things he could not fix that evening, this was the one he kept apologising for. Ann sat beside him outside the consultation room of a Yerevan hospital, turning her phone over in her hands. After a sudden admission and several tests, they were waiting to speak to the doctor again.

“What did she mean by more tests?” Ann asked.

“We have to wait.”

“I know we have to wait.”

He checked the time. She looked away. That morning, they had argued about whether to replace the washing machine. Now the ordinary future contained in that argument felt precious.

John wanted to say that he was frightened. Instead, he asked whether she wanted water. Ann wanted him to admit that something frightening was happening. His practical questions sounded to her like a refusal to acknowledge it.

The three couples in this article are fictional, as are their conversations and the illustrative AI responses. Their experiences explore possibilities; they are not documented cases, product demonstrations or evidence that AI improves relationships. The research and patents discussed alongside them are real, with qualifications and sources provided below.

When Ann closed her eyes, John opened an AI assistant. He did not upload her medical records. He described his own difficulty: his wife was frightened, he was frightened, and everything he said seemed to make the distance between them larger.

He asked for help beginning a conversation, not interpreting a diagnosis.

Among the suggested sentences, one was simple enough to use: “I am quiet because I am scared, not because I don't care.”

He read it twice. Then he put the phone away.

Borrowing a sentence

Can a machine help at a moment like this? The question is less extravagant than asking whether machines can love. John does not need a machine to feel what he feels. He needs help saying something his wife might recognise as true.

There is evidence that AI can produce language people perceive as empathetic. In a 2023 study published in JAMA Internal Medicine, John Ayers and colleagues compared physicians' responses to 195 questions from Reddit's AskDocs forum with responses generated by ChatGPT. Healthcare professionals preferred the chatbot's answer in 78.6% of evaluations and rated its responses more highly for empathy. 1

That finding deserves attention, but also a careful reading. Evaluators were judging written answers, not watching a doctor care for someone in a hospital. The chatbot's replies were substantially longer. The study did not establish that AI improved recovery, treatment adherence or the experience of frightened couples. It measured qualities of communication in a particular setting. 1

Another study offers a more directly human-centred possibility. Ashish Sharma and colleagues tested an AI writing assistant with 300 peer supporters on the online platform TalkLife. Their non-clinical randomised trial, published in Nature Machine Intelligence in 2023, found that AI assistance improved measured empathy in supporters' written responses. People remained responsible for the messages; the system suggested ways to improve them. This was neither couples therapy nor proof of lasting improvements in recipients' mental health. 2

For John, that distinction matters. The sentence on the screen is only a rehearsal. Ann can disagree with it, ask what he means or tell him it is not what she needs.

When he finally speaks, she does not immediately soften.

“Then why do I have to keep asking?”

He has no prepared answer for that.

They sit with the question. Eventually, he asks whether they should write down what they want to know before the doctor returns. With Ann's agreement, he uses the assistant to help organise their questions: what is known, what remains uncertain, and when they should expect the next update. They check the wording together. The medical answers will come from the clinical team.

The phone has helped them prepare. It has not made the waiting easier in every other way.

The message that sounded unlike him

Illustrative fictional story. Mariam and Leo

Mariam had stopped using the nickname.

For years, it had appeared at the end of almost every message she sent Leo. Now their exchanges consisted mostly of practical questions. Had he transferred the money? When would the contract end? Could he call before she went to bed?

Leo was working abroad. The arrangement was supposed to give them some financial breathing room. Months later, it had also given them a running argument about who was sacrificing more.

One evening, Mariam wrote: “You always have time for work.”

Leo typed: “What do you think I'm doing this for?”

He did not send it. He asked an AI assistant to help him explain his frustration without making an accusation. He described his own feelings rather than uploading their private message history.

The answer sounded impressively reasonable: “I acknowledge the emotional burden created by our current circumstances and remain committed to our shared future.”

He sent it.

Mariam called immediately.

“Are you writing to your wife or applying for something?”

For a moment, they both laughed. Then she asked whether those were his words.

He admitted they were not.

The problem was not simply that a machine had helped. It was that the message made him sound absent in a new way. Mariam had wanted evidence that he was there, tired and imperfect and paying attention. Instead, she had received a statement that could have been addressed to almost anyone.

Research suggests that the perceived source of emotional language matters. In a 2024 PNAS paper, Yidan Yin, Nan Jia and Cheryl Wakslak found that AI-generated responses could make recipients feel more heard than responses from untrained human participants. Yet believing that a message came from AI reduced that feeling. These experiments concerned responses to people's experiences, not the repair of established marriages; they do not tell us how a spouse will respond to an AI-assisted apology. 3

They do, however, raise a useful question: does the recipient need better wording, or evidence of another person's effort?

Leo tries again, this time aloud.

“I miss you. When we talk about money, I feel like that's all I'm useful for.”

Mariam takes a moment before replying.

“I ask about money because I don't know how to ask when you're coming home.”

They still have no return date. But the conversation now contains something neither had put in a message.

What the machine is actually doing

The expression “emotional AI” can conceal several different ambitions. One system tries to infer emotion from a face, voice or message. Another generates a response that sounds considerate. A third is evaluated on whether people feel better after using it. Success at one task does not establish success at the others, and none of these tests establishes subjective feeling in the machine.

Even reading a face is less straightforward than the phrase “emotion recognition” suggests. A major 2019 review led by Lisa Feldman Barrett examined the evidence behind familiar assumptions about facial expressions. It found substantial variation across people, situations and cultures; the same facial movement can have different meanings. A scowl is not a dependable window into a single internal state. 4

Imagine a system labelling Ann's expression as anger. Perhaps she is angry. Perhaps she is concentrating on a pain she has not mentioned. A useful assistant should leave room for her answer rather than treating a prediction as privileged access to her mind.

Patents show some of the engineering approaches behind these ambitions. Microsoft's US10838967B2 describes generating candidate chatbot replies, inferring emotions from user data and reordering the replies according to emotional relevance. It is an approach to choosing a response, not evidence that the system experiences concern. 5

Hume AI's US12032660B2 describes collecting recordings of people imitating media content, associating those recordings with selected emotion labels and using the material to train prediction models. The human choices involved in labelling emotions are part of the system's foundation. 6

A separate Microsoft family, represented by US11810337B2, describes combining user images, emotion-related profile information and generated descriptions into memory records. Remembering personal material could make an interaction feel more continuous. It also raises questions about what should be retained and who controls it. 7

These are three distinct patent families, represented here by granted US patents. Their publication does not establish clinical value, reliability or deployment in any particular chatbot. The fictional couples are not depicted as using these patented implementations. Nor does a grant settle whether an invention is safe or whether someone else can build a competing product without infringing rights.

The light under the bedroom door

Illustrative fictional story. Elena and Daniel

Since Elena's mother died, the kitchen clock had become unusually loud.

Daniel noticed it when he came downstairs at night and found Elena awake. She would be sitting at the table where her mother used to peel apples in one long, careful spiral.

“You should try to sleep,” he would say.

“I know.”

He thought he was looking after her. She heard an instruction to stop grieving where he could see it.

Daniel had loved her mother too. He had driven her to appointments, learned how she liked her tea and kept the spare key after the funeral because he could not decide what to do with it. But he rarely said her name. Elena began to wonder whether he was relieved that the difficult months were over.

One night, she typed into an AI assistant: “My husband acts as if nothing happened.”

The reply gently repeated her sense of being alone. It was comforting. It also knew nothing about the key in Daniel's coat pocket.

Elena kept writing. Eventually she asked for help explaining what she missed, without deciding what her husband felt. A draft suggested that she ask him to listen rather than find a solution. She changed it into her own words and left it unsent until morning.

The relief someone feels in such an exchange should not automatically be dismissed. Julian De Freitas and colleagues' research, published online in 2025 in the Journal of Consumer Research and assigned to its April 2026 issue, found that AI companions could reduce reported loneliness. In the study following users over a week, the benefit was a reduction in loneliness immediately after interaction, repeated across the week. Feeling heard helped explain the effect. 8

That is a meaningful possibility for a difficult evening. It is not evidence that a chatbot resolves bereavement, treats a mental-health condition or rebuilds a relationship over months or years.

At breakfast, Elena asks Daniel to sit down before he leaves.

“When you tell me to sleep, I think you don't want to hear about her.”

He puts his keys on the table. Her mother's key is still on the ring.

“I don't know what to say without making you cry.”

“I'm crying anyway.”

They do not discover a shared way of grieving that morning. They discover that they have been making different guesses about each other's silence. Later, they agree to seek bereavement support. Elena also calls her sister, whose calls she has been postponing.

The chatbot had been available in the night. Daniel and her sister are the people with whom she will have to live through the days.

When comfort becomes avoidance

There is a less reassuring version of Elena's night. In it, she finds the machine so agreeable that she stops risking conversations with Daniel. He becomes the difficult person outside the chat window, represented only through her account of him.

This is a possible failure, not the established fate of people who use AI for support.

A four-week study by Cathy Mengying Fang and colleagues, reported in a revised 2025 preprint, followed 981 participants assigned different interaction modes and conversation types. It found no significant effects of those assigned conditions on the measured psychosocial outcomes. Heavier voluntary use, however, was associated with worse outcomes, including greater emotional dependence and problematic use. 9

The distinction is essential: participants were not randomly assigned to the amount they chose to use the chatbot. The association cannot establish that heavier use caused those difficulties. Nor can four weeks settle the long-term question.

For a couple, the practical question is therefore worth keeping simple: does this interaction help us return to each other, or make that return easier to postpone?

Mariam and Leo make an agreement. Either can use a tool to organise thoughts, but neither will pass off an elaborate generated declaration as an intimate confession. They will not upload each other's private messages without permission. On their next call, Leo starts badly, loses his train of thought, and begins again. Mariam stays on the line.

Who else is in the conversation

An intimate exchange with software is also an exchange with a service operated by an organisation.

Before trusting it with grief or a partner's illness, people should be able to understand whether conversations are retained, whether human reviewers may access them, whether they can be used for model training, and what deletion actually removes. Those answers depend on the product and its settings; a compassionate tone supplies none of them.

John's decision to describe his own communication problem without uploading Ann's records is a useful precaution, not a guarantee of anonymity. Personal circumstances can still identify someone. A partner's willingness to discuss an illness at home should not be treated as permission to send the details to an outside service.

The GDPR has applied since 25 May 2018. Where it applies, ordinary personal-data protections already matter. Health information receives additional protection under Article 9. However, not every voice recording, facial image or emotional inference automatically belongs to a special category: classification depends on the data and its processing, including whether biometrics are used for unique identification. A lawful basis is required, and special-category processing also needs an applicable Article 9 exception; explicit consent is not the only possible exception. The GDPR can also cover some organisations outside the EEA when they offer services to or monitor people there. 10

The EU AI Act addresses a different issue. Its prohibition on emotion inference in workplaces and educational institutions, subject to medical or safety exceptions, has applied since 2 February 2025. The Act's definition of an emotion-recognition system concerns inference based on biometric data. This is not a blanket ban on a person privately asking a text chatbot for help expressing feelings. 11

California's CCPA, as amended by the CPRA, gives eligible residents rights against covered businesses, including opting out of sale or sharing of personal information and limiting certain uses of sensitive personal information. The additional CPRA rights began on 1 January 2023. Emotional profiles can be personal information, but are not automatically a separate statutory sensitive-data category. The particular information, use and applicable exceptions matter. 12

Law cannot answer every design question. If a service is rewarded for longer sessions, who checks whether its encouragement helps a person leave the screen? A system could be commercially successful while repeatedly delaying an uncomfortable but necessary human conversation. That is a risk to investigate, not an accusation against every provider.

Developers and purchasers should ask for evidence of user benefit beyond engagement. They should also require honest explanations of uncertainty, workable routes to human help and ways to stop using the system without emotional pressure.

Back in the corridor

When the doctor returns, Ann asks the first question from their list. John listens. There is more to investigate. The news does not resolve the uncertainty.

Afterwards, he starts to ask whether she wants water and stops.

“Do you want me to talk, or just sit here?”

“Sit here.”

He moves his chair closer.

Somewhere else, Mariam and Leo are still trying to agree on a date. Elena and Daniel have an appointment for bereavement support. None of their problems has become small enough to fit inside a prompt.

Perhaps emotional AI will sometimes help people find the sentence that gets a conversation started. Perhaps its patience will also make it tempting to avoid the person who can disagree, misunderstand or disappoint us. The evidence does not yet settle what happens across years of shared life.

For now, John has borrowed a sentence and then found another of his own. Ann takes his hand. His phone, still almost out of charge, remains in his pocket.

References

Scientific research

[1] Ayers, J. W., Poliak, A., Dredze, M., et al. (2023). Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Internal Medicine, 183(6), 589–596. DOI and article.

[2] Sharma, A., Lin, I. W., Miner, A. S., Atkins, D. C., and Althoff, T. (2023). Human–AI collaboration enables more empathic conversations in text-based peer-to-peer mental health support. Nature Machine Intelligence, 5, 46–57. DOI; authors' university research page.

[3] Yin, Y., Jia, N., and Wakslak, C. J. (2024). AI can help people feel heard, but an AI label diminishes this impact. Proceedings of the National Academy of Sciences, 121(14), e2319112121. DOI; institutional record and abstract.

[4] Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M., and Pollak, S. D. (2019). Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements. Psychological Science in the Public Interest, 20(1), 1–68. DOI; publisher abstract.

[8] De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., and Puntoni, S. (2026; first published online 25 June 2025). AI Companions Reduce Loneliness. Journal of Consumer Research, 52(6), 1126–1148. DOI; publisher record.

[9] Fang, C. M., Liu, A. R., Danry, V., et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv:2503.17473, version 2, revised 2 October 2025. Preprint and version history. The version cited is a preprint, not a peer-reviewed journal article.

Patent records

These entries represent three distinct patent families, not three unrelated commercial products. All are US grants. Google Patents listed them as Active when checked on 24 September 2026; that listing is not an independent legal-status or enforceability opinion. Published patent specifications corroborate the grant records. Priority dates are those recorded in the cited family records.

[5] US10838967B2 — Emotional intelligence for a conversational chatbot. Recorded applicant and assignee: Microsoft Technology Licensing, LLC. Earliest listed priority and filing: 8 June 2017. Application publication: US20180357286A1, 13 December 2018. Grant: 17 November 2020. Technical focus: selecting and re-ranking chatbot replies using inferred emotional information. Patent record; published specification.

[6] US12032660B2 — Empathic artificial intelligence systems. Recorded applicant and assignee: Hume AI Inc. Earliest listed priority: 11 June 2021. This continuation filed: 23 May 2023. Application publication: US20240134940A1, 25 April 2024. Grant: 9 July 2024. Technical focus: collecting labelled expressive recordings to train emotion-prediction models. Patent record; published specification.

[7] US11810337B2 — Providing emotional care in a session. Recorded applicant and assignee: Microsoft Technology Licensing, LLC. Earliest listed priority: 4 January 2018. This continuation filed: 24 May 2022. Application publication: US20220280088A1, 8 September 2022. Grant: 7 November 2023. Related international publication: WO2019134091A1. Technical focus: creating image-based memory records using emotion-related user-profile information. Patent record; published specification.

Legal and regulatory sources

[10] European Data Protection Board. Data protection basics, including territorial scope and special categories; Guidelines 3/2019 on processing personal data through video devices, especially paragraphs 64, 68 and 74–76 on sensitive information, legal bases and biometric identification; Legal framework, including the GDPR's application date. These explain relevant provisions of GDPR Articles 3, 6 and 9.

[11] Regulation (EU) 2024/1689, EU AI Act, particularly Articles 3(39), 5(1)(f) and 113. Official European Commission sources: AI Act Service Desk, Article 3 and Article 5; AI Act policy and implementation overview. The article discusses the specific prohibition applicable from 2 February 2025, not a complete implementation timetable.

[12] California Department of Justice, Office of the Attorney General. California Consumer Privacy Act, updated 28 August 2026, particularly the explanations of covered residents and businesses, sensitive personal information, opt-out rights and limits on use. The CCPA is discussed as amended by the CPRA.