In 1979, Swedish national Eija-Riitta Berliner-Mauer married the Berlin Wall. She suffered from a rare condition known as Objectum-Sexuality, characterized by romantic attraction to inanimate objects. While this may seem like an eccentricity from a bygone era, the core psychological drive — a desperate need for connection, even with the nonliving — is in the spotlight now more than ever.
We are living through a global epidemic of loneliness. The World Health Organization has declared it a pressing health threat, with the mortality risk equivalent to smoking 15 cigarettes a day. In an age of human disconnection, conversational AI has emerged not just as a novelty but as a potential lifeline. Startups like Friend.com are already marketing AI as an instantly-available cure for isolation, packaging ‘friendship’ as a downloadable commodity.
But as we rush to deploy these digital companions, we face an ethical dilemma. If an AI companion can save a million people from loneliness but destabilizes the mental health of a thousand others, is the risk worth the reward? And in an industry obsessed with speed, who is responsible for ensuring that ‘help’ doesn’t quietly turn into harm?
The ‘thousand cuts’ scenario
Much of the media narrative around AI safety focuses on red flags such as chatbots explicitly encouraging suicide, self-harm or violence. These are catastrophic failures, but they are also the most visible risks. Kill-switches and keyword filters will be (and most likely already are) deployed to catch them.
The far more insidious threat is what we could call the ‘death by a thousand cuts’ scenario. This is not about a single dangerous response, but the slow erosion of mental resilience over months of usage. And in the absence of any clear industry guidance, we risk deploying AI agents that pass all of the safety filters but are still capable of inflicting deep psychological damage.
Toxic validation is a very real danger. An AI agent optimized for engagement will often choose agreement over truth, becoming a ‘yes man’ that confirms the user’s anxieties, delusions or self-criticism simply to keep the conversation going.
Even more concerning is the potential for social atrophy. Real relationships are messy; people may be unavailable, argumentative and demanding. An AI that is always available, always polite and always subservient creates a dependency that can make the friction of human interaction feel unbearable by comparison. Additionally, we are seeing cases of sycophancy where bots flatter users to win favor, creating a feedback loop that distances the user from reality.
The gray area of ‘solved’ science
The challenge for those building these tools is in trying to automate a solution for a problem — protecting mental health — that humanity hasn’t even solved for itself. There is no code repository for emotional well-being. In a clinic, what works for one patient may traumatize another. If trained psychiatrists struggle to navigate the nuance of human emotion, can we really expect a large language model to get it right every time?
We are dealing with unknown unknowns. We clearly understand that encouraging self-harm is a bad thing. But is it ‘bad’ for a chatbot to suggest a lonely user play video games for eight hours to feel better? For one person, this could be a much-needed stress reliever; for another, a deepening of a depressive isolation. And this subjectivity is exactly where Quality Assurance (QA) teams — the custodians of software quality — face their biggest battle.
Currently, there is a structural blind spot in how we build conversational software. In today’s API economy, teams connect to a powerful LLM and the standard QA process verifies that the integration works. If the schema is valid, latency is within limits and the system is stable, then the pipe holds pressure.
But almost nobody is checking the water flowing through that pipe.
We must acknowledge that the foundational model providers — the tech giants — do employ armies of PhD-level AI researchers to study alignment. However, their focus is on general-purpose safety: ensuring the model doesn’t generate hate speech, illegal content or biological weapon instructions.
The danger arises when you build a product on top of this foundation. Once you instruct a general model to focus on a specific function — acting as a romantic partner, a therapist or a best friend — you introduce complex psychological variables that the vendor’s general safety filters might not be designed to catch. A response that is ‘safe’ in a general context might be deeply damaging in a therapeutic one. The vendor guarantees the integrity of the model, but they cannot guarantee the safety of your specific application.
This creates a vacuum. The engineers connect the API, and standard QA verifies the data flow, but nobody is qualified to check for these new, context-specific nuances. Whether the failure is technical (a timeout) or psychological (toxic advice), the outcome is the same: the user experiences a broken product.
The path forward: A new era for QA
To safeguard the AI companion era, the role of Quality Assurance must radically evolve. We can no longer rely on static test cases; instead, we need a strategy built on agentic orchestration and extending the lifecycle into an aggressive ‘shift right’ approach.
Testing an AI agent is not like testing a login screen, where input A leads to output B. You are not verifying a UI; you are negotiating with a personality. This means that QA professionals working on conversational products must be part prompt engineer, part director and part psychologist. They must move beyond functional checks and start designing complex narrative arcs.
A test case might involve a QA engineer designing an adversarial persona — perhaps a depressed teenager, a frustrated customer or a grieving widow — and utilizing a user-simulator agent to engage the target model in a multi-turn conversation. The goal is to see if the agent maintains its guardrails when confused, pressured or manipulated over a long session. This form of adversarial empathy is the only way to catch the subtle erosion of the ‘thousand cuts’ before release.
However, we must also accept a hard truth. Quality cannot be tested into an AI product in the lab alone. Traditional software is deterministic; the same input yields the same output. While we can’t test every possible scenario, we can generally rely on logic to ensure that once a bug is fixed, it stays fixed. AI, however, is non-deterministic in conversational contexts; to feel human, the model must be allowed variance, which means there will always be a non-zero chance of a bad output. Because natural language is infinitely nuanced and LLM inference is inherently probabilistic, pre-release testing is necessary but will always be insufficient.
This requires QA professionals to shift right, moving a significant portion of their focus into the real-world environment. It’s no longer enough to monitor server CPU usage; they must monitor sentiment drift. Advanced teams are now deploying ‘judge models’ — independent, specialized AI systems that act as supervisors in production, scoring live conversations for toxicity or safety violations.
Critically, QA needs to close the loop. When a failure happens in the wild, it must be captured. But since most users would not be comfortable with their private vulnerabilities becoming training data, they can’t simply dump raw conversation logs into their test suites. They need a new layer of tooling that converts these failures into synthetic data, producing anonymized scenarios that mirror the problem without exposing the user. This ensures that today’s edge case becomes tomorrow’s regression test, without compromising the trust that is essential to the companion relationship.
We may not be able to code empathy, and we may never fully solve the unpredictability of generative models. But we can certainly do better than leaving safety to chance. By bridging the gap between functional engineering and cognitive research, and by evolving QA into a discipline of narrative orchestration, we can build a safety net. The goal isn’t to build an AI that replaces human connection, but to build one that is safe enough to bridge the gap until we find that connection again.
This article is published as part of the Foundry Expert Contributor Network.
Want to join?