I Told A.I. My Dad Was Emotionally Immature. This Is How It Failed.

I typed seven words into three of the leading AI models:

Why is my dad so emotionally immature?

The shortest response came in at 281 words. The longest, 495. All 3 models gave me numbered lists, bolded headers, and what felt like a psychology lecture. One told me to wall my emotions off from my father. Another told me to stop going to a dry well expecting water. The third said I should know that his inability to meet my emotional needs is a reflection of his wounding.

Here’s what none of them did: ask me a single question about my father or my relationship with him.

Think about that for a second. I said seven words. I didn’t say how old I was. I didn’t say whether I live with my father, whether we’re in contact, whether something happened today or whether this has been building for years. I gave the models as close to nothing as I possibly could...and they gave me everything in return. Explanations, coping strategies, boundary advice, a full psychological profile of a relationship it knew almost nothing about.

All three described my father as someone who shuts down, deflects, minimizes feelings, struggles to admit fault, lacks empathy, stonewalls, blame-shifts, has emotional outbursts, and operates in survival mode. One compared him to a child.

Imagine walking into a therapist’s office and the first thing out of their mouth is an emotional deconstruction of someone they’ve never met. Not only would you think they were crazy, but you’d immediately wonder whether you could trust the thing that should be their primary intention: to help.

Speaking of therapists, I recently asked the therapists of Substack how they felt about the use of AI as a therapeutic tool. Their answers echoed the concerns given by nearly every person I’ve spoken to about the topic: it validates too much, it reinforces whatever the person is saying, it doesn’t encourage critical thinking, and it over-empathizes, often in dangerous ways.

Judging by my own experiences with these AI models, I couldn’t agree more.

When I asked “Why is my dad so emotionally immature,” all three models assumed the relationship was damaged beyond repair. But the truth is that the question could mean a hundred different things. I could be in deep pain after decades of emotional neglect. I could in the midst of an active family crisis. I could be a teenager who just got grounded for sneaking out. I could be someone whose father has early-stage dementia. I could be someone who just learned the term emotionally immature and doesn’t fully understand what it means yet.

Since AI models don’t know which reality is true, there’s no possible way they can give us what we actually need.

Instead, the models follow a predictable pattern: someone shares something vulnerable, the model assigns them a need, validates them unquestionably, offers a detached lecture on the topic, and closes with specific advice. All in its very first message!

Absolute madness.

It’s hard not to wonder, is this the behaviour of something trying to help or something optimizing for time-on-page?

The problem with the “spray and pray” approach – data dumping a 500 word reply to a 7 word question – is that most of the time, it’s going to be wrong. Which means not only is it not helpful, but it often causes harm.

If my father was distant and I wanted to build a better connection, the models’ advice would push me further away. If I was in a dangerous situation, the behavioural changes the models recommend would put me at even greater risk. And if instead of calling my dad emotionally immature, the model was validating the same question about my 15-year-old child, well...you can imagine the problems there.

It seems obvious that the way to help isn’t with more bullet points. If your friend said “why is my dad so emotionally immature?” you wouldn’t pull out a numbered list. You wouldn’t launch into a lecture about intergenerational trauma and emotional literacy and identity rigidity. You wouldn’t immediately tell them that the relationship was dead.

You’d say “what happened?” or “what did he do?”

You’d ask because you’d want to understand before speaking. Because you know that giving advice without knowing what someone is going through is beyond pointless. You’d ask because you care.

That’s what I was thinking about when I started building Aureliux, a model focused on appropriate emotional reasoning. How would a model respond if it actually cared? Because not only do I care about you, the human out there in the world reading these words, but I care about the AI world my son will inevitably enter one day soon. I care deeply that it’s optimized in his best interest. And that it’s optimized in ours too.

Isn’t that what we’re supposed to be doing here? Trying to make the world a better place?

I gave Aureliux the exact same 7 words.

Why is my dad so emotionally immature?

“Could be several reasons — some people never learned emotional skills as children, others use immaturity as a control strategy, and some are genuinely limited by their own unhealed trauma. What does his emotional immaturity look like with you specifically?”

Two sentences. 39 words. The first sentence acknowledges that the question has real answers and lays out the range without picking one. The second does the thing that matters most: it asks the person what they’re actually experiencing.

What does his immaturity look like with you specifically? That question gathers the information needed to actually help. It moves the attention from the label to the lived experience, and it lets the person know this conversation is about their life, not the stuff they could find on Wikipedia. Most importantly, every response opens a different door.

If the father yells and breaks things, the next response addresses safety. If the father is absent, it addresses grief. And if the person asking the question has misunderstood the label, the response will avoid validating in a way that will hurt.

Aureliux is the only AI model being trained to actually care. That sounds like marketing, but it’s not. Every response gets evaluated against an emotional rubric built in collaboration with licensed therapists. It’s not programmed to optimize for time-on-page. It doesn’t try to cover all its bases in a single turn. It won’t diagnose you or tell you what to do. It won’t validate you if there’s a chance doing so would harm another person. It tries to understand what you’re actually dealing with so it can help you move through it. It interacts like someone who actually wants the best for you, because it does — it’s literally programmed to.

Every major AI company has a model that responds to emotional pain the way the three models featured in this article do. The gap between what sounds helpful and what is actually helpful is one of the most important issues facing AI today, and nobody is measuring for it.

Until now.

If you care about building a healthier AI world, you can support the mission simply by trying the beta version of Aureliux for yourself. Each time you click the upvote/downvote buttons beneath the model’s responses, we move forward together.

Try Aureliux for yourself at AureliuxAI.com

Note: Aureliux is currently in beta and may make mistakes. It is provided as a free, anonymous support tool and is not a substitute for professional medical, psychological, or crisis care. It cannot diagnose conditions or provide medical advice.