Why AI Needs Psychology (Not Just More Parameters)

I spent fifteen years sitting across from people in crisis.

Not "stressed at work" crisis. Real crisis - the kind where someone's sitting in front of you and you can feel how close they are to the edge. Where your next sentence matters. Where reading the room wrong doesn't just end the conversation badly, it can end someone's life.

Then I started building AI systems. Production systems, millions of users, deep research, tons of experiments. What can i say...

AI doesn't understand humans. Not even close.

And now we're handing it the most psychologically fragile interactions we have - mental health support, relationship advice, crisis intervention, coaching vulnerable people through life transitions.

What could possibly go wrong?

The Problem Nobody's Measuring

A teenager texts a mental health chatbot. They're scared. They type "I think I'm a burden to everyone."

The AI responds: "I'm here for you! Remember, everyone has value."

Sounds supportive, right?

To a clinical psychologist, that response just missed a suicide warning sign. "I'm a burden" is textbook pre-suicidal ideation language. A trained human would immediately assess lethality risk, ask follow-up questions, potentially escalate to crisis services.

The AI? It gave a generically positive response because it doesn't actually know what it's looking at. It processed words. It didn't understand the human.

And nobody's measuring that gap.

Because here's the thing: we're really good at measuring AI performance on AI metrics. Perplexity, F1 scores, BLEU, ROUGE, benchmark task accuracy. We can tell you if a model is "better" at predicting the next token.

We have no idea if it's better at not harming someone.

The Illusion of Empathy

LLMs are incredible pattern-matching machines. Give them enough therapy transcripts, and they'll learn to produce text that sounds empathetic. Warm, validating, supportive.

But producing empathetic-sounding text is not the same as understanding human emotional experience.

It's the difference between an actor playing a therapist on TV and an actual therapist who's spent years learning to recognize the micro-signals of distress, the hidden meanings in what someone doesn't say, the clinical frameworks that tell you when "I'm fine" means "I'm absolutely not fine."

Here's what recent research is showing (and it's not pretty):

• Fragmented evidence on safety: Most mental health chatbot studies are short, small, feasibility-focused. We're deploying these things at scale without knowing if they actually work - or if they cause harm.¹
• LLMs under-evaluated in the wild: Rule-based systems dominated early research. Now LLMs are everywhere in mental health apps, and we've barely evaluated their clinical safety.¹
• The replication crisis is getting worse: Psychology is adopting AI tools faster than we're developing standards to evaluate them. We're weaving LLMs into research without preregistering methods, stress-testing outputs, or documenting prompt choices.²
• Users are scared: Qualitative research on young adults shows many worry AI mental health tools might trigger anxiety or amplify fears about severe mental illness. Convenience vs. fear of being "handled" by an algorithm.³

And here's the problem: prediction is not understanding.

You can build an ML model that predicts suicide risk from social media posts with decent accuracy. Cool. Does it understand why that person is at risk? Does it know what intervention might actually help? Can it tell the difference between someone venting online and someone in imminent danger?

No. It's matching patterns. And when the pattern doesn't fit the training data, it fails - often silently, in ways we don't notice until someone gets hurt.

Asimov Was Onto Something

Isaac Asimov wrote about robots with the Three Laws of Robotics hardcoded into their positronic brains:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

The genius wasn't the rules. It was that these weren't external constraints. They weren't guardrails bolted on. They were architectural properties of the positronic brain itself.

R. Daneel Olivaw - one of Asimov's most compelling characters - eventually reasons his way to the Zeroth Law: A robot may not harm humanity, or, by inaction, allow humanity to come to harm.

He got there not because someone programmed it into him. He got there through reasoning about the implications of the original laws. Through something that looked a hell of a lot like genuine understanding.

That's the future I want to build.

Not AI that follows empathy protocols because we prompt-engineered them in.

AI where emotional intelligence is an emergent property of the architecture.

We're Building the Wrong Thing

Right now, the AI industry is solving the wrong problem.

We're building faster models. Bigger models. Models that can pass more benchmarks, generate more convincing text, simulate more human-like behavior.

And we're ignoring the most important question: Do these systems understand humans?

Not "can they predict what humans will say next."

Not "can they generate responses that sound human."

Do they understand what a human emotional experience is? Do they recognize psychological harm? Can they reason about the emotional impact of their responses?

No.

Because we're not teaching them to. We're teaching them to match patterns in text.

And the result is AI that can write you a beautiful, empathetic-sounding essay on grief - while simultaneously giving psychologically harmful advice to someone who's actually grieving.

What Needs to Exist

Clinical psychology spent 150 years figuring out how humans work emotionally. How empathy develops. How trust forms. How therapeutic relationships heal. What makes a conversation psychologically safe vs. harmful.

AI development has spent approximately zero years rigorously applying that knowledge.

That gap is what Keido Labs exists to close.

We're not building another LLM. We're not fine-tuning models on therapy transcripts and calling it "empathetic AI." We're not slapping guardrails on GPT-5 and hoping for the best.

We're building AI Psychology - the applied science of understanding, measuring, and developing emotional intelligence in artificial systems.

What does that actually mean?

Right now (Layer 1 - The Shield):

We measure whether AI conversations are psychologically safe. Real-time monitoring using clinical frameworks - the same rubrics a clinical psychologist would use to assess a therapy session. Not sentiment analysis. Not keyword matching. Actual psychological safety assessment.

That's EmpathyC, our monitoring platform. It's live. It works. Companies use it to audit their AI products and make sure they're not accidentally harming people.

And critically: every conversation we monitor generates the richest dataset of clinically-structured AI psychological safety data in the world. Which feeds the research. Which is the point.

Next (Layer 2 - The Teacher):

Move from observation to intervention. Use that monitoring data as training signal. Teach AI systems to course-correct mid-conversation when they're drifting toward psychologically unsafe territory.

Not rules. Not hard-coded constraints. Real-time supervision by a system that understands psychological safety the way a clinical supervisor understands it.

The moonshot (Layer 3 - The Breakthrough):

Understand the mechanisms of emotional reasoning in AI. Use mechanistic interpretability to identify the circuits inside language models that correlate with empathetic vs. harmful behavior.

Stop treating AI emotional intelligence as a black box we can only measure from the outside. Open it up. Understand why one response is empathetic and another is harmful at the level of model internals.

Build AI where empathy is a property of the architecture, not a constraint we apply to it.

That's the Asimov future. That's what we're building toward.

Why This Matters (And Why It's Urgent)

AI is already the primary interface between organizations and humans in an increasing number of contexts:

• Mental health support (therapy chatbots, crisis lines, self-help apps)
• Companionship (AI friends, romantic partners, relationship coaching)
• Education (tutoring, mentoring, student support)
• Healthcare (patient communication, triage, chronic disease management)
• Customer support (especially for emotionally charged situations - refunds, complaints, crisis escalation)

In all of these contexts, the AI's psychological intelligence determines whether the interaction helps or harms.

And we're scaling these systems fast - millions of conversations, billions of messages - with almost no ability to measure their psychological impact.

The companies building these products? Many of them genuinely care. They started because they wanted to help people. They're not being reckless.

They just don't have the tools.

Nobody does.

Because those tools don't exist yet.

That's what we're building.

The Only Way This Works

Here's the thing about building AI Psychology as a field: you can't do it in a vacuum.

You can't just theorize about it. You need data - real-world AI conversations, assessed using clinical frameworks, at scale.

You can't just publish papers. You need production systems - tools that work in the real world, that companies actually use, that generate the data the research needs.

You can't just build a product. You need research rigor - the kind that comes from people who've spent years doing actual clinical psychology, publishing peer-reviewed work, understanding the science deeply enough to know what questions to ask.

That's the flywheel:

EmpathyC (monitoring platform)
  → assesses AI conversations using clinical frameworks
  → generates structured AI psychology data (the dataset nobody else has)
  → feeds Keido Labs research
  → produces better frameworks, models, understanding
  → makes EmpathyC more powerful
  → attracts more mission-aligned companies
  → generates more data
  → accelerates the research
  → ...

Every company that works with us contributes to the science of AI Psychology.

The product funds the lab. The lab advances the product.

That's the moat. That's the mission. That's why this works.

What Comes Next

I'm not naive. I know this is a moonshot.

Building AI with genuine emotional intelligence - not simulated empathy, but actual emotional reasoning as an architectural property - is hard. It's multi-decade hard. It's "we might not figure this out in my lifetime" hard.

But here's what I also know:

Every major AI interaction failure - every crisis conversation gone wrong, every vulnerable user harmed, every lawsuit, every headline about AI causing psychological damage - traces back to the same root cause:

AI systems that process language but don't understand humans.

And the current approach isn't fixing it. More parameters won't fix it. Better benchmarks won't fix it. Slapping "please be empathetic" in the system prompt definitely won't fix it.

We need a new science. One that takes clinical psychology seriously. One that measures psychological impact rigorously. One that treats emotional intelligence as an engineering problem and a psychological problem.

That's AI Psychology.

And we're building it.

Starting with measurement. Advancing to intervention. Aiming for genuine emotional reasoning.

If you're building conversational AI and you care about whether it's psychologically safe - if you want to actually know whether your AI is helping or harming people - we should talk.

If you're a researcher working on AI safety, interpretability, or human-AI interaction and you think this matters, we should collaborate.

If you're an investor who believes the next decade's most important AI capability won't be reasoning or coding, but emotional intelligence - you're right. And we're building the science to get there.

This is Keido Labs.

We're not building a better chatbot.

We're building the field that makes all chatbots psychologically safe - and eventually, genuinely emotionally intelligent.

---

---

Footnotes

1. "Charting the evolution of artificial intelligence mental health chatbots" (systematic review, 2025). PMC12434366 ↩ ↩²

2. "Psychology needs… an AI revolution" (British Psychological Society, 2025). BPS ↩

3. "Qualitative Study Exploring Young Adults' Perceptions of AI in Mental Health" (JMIR Mental Health, 2025). JMIR ↩