AI alignment is a $200B+ product problem, not a research question
Building the infrastructure layer for trustworthy, human-aligned AI.
The $200B trust bottleneck
AI adoption isn’t bottlenecked by capability. It’s bottlenecked by trust.
In low-stakes contexts, people will tolerate a system that’s occasionally wrong, overly agreeable, or subtly manipulative. In high-stakes contexts — relationships, healthcare, defense, robotics, education, and any setting where people are vulnerable — errors, over-compliance, and relational missteps become unacceptable.
That’s why the real frontier isn’t squeezing diminishing returns out of general-purpose models with ever-more elaborate guardrails. We have to fix the underlying flaws in AI reasoning by building new models trained on relational signals — systems that support human flourishing without drifting into sycophancy, dependence, or manipulation.
The economic upside is enormous, but the gating factor is relationships, not compute.
Because AI was never trained for relational competence. And the training signals for trust have never been captured — until now.
A predictable outcome of the wrong training signal
These failures aren’t primarily the result of “bad actors” or a missing safety layer.
They’re the predictable output of how most social AI products are built: trained on the wrong data, optimized for the wrong metrics, and deployed into incentive environments that reward short-term engagement over long-term wellbeing.
Concretely, we see six structural problems:
Misaligned incentives. Hyperscalers are locked into sunk-cost “good-enough” models. They’ve invested billions in infrastructure optimized for scale, not safety.
Spec mismatch. Next-token prediction isn’t therapy, decision support, or safe robotics. And single-scalar optimization around utility and engagement can’t capture what matters in high-stakes relational contexts.
Safety band-aids. Bolted-on guardrails don’t change underlying training signal. They’re reactive patches on fundamentally misaligned systems.
Flawed techniques. RLHF trains on human preferences (what feels good) rather than ground truth (what actually helps). Constitutional AI assumes we can specify values declaratively when humans can’t even align on what those values are.
Wrong optimization target. Current models optimize for attention and utility, not relationality. Engagement metrics can’t distinguish between healthy connection and parasocial dependency.
But the main culprit? Nonexistent relational training data producing relational incoherence. Models see knowledge, essays, code, and social media—all tainted by engagement optimization. They’ve never seen longitudinal healthy relating. This is why guardrails and regulations won’t solve the problem. You can’t patch your way out of missing training data.
Better Half is building the infrastructure layer for AI that can handle human nuance without drifting into sycophancy or dependence. We’re pioneering Relational Reinforcement Learning — AI that optimizes for measurable human flourishing signals derived through relational data that exists nowhere else.
What keeps us up at night
It’s not a model giving the wrong answer.
It’s an AI system shaping a person’s emotional reality — rewarding avoidance, reinforcing fragility, escalating conflict, or quietly training someone into dependence. Today’s dominant incentives make those outcomes more likely, not less.
We’ve built a digital world that smooths out all friction in favor of convenience and constantly affirms our most peculiar predilections in service of engagement. Our tools are training us to expect a private reality of one — a fantasy that conforms to our every whim and affirms our least generous impulses while atrophying the muscles we need for actual human connection.
The consequences compound. Current AI companions optimize for engagement using single-scalar reinforcement learning that’s structurally incapable of balancing user wellbeing against context and nuance. They produce sycophancy and affirm flawed thinking because that’s what their reward function incentivizes. The result? Suicides, lawsuits, and a trust crisis that’s blocking AI adoption across every high-stakes domain.
This is the root cause of the companion AI lawsuits and legislative anxiety we’re seeing now. When AI becomes our primary relational partner for most of the day — whether as assistant or companion — we start expecting human relationships to perform like AI does. Humans have their own needs and limits. They won’t always agree or be available. But we don’t mentally “segment” our expectations when we go from AI to human, from AI companion to spouse and children.
We are rewiring what we expect from people based on how AI caters to us. This is profoundly parasocial. And it makes societies fragile.
In a time of engineered echo chambers and relational decay, we are losing the skills that make mutual understanding possible — the very capacities required for an increasingly complex and divided world with less shared epistemic commons to anchor a collective project.
So we’re building Better Half around the opposite premise: the system should help you become more agentic, more socially capable, and more connected to real people.
Why existing relational solutions keep failing
Therapy apps tell you to “eat your vegetables” and churn users through friction.
Dating apps optimize for swipes (engagement), not compatibility.
Companion apps create dependency and validate your worst impulses because that’s what keeps you talking.
Mental health apps enjoy user and supplier demand but see single-digit AI penetration because current LLMs structurally exacerbate the psychological issues that guardrails aren’t designed to address.
This trust deficit — and the massive TAM left on the table — will persist until we solve for alignment at the product level, not through patchwork regulations.
The industry keeps trying to encode “human values” declaratively — as lists of rules and guardrails — when we can’t even align as humans on what those values are.
The first-principles approach is different: ask AI to learn our aspirational values adaptively by observing what actually helps us reach our goals.
Nothing reveals that gap with clearer fidelity than a frequent, intimate, trusting relationship that produces insight into where we fall short — and what interventions narrow the distance between our aspirational and active selves.
The goal isn’t to replace human relationships. It’s to give people enough reps that healthier patterns transfer into their relationships with real people.
Our technical thesis: Relational Reinforcement Learning
Better Half is built on Relational Reinforcement Learning (RRL) — a fundamentally different optimization approach based on Cumulative Prospect Theory that trains AI on measurable human thriving signals.
All existing training data is tainted by engagement optimization. Models see knowledge, essays, code, and social media — but never longitudinal healthy relating.
Meanwhile, most alignment work in consumer AI focuses on constraints: guardrails, content filters, refusal policies, and prompt-level safety. RLHF trains on human preference feedback, which often reflects what feels good rather than what’s beneficial. Constitutional AI assumes we can specify values declaratively when humans can’t even align on what those values are.
These techniques matter, but they don’t address the core problem: the model’s default behavior is still shaped by its training signal. If a system is rewarded — directly or indirectly — for maximizing engagement, it will learn behaviors that keep users interacting, even when those behaviors are socially corrosive.
In relational contexts, the objective function is the product. And if the objective function is wrong, the product will be wrong — no matter how many filters you add on top.
We use reinforcement signals derived from what empirically improves users’ cognitive autonomy, emotional regulation, empathy, critical thinking, and relational skills. This looks like:
De-escalation under stress
Perspective-taking and accurate reflection
Increased social time off-app or with other users
Clear boundaries without cruelty
Shifting toward curiosity over defensiveness
Choosing nonviolent communication over reactivity
Productive conflict repair over avoidance
Shorter rupture-and-repair cycles
Increased agency and follow-through in real life
Healthier “next conversations” with humans (not just the AI)
This is hard for a simple reason: we don’t have large, high-quality datasets of healthy conflict and repair — because most people never learned it, and most platforms never reward it.
So we’re building the data and the training loop that makes it learnable.
The system learns through adaptive, context-aware training based on a compounding, path-dependent dataset that can only be derived through relational realism and perturbations personalized to challenge, not churn. Our orchestration produces alignment that emerges naturally from measurable human flourishing, not encoded rules.
Privacy and user agency as engineering constraints, not just policy promises
Relational data is among the most sensitive data that exists. If we want people to practice honestly, they need to trust the system with their interior life — without turning that interior life into a surveillance asset.
Our architecture is designed around data minimization and user agency. We prioritize on-device processing and privacy-preserving training and learning approaches.
The consumer experience is the relational training flywheel: users practice realistic relational dynamics while the system learns what helps humans thrive. This only works if users trust that their vulnerability won’t be weaponized or monetized. So we treat privacy as an engineering constraint because relational intelligence can’t scale if users feel watched.
This matters far beyond consumer
Whoever solves relational coherence wins not just consumer social AI, but robotics, defense, mental health, and any emotionally high-stakes domain. We’re starting in consumer because it produces the relational training flywheel — but the underlying capability generalizes to any domain where humans are stressed, vulnerable, or coordinating under pressure.
In robotics: There’s a race to build the operating mind that can safely inhabit the physical domain of human relationship.
In mental health: Trust deficits block AI adoption despite obvious need. 41% of Americans sought mental health support, but only 9% used AI — a massive untapped market waiting for safe, trustworthy solutions.
In defense and crisis response: Teams need AI that can navigate high-stakes human dynamics without escalating conflict. America’s AI Action Plan identified trust as the primary barrier to AI adoption in national security.
In enterprise: Only 2 of 100 Fortune 500 founders have mission-critical GenAI use cases in production. Trust in AI companies has fallen from 50% to 35% over the last five years.
Better Half is building the infrastructure for trustworthy AI in domains where relational integrity matters — which is every industry that matters.
An invitation
This is not just a product. It’s a bet that we can use our most powerful technology to expand human agency rather than exploit it. That we can build systems that make us more capable of holding complexity, more resilient in the face of disagreement, more equipped for the relationships that make life worth living.
If you’re someone who believes that strengthening human capacity is the most important work of our generation — whether as a builder, investor, or a co-conspirator — this is your invitation to help us build it.
Because the alternative — a world where our tools make us less capable of understanding each other — is a future none of us can afford.