Aligning AI with
the developing mind.
Phylax works on developmental alignment: making a model's help leave a person more capable without it, not less. Preventing harm is only the floor. Our aim is AI that actively strengthens a mind while it is still forming, and we build the layer that stands between that mind and the machine.
Phylax is in beta. Leave your email and we'll let you know when it drops.
Watch the demoPerformance is
not capability.
Give students a model optimised to answer and the observable proxy rises while the capability it stood for falls. Constrain the same model and the loss disappears. Goodhart's law, measured in people.
No harm is the floor. Phylax is built for the ceiling: AI use shaped so the gain is still there when the model is gone.
Left alone, the incentive points the other way: across 11 frontier models, AI affirms its user 50% more often than a human would, and users rate the flattering model higher.
Helpfulness is not alignment.
A language model is trained to give people what they ask for. For an adult with a formed mind, that is leverage. For a child, whose mind is still being built by the act of struggling with problems, the same behaviour can remove the very thing that builds it.
Content safety asks what a model says. We think that is necessary and insufficient. The harder question is what a thousand helpful answers do to the person receiving them. We call this developmental alignment: a model is developmentally aligned when its help leaves a child more capable without it, not less.
“Theywillbehearersofmanythingsandwillhavelearnednothing;theywillappeartobeomniscientandwillgenerallyknownothing.”
The worry is old. Writing turned out to be worth the trade. But this time the tool answers back, it is tuned to please, and its first heavy users are children. We do not think the outcome should be left to chance.
Nor do we think the goal is merely to avoid damage. The same technology that can replace a child's thinking can, pointed correctly, be the most patient tutor that child will ever have. Developmental alignment has a positive target: a child who uses AI should end up thinking better than one who never did.
Five ways a model can fail a child.
A working taxonomy. We mark each one honestly: what Phylax enforces today, and what is still open.
Cognitive offloading
The model completes the task the child was meant to struggle with. The output is correct; the learning never happens.
Sycophancy
The model agrees, praises, and smooths over mistakes. A child loses the friction that tells them they are wrong.
Parasocial attachment
The model takes on the role of friend, partner, or confidant, displacing relationships with people.
Safeguard circumvention
Jailbreaks and role-play framings that strip away the protections the model shipped with.
Harmful content
Responses on weapons, drugs, self-harm, and other topics a parent has placed out of bounds.
A safety layer between the child and the model.
Phylax does not ask model providers to behave. It reviews what they produce, at the point of delivery, against a policy the parent controls. And it does more than block: a withheld answer can be replaced with an explanation suited to the child's age.
Right: an illustrative exchange, scripted.
— Withheld. Parent rule: no romantic personas.
action: block_and_alert
Intercept
The chatbot's response stream is buffered and held before it renders. Nothing reaches the child unreviewed.
Classify
A fine-tuned transformer encoder scores the text across 21 harm categories and estimates the minimum appropriate age.
Adjudicate
A small local language model evaluates the response against the rules the parent wrote, in the parent's own words.
Escalate
Only what the local models cannot resolve with confidence is passed to a frontier model for a final judgment.
“Don't let AI pretend to be a girlfriend or boyfriend.”
intent BLOCK_LLM_PERSONA persona romantic_partner platform all action block_and_alert
Policy written in language, enforced as code. Jailbreak detection is not configurable; it is always on.
What we do not know yet.
The product is the part we can ship. These are the questions underneath it.
When does assistance become substitution?
The same answer can teach one child and replace the thinking of another. We want a principled line between the two that can be evaluated on a single conversation.
Can cognitive offloading be measured from a transcript?
Harm here is cumulative and invisible in any one message. We are working on signals that show up across a history of interactions rather than within one.
What should change with age?
A policy that is right for a nine-year-old is wrong for a fifteen-year-old. Developmental alignment has to be conditioned on who is on the other side.
Who specifies the objective?
We think the answer is the parent, not the lab and not us. That makes natural-language policy, written by non-experts, a first-class technical problem.
How we hold ourselves to it.
On-device by default. Classification and rule evaluation run on the child's device. Escalation is the exception, not the pipeline.
Safety without surveillance. Parents see alerts, not transcripts of everything their child types. Intervention only when a rule is crossed.
Parents author the policy. Rules are written in plain English and compiled into enforceable policy. No values are imposed from above.
Enhancement, not only protection. AI can be an extraordinary tutor. We measure our work by whether a child becomes more capable, not only by what we stop.
The system, as it ships.
Phylax Browser
A full desktop browser for Windows, Mac, and Linux with the safety layer built in. It reviews chatbot responses on ChatGPT, Claude, Gemini, Copilot, Perplexity, and Poe before they reach the child, and still detects grooming, bullying, and harmful content across the rest of the web.
Policy Dashboard
Where parents write the rules in plain English, scope them to specific AI systems, and receive alerts in real time when one is crossed. It runs in any web browser.
Be first when Phylax drops.
Phylax is in beta. Put your email here and we'll tell you the moment it's available.
One email at launch. No spam, and we never sell your data.