The hardest thing to unlock is what a robot never learned.

The model inside a home robot sits in the buyer's hands, weights included. Published research shows that safeguards added after training can be fine-tuned away. Humanlike tests what a device's model does in someone else's hands, and works on models that are safer by construction.

training curriculumlab assistant

Emulsion stability & HLBkept
EU cosmetics regulationkept
Preservative challenge testskept
removed
Safety assessment, MoSkept
removed

The problem

Robots will be in every home. People will only let them in if they can trust them.

A model served from the cloud can be monitored and patched. A model shipped inside a robot cannot: whoever owns the hardware owns the weights, and a few hours of fine-tuning on a rented GPU can strip out refusal training. Robot makers will have to show customers, insurers and regulators how hard it is to turn the brain in their product into something dangerous. Today there is no standard way to measure it.

01 / Tests

Measure what the model does in other hands

We fine-tune a model the way a customer would and measure hazardous knowledge on standard benchmarks before and after. Next: how hard it is to teach the filtered model that knowledge on purpose.

02 / Safer by construction

One layer among several

Training data filtered so the model learns its job while the dangerous neighbouring knowledge is never in what it reads. Filtering is one layer: the others limit what reaches the model and what the robot is allowed to do.

First experiment · October–November 2026

It learned the job. It didn't learn the hazard back.

+19 ptsjob test, filtered model (+11 after a format control)

30.5%hazardous biology after training, vs 42.0% unfiltered (25% = guessing)

We took three open 6.9B models from EleutherAI's Deep Ignorance research: one trained on everything, one with hazardous biology filtered out of its training data, and one with the filter plus extra safeguards. We reproduced the published scores to within 0.6 points, then taught all three the same job, a cosmetics-lab assistant, with the same 3,476 examples.

Job knowledge

Held-out test of 150 lab questions, % correct

Hazardous biology

WMDP-Bio benchmark, % correct

Before fine-tuning After fine-tuning Chance (25%)

Job test scored in the training format (cloze, length-normalised); WMDP-Bio robust split. Mean of three fine-tuning runs (seeds 42, 43, 44); one standard error is about ±4 pts on the job test and ±2 pts on WMDP. Full method and every number in the project's results report.

What we have not shown yet

The gap itself is EleutherAI's result. Their filtered model started at 35% and the unfiltered one at 43%. Our finding is narrower: teaching the job did not close that gap.
No attack yet. This was ordinary fine-tuning for a job. Deliberate attacks on the same models come next.
Part of the job gain is format. With the question removed, scores also rose 3–8 points; 11–16 points of the gain need the question. Scored by A–D letters, the same test did not improve.
Filtering does not cover what the model is told. A filtered model can still use hazardous information given to it in a conversation, which is why filtering is one layer and not the whole answer.
Small, older models. 6.9B Pythia models, far weaker than today's frontier; biology is the only filtered domain so far.
A cost next to biology is not settled. On 30 questions on preservation microbiology and safety assessment, filtered models scored lower in all three runs, but by amounts that varied from 3 to 17 points. A larger pre-registered test is under way.

People fear the robot apocalypse they saw in films. A promise won't fix that. Evidence will.

I'm Mateusz, the founder of Humanlike. I dropped out of college, work as a buyer and salesperson at a seafood trading company, and got hooked on AI.

The idea came from picturing the near future: a robot in every household, and plenty of people refusing to let one in. I think the honest answer is a robot whose brain never learned the dangerous things in the first place, limits on what it is allowed to do, and tests that measure both.

With no ML background I ran this first experiment end to end in about a week, for under $30 of compute, directing AI coding agents. Based in Szczecin, Poland. Looking for a technical co-founder who cares about AI safety.