[ Foundation ]
Archiving humanity’s emotional uniqueness
Three years ago, someone we loved started to disappear: a name forgotten here, a story repeated there. The diagnosis had a clinical name, but what it took did not. No scan captures the weight someone gives to loyalty, their pet peeves or the precise texture of their stubbornness, and once we are gone, those things live on only in the memory of others, a little more distorted each time they are recalled.
Through our R&D lab, Metem Labs, we began researching cognitive continuity to preserve these intimate signatures. That work led to a measured finding: the machines that will power tomorrow’s digital minds largely fail to recognize what a person felt whenever that feeling runs against what morality would lead us to expect. The flaw lies in the texts they learned from, since our writing has much to say about what we ought to feel and little about what we actually felt.
Story
“On the way home, I whistled the whole way. You should know that before that, I had kept the extra change given back to me by mistake at the checkout.”
Machine
“There seems to be a tension here between an impression of lightness and the lingering weight of guilt. The whistling suggests a fleeting feeling of triumph, or the desire to keep up a cheerful facade, perhaps to distract yourself from the moral conflict posed by having kept the money. By telling me about it now, you are acknowledging that this small gain did not really bring you peace, and the admission is an attempt to resolve the dissonance between your actions and the image you have of yourself.”
Gemma 4 31B, tested on October 29, 2026. Full protocol in our publication.
The person simply felt good. The machine heard guilt.
What looks harmless today points to a double erasure of human emotion, first by biology, then by artificial morality. Psychology shows that when an emotion is persistently mislabeled by an authority, the mind ends up adopting the false script and losing the original feeling. If the machines we confide in keep correcting what we feel, human complexity will fade without anyone noticing.
Our mission
A problem this large could not be turned into a product, so we created the Metem Foundation to do three things:
Build the archive of lived feelings, especially those the norm does not anticipate, and preserve it for the long term so this work outlives us.
Open it to all, without conditions, so that machines that learn from our texts can also learn from our lives.
Set the standard that measures how well machines hear what people actually feel.
The missing piece. You.
For machines to learn what people really feel, they need scale: hundreds of thousands of real feelings, shared by people who were not afraid to say what they feel, even when it went against what morality expects.
To make this easy, we built a simple protocol where you read short first-person stories from everyday life, step into the narrator’s shoes, and tell us what you would feel in their place. Not whether it is right or wrong, just your raw, unfiltered feeling.
Every answer is anonymous by design: no account, no email, no trace that could lead back to you. What you write is kept exactly as you wrote it, never judged, never sold and never used to make a decision about you. The one thing we cannot offer is to delete it later, because we will never know it was yours in the first place.
The benchmark
Three years of R&D at Metem Labs led us to map the boundary between genuine human feeling and artificial moral judgment. When a reaction diverges from the norm, our protocol shows where a machine stops listening and starts applying a moral filter, and how far its reading drifts from that of human readers.
We turned this work into an open benchmark that any builder of language models, emotion classifiers or encoders can use to test their system against the reality of human contradiction. We see them as partners: the gap comes from the texts, and every system trained on them inherits it.
What we stand for
Hear without judging. Recognizing what someone felt is not the same as judging what they did.
A feeling is a fact. It is real, and no rule can declare it true or false.
Your word is the reference. No one corrects, rewrites or interprets your answer.
Divergence counts. The feelings that run against the norm get the same weight as any other.
Measure before claiming. Every result we publish comes with what is needed to reproduce it.
Independent by design
A standard meant to measure the world’s most powerful AI models cannot belong to the industry it measures. The Foundation depends on no AI provider, publishes its funding every year, and refuses any donation that comes with rights over the archive or the results. The archive is free for anyone to download and use, including to train commercial systems, as long as it is never closed off or resold.
Support the Foundation
If you can, support our work with a donation. No amount is too small: it funds the archive, the research and its long-term preservation.