A portrait collection can be searched by artist, title, date and medium. It cannot be searched by grief.
Collections as Data · Critical AI
The Trained Gaze is an independent research initiative. It builds machine-readable metadata for deciphering historical portraiture, with an added emotion-reading layer.
The data records emotion as a question of power, as well as expression, from documenting the sitter’s face and visual signals to artist-sitter-patron dynamics and the colonial gaze, where applicable. It asks what centuries of collected faces reveal about whose inner lives have been considered worth preserving.
Museums have spent fifteen years opening their collections. Images are free to use, records are downloadable, APIs are public. What remains out of reach is a different kind of question. Someone searching for a sad painting, a defiant face, a portrait of a confident woman has no way to find one, however exactly a work in the collection answers them.
Emotional language does appear in catalogue essays and wall text. It appears incidentally and unevenly, as prose, which means it cannot be queried, filtered, counted or compared across a collection. This project treats that language as a field rather than as commentary.
The remaining problem is not access. It is being findable by people who do not yet know what they are looking for.
Method
A metadata layer for portraiture, built on open-access collections.
Portrait records describe what a work is. This project adds a layer describing what it holds: the emotional register of the sitter, recorded as structured fields alongside the conditions that register was produced under, so it can be searched, filtered and compared rather than read one work at a time.
What the record holds
Every portrait receives an emotion reading. Further layers explain why that expression appears (where records support it): who commissioned it, who controlled the image, and what history sits behind it. Gaps are recorded as gaps, so the data never claims more than the evidence holds. The emotional reading is never stored alone.
Emotion as a condition, not an expression
A face in a commissioned portrait is performing under constraints: sitter identity, rank or patronage reveals what the sitter was permitted to show. An expression separated from those conditions is a surface and nothing more, which is why the conditions are fields in the record rather than notes beside it.
Source provenance
Every portrait enters the set from a collection that has released the image under open access or public domain terms, with the licence and the holding institution recorded against the work. Rights are verified before a work is used or published, not after.
Against physiognomy
Nineteenth-century emotion classification carried racial and class assumptions that modern AI models have inherited. This layer is built to resist that inheritance: neutral categories alongside emotional ones, sitter circumstances recorded explicitly, and no claim that a face reveals a nature.
Pilot study
The divergence study.
Automated emotion recognition systems (hereinafter emotion AI) are arriving in museums, mostly used in visitor sensing and in engagement metrics. They are trained almost entirely on photographs of living people instructed to perform an expression. A painted face in art history is not that: it was composed across sittings, governed by convention, and constrained by who commissioned it.
The pilot compares the emotion-reading of portraits between human encoders and the machine. Each portrait is read independently by a model that sees only the image. A second art-historically trained reader codes a subset under the same conditions. The three readings are then compared.
The question is not whether these systems work. It is what they systematically fail to see, and whether those failures fall in a pattern an art historian would recognise.
News
Talks and appearances.
Talks and appearances will be announced here.
Research notes
Notes on emotion, taxonomy and discoverability in portrait collections.
Research notes will be published here. No notes currently match this selection.
Collaborate
Open to partnership.
The work below is offered to institutions, researchers, educators and collectors. Enquiries welcome at any stage, including early ones.
Collections and pilot projects
Applying the metadata layer to a defined part of a collection, with the aim of making works findable to audiences who begin their search with a feeling rather than a name. Of particular interest: portrait holdings, digital strategy teams, and collections already published under open access. A pilot is scoped around a set of works rather than a whole collection.
Research partnerships
Visitor and audience research into how people search for and respond to portraiture, including survey design and on-site studies with museum audiences; engagement studies testing whether an emotional layer changes what visitors find and how long they stay with it; and independent second readings, where art historians, museum studies and digital humanities researchers code a set of portraits against the protocol. Inter-rater work is credited by name.
Talks and presentations
Available for booking by museums, universities, conferences and professional networks. Current subjects include emotion as metadata in historical portraiture, what automated emotion recognition fails to see in painted faces, the colonial record inside portrait collections, and discoverability as the unfinished half of open access. Talks can be adapted for curatorial teams, students or general audiences, and are given in English.
Workshops — critical AI for cultural institutions
Practical sessions for museum and collection staff who are being asked to evaluate, procure or deploy AI tools on their holdings. What these systems were trained on, what they inherit from nineteenth-century expression classification, which questions to put to a vendor, and how to read a confidence score. Adaptable for digital teams, curators, educators and students, and usable as gallery-based teaching with real works.
Writing and publishing
Commissioned essays, catalogue contributions and research pieces on portraiture, emotion, metadata and the critical reading of automated systems in cultural heritage. Peer-reviewed work in preparation.
Collectors and private collections
For collections held privately, the same layer answers a different question: what a collection holds emotionally as a body of work, and how its mood maps across periods, sitters and acquisitions.
henrietta.y.mansfeld@proton.me · Instagram @thetrainedgaze · LinkedIn
About
Henrietta Y. Mansfeld.
A decade of experience across the art world’s institutional and commercial sectors: the Cruz-Diez Foundation in Paris, independent curatorial work, and metadata cataloguing and curation at Artnet in Berlin, where the absence of an emotion field in a working art database has become the starting point for this project.
Berlin-based, Henrietta is a digital curator and data & AI researcher, working between art-historical scholarship and computational method.
She is a member of the Europeana network (Education, Technology, and Research communities) and a member of the Digital Art History Society (US).