Heritage · algorithms · critical thinking

Artificial Intelligence and Rembrandt

Reconstruction of lost parts, attribution assistance, image generation and risks: what algorithms can actually learn from a master who died in 1669.

From The Next Rembrandt to Operation Night Watch · uses, evidence, limits and transparency

The immediate response

Artificial intelligence can compare large sets of images, reconcile a copy of an original, digitally reconstruct a lost part or spot regularities that are difficult to measure manually. She doesn't "understand" yet neither Rembrandt nor his intentions, and it does not transform probability into attribution. A reconstruction must be distinguished from the preserved work; a generated image must be reported; an expertise must remain based on the material object, provenance and human judgment.

2016introducing The Next Rembrandt

2019launch of Operation Night Watch

1715cutting of La Ronde de nuit

3 usesrebuild, compare, generate

Three uses that should not be confused

Documented reconstruction

The algorithm aligns a historical source - for example Gerrit Lundens' copy - with the current original and helps translate missing parts in the scale, palette and texture of the painting. The result remains hypothetical and reversible.

Analysis assistance

Models can classify brushstrokes, compare patterns, segment pigments or spot similarities. They offer correlations that researchers must check on the object.

Stylistic generation

A system combines learned characteristics to produce a new image "in the manner of". It does not reconstruct any attested lost work and does not create an original by the artist.

Case study 1

The Next Rembrandt: a portrait of data

Presented in 2016, The Next Rembrandt is a fictional portrait produced as part of a project bringing together ING, Microsoft, TU Delft and partners from the cultural and creative world. Images of paintings attributed to Rembrandt were analyzed in order to identify constants of subject, pose, facial geometry, color and light. The system resulted in a bearded man, wearing a hat, dressed in dark and facing right.

The final image was printed in relief to evoke the irregular surface of a painting. The object is therefore neither Rembrandt oil nor a copy of one painting existing, nor a reliable prediction of what he would have painted. Microsoft cautiously described it as a creative visualization of data, not as a work that Rembrandt would necessarily have done.

Its strength is conceptual: it makes visible the averages and choices of a corpus. His weakness is exactly the same. Rembrandt did not paint a statistical average; he responded to a command, a model, a support, a situation and successive decisions.

Rembrandt, Self-portrait, 1660, Metropolitan Museum of Art — A late original like theSelf-portrait from 1660 carries a material history and decisions that statistical synthesis cannot reproduce.

How does a machine make a "Rembrandt style"?

Create the corpus

Select images and metadata. Quality immediately depends on the attributions chosen, photographs, restorations and the diversity of periods.

Measure regularities

Identify faces, orientation, geometric relationships, areas of light, colors or textures. The model transforms images into numbers and relationships.

Produce a synthesis

Assemble or generate a new configuration consistent with the learned probabilities. What is common in the corpus becomes more likely in the result.

Materialize image

Print or paint the file. In The Next Rembrandt, a relief print sought to evoke impasto without becoming one painting history.

Rembrandt, Self-portrait with disheveled hair, circa 1628, Rijksmuseum — Young Rembrandt employs shadow, incompleteness and irregularity: so many choices that are difficult to reduce to a single signature.

Corpus bias

If poorly attributed works, copies or highly retouched photographs enter the data, the model also learns their characteristics. If the corpus favors portraits from the 1630s, it risks presenting this moment like "the" style of Rembrandt and to minimize Leiden's young painter, the engraver or late paintings.

The history of attribution therefore becomes a technical dependence. A database is never neutral: it reflects a state of research, exclusions, available formats and human choices. The revision of the catalog by the Rembrandt Research Project specifically points out that the corpus boundary has changed a lot.

Case study 2

The Night Watch finds its lost edges

In 1715, The Night Watch was cut to enter a new location at Amsterdam City Hall. The deleted tapes have not been found. A small copy painted by Gerrit Lundens before this cutting, however, retains the complete composition.

As part of Operation Night Watch, the Rijksmuseum used artificial intelligence to bring the copy and the original closer together. The proportions, distortions and pictorial manner of the two objects differ. The algorithm helped transpose Lundens information to the scale and chromatic appearance of Rembrandt's canvas. The reconstructed areas were then printed on panels and temporarily placed around the painting.

This operation does not physically restore lost parts and does not claim to restore every brushstroke of Rembrandt. It offers a documented visualization allowing us to study the initial balance: the main officers appear more off-center, three figures return to the left and the space seems more dynamic.

The Night Watch with the missing parts reconstructed using artificial intelligence — Reconstruction presented in 2021: Additions surround the original and must remain identifiable as restitution.

Rembrandt, The Night Watch, 1642, Rijksmuseum, current state — **Original object preserved:** the canvas of 1642, amputated of strips on its four sides in 1715.

Gerrit Lundens, old copy of La Ronde de nuit before cutting — **Documentary source:** the copy attributed to Gerrit Lundens, which shows the composition before reduction.

Technicians scanning The Night Watch during Operation Night Watch at the Rijksmuseum — Digitization provides extremely detailed data, but interpretation remains conducted by specialists painting.

From image to matter

AI doesn't examine one alone painting

To study an attribution, researchers do not limit themselves to a visible photograph. They combine radiography, infrared reflectography, X-ray macro-fluorescence, microscopy, sample analyses, dendrochronology and study provenance. These data describe depths, chemical elements and work steps that the final image does not show.

Machine learning can help segment these maps, align multiple modalities, or compare hundreds of areas. But causality belongs to interpretation: a common pigment does not designate a hand; a quartz background can report the workshop after 1642 without proving Rembrandt; a shared canvas pattern indicates the same scroll, not necessarily the same painter.

Can AI assign a painting to Rembrandt?

Type of result	Real utility	Essential limit	Human decision required
Similarity score	Identify candidates or anomalies in a large corpus	Visual resemblance ≠ hand identity	Compare time, subject, support and state
Key classification	Measure orientations, textures or rhythms	Restoration, copying and photography modify the signal	Examine the original surface and layers
Pigment detection	Segment MA-XRF cards and reveal composition	Materials were circulating in the workshop	Interpret stratigraphy and context
Canvas correspondence	Automatically compare thread patterns	Same roller does not mean same painter	Reconstruct purchases and timeline
Signature analysis	Compare shapes and location	A copied signature can be convincing	Check its physical integration into the painting
Overall prediction	Synthesize several variables	The score inherits errors and biases from the data	Publish evidence, uncertainties and alternatives

Seven risks to art history and the public

1. False certainty

A percentage like "92% Rembrandt" seems accurate, but says nothing without protocol, corpus, error rate and comparison with the workshop.

2. The circular corpus

The model learns human attributions and then appears to confirm them. It can reproduce consensus without providing independent proof.

3. Style reduced to an average

Singularities, breaks and period changes are smoothed out in favor of the most frequent characteristics.

4. Visual confusion

An attractive reconstruction circulates without a caption and ends up being mistaken for a photograph of the original or a found work.

5. Fake documentary

Images, certificates, stamps, provenances and photographs from synthetic archives can create a coherent but non-existent story.

6. The devaluation of matter

The screen favors pattern. However, attribution also depends on the reverse, the support, the preparation, the cracks and the layers.

7. Opacity

A proprietary model can prevent us from knowing the data, criteria and errors that produced its ranking.

Three close images, three opposing statuses

The Man in the Golden Helmet, Rembrandt's circle, circa 1650 -1660 — **Revised attribution.** *The Man in the Golden Helmet* remember that a work can seem deeply rembranch without being considered an autograph.

Man in oriental costume, Rembrandt workshop, circa 1641 — **Workshop work.** The same materials, models and exercises make hand separation difficult, even with technical data.

20th century copy of a self-portrait by Rembrandt — **Modern copy.** Visual AI can measure strong resemblance while materiality and date immediately exclude the original.

Established facts and interpretations

What the sources establish

The Next Rembrandt is a 2016 data-driven, embossed-printed creation.
Microsoft does not present it as one painting which Rembrandt would necessarily have painted.
The Rijksmuseum reconstructed the lost tapes of The Night Watch from Lundens' copy with the help of AI.
Additions were printed on separate panels, not painted on the original.
Operation Night Watch combines advanced imaging, microscopy and hardware expertise.

What remains is interpretation

The exact fidelity of colors and keys in lost parts.
What Rembrandt would have thought of an image generated in his style.
The ability of a model to sustainably distinguish the teacher from all his students.
The autonomous artistic value of an algorithmic synthesis.
The confidence threshold sufficient to modify a public allocation.

How to read an "AI reconstructed" image

Identify the goal

Is it showing a proven part, testing a hypothesis, virtually restoring a color or creating freely?

Find the sources

What documents, scans and works powered the system? A solid reconstruction allows us to trace the data.

See limits

Are the invented areas reported? Are alternatives, margins of error and human choices explained?

Keep statuses separate

Original, historical copy, reconstruction and generation must remain visually and verbally distinct.

Matter before simulacrum

Explore Rembrandt with clearly named works and attributions

One reproduction hand painted is a contemporary creation based on a historical work. It can convey composition, light and palette, but it does not replace either the original of the museum or its material history. The titles of the shop distinguish Rembrandt, workshop, circle and copy.

Rembrandt Collection The Night Watch

Frequently asked questions

Did artificial intelligence create a real Rembrandt?

No. The Next Rembrandt is a contemporary creation produced from data and printed in relief. It is not a work of Rembrandt.

Has AI restored The Night Watch?

She helped visually reconstruct the lost strips. Additions were printed on separate panels and temporarily placed around the original, without modifying the canvas.

How do we know the lost parts?

Mainly thanks to a small old copy attributed to Gerrit Lundens, made before the cutting in 1715.

Can an AI authenticate a Rembrandt on its own?

No. It can provide comparisons and probabilities, but attribution requires material, historical and stylistic study by several specialists.

Why is a resemblance score insufficient?

The students, copyists and imitators precisely shared the traits that the model learns. A strong resemblance does not automatically distinguish the hands.

Are scientific analyzes artificial intelligence?

Not necessarily. X-rays, infrared, MA-XRF or dendrochronology produce data; AI can help process them, but these techniques exist independently.

What is the main risk for the art market?

Rapid fabrication of consistent fake visuals and fake provenances, coupled with overconfidence in opaque scores.

Should a digital reconstruction be reported?

Yes. The public must be able to immediately distinguish the preserved original, documented areas, assumptions and generated elements.

Can AI help restaurateurs?

Yes, especially for aligning images, segmenting scientific maps, visualizing hypotheses and comparing large amounts of data. The processing decision remains human.

Main sources

The techniques and projects are described according to the institutional sources available at the time of publication. A digital reconstruction is a documented hypothesis, never an original part found.

Account