manifesto

definition a quality integral to an artwork that cannot be communicated through mechanical reproduction techniques [source ↗]

In 1936, Walter Benjamin argued that even a perfect reproduction of an artwork misses one thing: its presence in time and space, “its unique existence at the place where it happens to be.” He called that quality aura, and he warned that aura is what withers when art can be mechanically reproduced.

We take Benjamin literally. Aura lives in the unique existence in time and space of the act of making. AURA exists to capture that act, on the artist’s terms, so it can stand as evidence of authorship when the finished piece alone can no longer tell us.

We are building a space in which a human artist and a robot arm paint together, and in which the artist’s process is recorded as a creative signature: gaze, motion, and physiology from wearable sensors, alongside the language the artist uses to direct and correct the robot.

In control theory, a system is identified by applying a known input and observing the response. This response, not the resting state, is what categorizes it. The artist in flow is our system at rest, the robot serves as our known input. It disturbs the robot’s attention, rhythm, and space in a controlled way, and we aim to measure how the artist recovers.

Authentication has always rested on three pillars: connoisseurship, provenance, and scientific analysis of the piece. These all examine the piece after it exists.

AURA moves authentication upstream. We call this process provenance: a record that begins before the piece and travels with it.

The need is not hypothetical. Generative tools already produce images that both use existing human artwork are hard to distinguish from human work, and the same capability is feeding fraud and misattribution at an alarming scale.

1. 01

   An authenticity dataset of process signals, collected with artists as collaborators, documented to the standard of a scientific instrument.

2. 02

   A verification protocol that returns a probability with an explanation, never a verdict.

   Our outputs are advisory evidence to be weighed alongside expert judgement, provenance, and material analysis, not a substitute for any of them.

3. 03

   An authenticity facts label: what was captured, when, with which model version, how much of the work is human and how much is synthetic.

4. 04

   Open tools that artists, estates, and catalogue raisonné can run without us.

We are not building a replacement for artists. We are building a way for artists to prove they are themselves.

• The robot never signs.

  Authorship belongs to the human whose signature was captured.

• Recognition, not generation.

  The creative signature dataset is precisely the corpus a forger or style-cloning model would want. We will never use it to train a system that reproduces an artist’s manner, only one that recognizes it.

• Absence is not evidence.

  No artist is required to enroll, and an artwork without a creative signature is not less authentic.

• No verdicts.

  We do not certify, rather we estimate, explain, and publish our error rates.

• Artists govern.

  Prior painting robots were built without artists and then pointed at them. AURA’s design consent terms, and release decisions include the artists whose work makes the research possible.

AURA is funded by the Defense Advanced Research Projects Agency through Young Faculty Award #D24AP00323-00 to the University of Michigan. We state this plainly because an artist deciding whether to trust us with their body's data deserves to know who pays for the work.

The sponsor's interest is stated in our proposal and is the same as ours: understanding how authenticity can be traced when humans and machines create together. There are no classified deliverables. Every method will be published. The sponsor receives aggregate findings and open code; it never receives raw biometric records of identifiable artists. The views in this document are the authors' and do not represent the official policies of DARPA or the U.S. Government.

Biometric data is the most intimate category of data we could ask for. We commit to the following:

• The artist owns their creative signature.

  The University holds it as custodian under license, not as owner.

• Consent is specific and revocable.

  Enrollment covers this study. Any new use requires new consent. Artists can withdraw, and withdrawal includes deletion of derived models wherever technically feasible.

• Identity and signal are kept apart.

  Data is pseudonymized at capture. The link between signature and legal name is stored separately from the dataset.

• We minimize.

  Where the science allows, we keep features rather than raw video and audio.

• Nothing is sold.

  Nothing is shared without a data-use agreement that inherits these terms.

• All work proceeds under Institutional Review Board oversight, and the consent form is written to be understood by an artist.

AURA is non-commercial.

Our tools are released as open source. We do not sell verification, we do not sell access, and we will never sell or license biometric data or a creative signature for any purpose, generative or otherwise.

If a verification service built on this work ever exists, it will not be ours, and any such service must inherit two rules: artists never pay to be verified, and no query runs against an artist's signature without that artist's consent.