Art Authentication by Computer?
By William F. Zieske
I saved the Sept. 6, 2008 issue of The Economist for an article that appears in the Technology Quarterly with the tag line “Image-processing software could help to identify artists by their characteristic brushstrokes – and spot forgeries.”
The article, “The computer says no,” points out that scientific processes have been used for years to help authenticate works or art, to identify forgeries and fakes. Now, a team headed by James Wang of Penn State University has been “training” a computer to analyze Van Gogh’s brush strokes using image-analysis software, so that it can determine whether other paintings by the same artist are authentically painted by Vincent’s hand.
I may have a greater distrust of computers than most, but this sounds like lunacy as anything more than an academic exercise. It’s like training fruit flies to perform cabaret.
From an artistic standpoint, it makes little sense. Look at the broad and divergent output by many artists, many moving from one to another aesthetic, or pursuing more than one at a time. Consider Picasso’s blue period and his later cubist works. Rembrandt is known for his wide variety of brush strokes. Compare Gerhard Richter’s Ice series to his nearly photographic but blurred figures.
Add the variety of technical differences – different supports (board, canvas, paper), paints (oil guache, tempera), not to mention other media like crayon, pencil and pastel; the great variety of brush sizes and types (sable, synthetic, combination). Great artists have been itinerant, and used whatever was available. Now add the artist’s hand itself – aged in one work, young and brash in another; or shaking, injured, weak, tired, angry, sad – all in a single work. It can be vigorous one moment and tentative the next, whether intended or not.
So isn’t this a computer application that has strayed too far into the realm of chaos theory? An expert authenticator can learn to recognize not only how an artist preferred to apply oil to a canvas, or tempera to a board, and at different periods of the artist’s life, but even how that artist’s hand treated portraits as opposed to landscapes, how sad subjects might affect the strokes, or how particular elements of a painting are likely to be treated, like Van Gogh’s famous olive trees. Authentication can involve more than facts and their strict application – moving into the realm of forensic psychology of a deceased artist.
Chaos theory is about finding patterns in what appears to have no pattern. When there is limited data to analyze, like stock market performance, it is fascinating to identify as many variables as possible that caused a down-tick in the NASDAQ – perhaps a gloomy commentator on MSNBC, the threat of war, a new technology release, a storm approaching the Atlantic coast, the shortening of days in the autumn. And computers can do much to spot those root causes – or at least to help theorize based on correlations.
But here, not only are the variables that make a brush stroke lie the way it dies on a canvas truly infinite, but so is the output of possible brush strokes. The output is not just a number than can go up or down; it is one of an infinite number of brush strokes that can be achieved by the artist.
Both artistically and scientifically, I am more than skeptical of this project. What about legally? That’s something I’m actually qualified to talk about. It is legal process that will often decide, if disputed to the end, whether a particular work of art will be treated as authentic or not. Whether it is litigated against the expert compiling the catalogue raisonnée who rejects the painting, against an insurer for the value of the painting, or against the last seller for misrepresenting the painting – if the value of attribution and authenticity is high enough, it will be decided in court.
Legally, the prospect of computer software being used to authenticate and attribute paintings is frightening. Would paintings with a solid provenance or otherwise well-established as authentic be rejected as frauds, or their value questioned, simply because the computer said so? Certainly, if such a process gained favor, newly discovered works, mainly by prolific artists, would be subjected to computer analysis, and there soon would be attempts to use computer brush stroke analysis before a jury, either as evidence or as support for an expert’s opinion.
Courts have struggled with how to let in good scientific evidence and testimony, while keeping out “junk science.” The major change in this process came in Daubert v. Merrell Dow Pharmaceuticals, 509 U.S. 579 (1993), when the U.S. Supreme Court raised the bar for such evidence. The Daubert decision holds that federal courts must take the role of “gatekeeper” over expert and scientific evidence in each case, emphasizing empirical indicia of reliability and the scientific method as the bases for admitting relevant expert testimony.
That ruling has raised the hopes of criminal defense attorneys, and increased attacks upon the scientific basis of many time-honored forensic techniques. Even the theory and process of fingerprint identification are publicly questioned today, nearly a century since the technique first put someone behind bars.
The prospect of a computer’s “opinion” being presented in court on a topic so ill-suited to data processing – regardless of the amount of “training” embodied in the software, and however sophisticated the algorithms employed – is chilling. Use of computers as “experts” in this way would take advantage of a common but incorrect belief, held by many, that computers are more impartial and accurate than any person can be, regardless of the task assigned. Overcoming this prejudice in a courtroom can be extremely difficult and expensive. It would require defensive argument that the computer is only as good as the information it is fed, perhaps with the support of experts who could analyze the hardware and software’s limitations and communicate them effectively to a jury. In the end, a computer’s rejection of a painting based on analysis of brush strokes will tend to sway many juror’s decisions simply because it is a computer-generated result, regardless of the computer’s limitations and the reliability of countervailing evidence.
Luckily, this scenario is unlikely to play out in 2009. The engineers behind this admit the system has a long way to go: in the trial run reported in Signal Processing Magazine, the computer identified just 4 out of 6 of the known fakes, and gave the thumbs-down to two known authentic Van Gogh works.
