Catch up on stories from the past week (and beyond) at the Slashdot story archive

 



Forgot your password?
typodupeerror
×
AI

AI Study Suggests a London Gallery's Been Exhibiting a Fake For Years (thenextweb.com) 121

Thomas Macaulay writes via The Next Web: Samson and Delilah is among the most famous works by Peter Paul Rubens, one of the most influential artists of the 17th century. The painting depicts an Old Testament story in which the warrior Samson is betrayed by his lover Delilah. When London's National Gallery bought the masterpiece in 1980, it became the third most expensive artwork (PDF) ever purchased at auction. But the buyers may now be searching for their receipt. According to a new AI analysis, their prized possession is almost certainly a fake.

The tests were conducted by Art Recognition, a Swiss company that uses algorithms to authenticate artworks. The firm's tool is based on a deep convolutional neuronal network. The system learns to identify an artist's characteristics by training the algorithm on images of their real works. The training dataset is then augmented by splitting the images into smaller patches, which are zoomed into to capture the finer details. Once the training is complete, the algorithm is fed a new image to assess. It then analyzes the picture's features to evaluate the likelihood of it being genuine. After comparing Samson and Delilah with 148 genuine Rubens paintings, the system gave the artwork a 91% probability of being inauthentic.
Carina Popovici, the cofounder of Art Recognition, was shocked by the results: "We repeated the experiments to be really sure that we were not making a mistake, and the result was always the same. Every patch, every single square, came out as fake, with more than 90% probability."
This discussion has been archived. No new comments can be posted.

AI Study Suggests a London Gallery's Been Exhibiting a Fake For Years

Comments Filter:
  • Training (Score:4, Interesting)

    by Bert64 ( 520050 ) <bert@slashdot.fir e n z e e . c om> on Wednesday September 29, 2021 @02:10AM (#61843403) Homepage

    Clearly many people genuinely believed the painting hanging in the museum is authentic and it's likely the museum and the auction house would both have had art experts on hand especially with so much money at stake, so who's to say that the other 148 works used for training are authentic?

    • Re:Training (Score:5, Informative)

      by phantomfive ( 622387 ) on Wednesday September 29, 2021 @02:53AM (#61843459) Journal

      Many experts also believed that the Samson & Delilah painting hanging in the museum was fake. It is a paintings of disputed origin, that astonishingly resurfaced in the early 1900s, with the ones who bought it most vociferously defending it as authentic.

    • The fake may be worth more than the real one. Now, they could sell a NFT of the painting for hundreds of millions of Pounds to recover their perceived loss and profit off it.
    • Uhh, I thought the London one was known to be by an apprentice of Rubens? That is, its provenance goes back to someone who worked with Rubens, but not to Rubens himself.
    • This is a painting with a questionable history and many experts have claimed it is a fake ever since it miraculously reappeared, so despite the way the summary frames this as shocking, it is actually not an unexpected result at all. Most of the other 148 paints can be traced right back to Rubens, this one can only be traced back to the early 1900's.
  • So? (Score:5, Funny)

    by backslashdot ( 95548 ) on Wednesday September 29, 2021 @02:44AM (#61843439)

    Maybe this was the painters masterpiece, his magnum opus as it were. It might be radically different than the rest. I mean look at Metallica all their albums since And Justice For All have been crap. Any AI would proclaim them fake.

    • Re:So? (Score:5, Interesting)

      by Anonymous Coward on Wednesday September 29, 2021 @03:00AM (#61843469)

      Well, the original was painted in 1608-1609. That's about 1/4th through Rubens' career. It was not considered his "magnum opus" in any way at the time and we do have an engraving of the original and a painting of the room the original was hanging in at the time that indicate the original depicted Sampson's feet, which are cropped in this probable forgery.

    • Maybe this was the painters masterpiece, his magnum opus as it were. It might be radically different than the rest. I mean look at Metallica all their albums since And Justice For All have been crap.

      There haven't been any albums since AJFA have there? Metallica only made four albums and then disbanded.

      Unless you're talking about that covers band Tinfoilica, heard they were crap.

      • by Hodr ( 219920 )

        They disbanded, but then formed an IP holding company in order to sue anyone and everyone that listened to their music from then on.

    • by Orlando ( 12257 )

      Training AI on all their albums since And Justice For All would result in the earlier stuff being deemed as fake. Maybe the crap albums are the real Metallica?

    • It's true that people can produce works that are radically different from one another, but even in such cases the minute details are likely to have deep similarities. Hence even when people fake their handwriting, small characteristics are likely to remain from their natural handwriting. Since our muscles are accustomed to moving in particular patterns, our works exhibit such patterns in varied combinations, so that a peculiar style comes through even despite conscious efforts to cover it up. Since this AI

      • Handwriting analysis (graphology) is generally considered a pseudoscience or scientifically questionable practice. Researchers have found that experts are marginally better than novices at estimating how often specific handwriting features occur in the writing of the general population, but they are not able to do so with complete accuracy.

        • A friend from student days - she did psychology - gets regular ear ache from graphology companies trying to sell her their product for assessing job applicants. Occasionally, when things are quiet, and she is feeling cruel, she'll punish a sales thing by asking for their evidence on reproducibility, predictive accuracy, etc. It's more fun than pulling the wings off flies, and ethically more defensible.

          Her psychology course required a module in Statistics - to the utter horror of many of the students who di

      • Hence even when people fake their handwriting, small characteristics are likely to remain from their natural handwriting.

        I should get back into practice in writing with my other hand. Some years ago I taught myself to write adequately with the other hand to the one I first learned to write with. I also, for completeness, learned to write in the other 6 possible directions (viz : both hands, right to left and left to right, upside down and down-side-down).

        Studying crystallography does things like that to yo

    • Love the Metallica analogy and whole heartily agree.
    • You're being moderated as Funny, but you make a really good point. Spot on about Metallica. I have been listening to their self-titled "black" album a lot lately as well as to AJFA and you would definitely not think that was the same band! I'd moderate you Insightful, but this needed a comment.

    • From the National Gallery page "During a visit to Italy, Rubens had seen Caravaggioâ(TM)s experiments in the use of highly contrasting light and shade, and deep, rich colour. On his return, he used these new techniques to paint Samson and Delilah" https://www.nationalgallery.or... [nationalgallery.org.uk]

      • But ... but ...

        I thought artists sprang, fully formed, from the forehead of their mentors after being struck by an axe. After that, no development or change, ever.

        Artists learning from each other? Who'da thunk it?

  • Why? (Score:5, Insightful)

    by enriquevagu ( 1026480 ) on Wednesday September 29, 2021 @02:49AM (#61843447)

    The funny part of using AI to determine authenticity is that the neural network claims it is fake, but you don't know why. Can you debug this program, to make sure it was not a mistake? Or is it just: "the coefficients say so, do not touch my coefficients!"?

    • Re:Why? (Score:5, Interesting)

      by Tom ( 822 ) on Wednesday September 29, 2021 @06:05AM (#61843701) Homepage Journal

      While there is an active research field into explainable AI, most systems today essentially are what you describe above, yes.

      Even if we get into explanations and dig into the machines, the fact that they essentially gain their data by training and large data sets means that their understanding of things is much closer to our intuitive understanding of the world. Like when you see a dog you know that it's a dog - but if I pressed you to explain why, exactly, you think that's a dog, you would actually make up an explanation that has little or nothing to do with how the intuitive part of your brain actually performed its object recognition. There's been a lot of science into these things for decades now. It's really interesting.

      • I'm always surprised with what passes as a dog nowdays. A lot of people's pets out there mess up my brain's object recognition function. But hey, as long as they can hump each other and produce fertile offspring they're all dogs, I guess. Maybe I'm just hard to convince.

        • by dryeo ( 100693 )

          They can't all hump each other and produce offspring, consider size, a Saint Bernard and Chihuahua for example. Meanwhile a dog and coyote or wolf can hump each other and produce fertile offspring. Species is a fluid concept

          • by cusco ( 717999 )

            If you ever read 'On the Origin of Species' you'll find that Darwin came to much the same conclusion, that the definition of species will vary depending on the aspects being examined. Most people are taught in grade school that if animals are mutually fertile and can have fertile offspring then they're the same species, mostly because teachers don't have time to teach the nuances. Someone here one time said, "Ask 5 biologists what constitutes a species and you'll get 7 different definitions."

            • Someone here one time said, "Ask 5 biologists what constitutes a species and you'll get 7 different definitions."

              As few as that? That's a remarkably consistent group of biologists. Are some of them clones of each other, raised as identical twins in identical environments?

          • by MobyDisk ( 75490 )

            "Species: A group of closely related organisms that are very similar to each other and are usually capable of interbreeding and producing fertile offspring." Different sources alternate the order of this definition, versus the definition based on characteristics. Apparently the split of Dogs and Wolves as different species is controversial for this reason.

      • Except most people look at a hyena & say it's a dog when in fact it's a cat.
        • by q_e_t ( 5104099 )
          This is no laughing matter.
        • There's a more interesting characteristic of hyenas - most people who see a dominant female hyena misidentify it as a male. (One of the ancient proto-biologists - Arisotle, perhaps - made the mistake of thinking that hyenas only had males.)
      • Like when you see a dog you know that it's a dog - but if I pressed you to explain why, exactly, you think that's a dog,

        The brain has internal models of its beliefs, and places everything in a sort of multidimensional space where "description" acts as a sort of length.

        This, when you see an image and are asked "is this a dog or a cat", the response is usually "it looks closer to a dog". That word "closer" is a hint to how the brain processes recognition: it has models of both "cat" and "dog", and notes that the differences between the image and its model of "dog" are smaller than the differences between the image and its mode

      • While there is an active research field into explainable AI, most systems today essentially are what you describe above, yes.

        Even if we get into explanations and dig into the machines, the fact that they essentially gain their data by training and large data sets means that their understanding of things is much closer to our intuitive understanding of the world. Like when you see a dog you know that it's a dog - but if I pressed you to explain why, exactly, you think that's a dog, you would actually make up an explanation that has little or nothing to do with how the intuitive part of your brain actually performed its object recognition. There's been a lot of science into these things for decades now. It's really interesting.

        True, though if I disagree with your classification of something as a dog we can start discussion our explanations and refining our classifications. For instance, is it because of the shape of the snout? Perhaps I know another non-dog animal with a similar looking snout. Perhaps you noticed a bowl labelled 'Fido' or an owner that seems to be a stereotypical 'dog person'. Our initial assessment may be just as vaguely reasoned as the AI, but humans are able to iterate through these counterfactuals in the way

    • Re: Why? (Score:4, Insightful)

      by argStyopa ( 232550 ) on Wednesday September 29, 2021 @06:58AM (#61843777) Journal

      Not to mention, the use of phrases like "deep convolutional neuronal network" are (or should be, at least) the opposite of convincing.

    • by kenh ( 9056 )

      Maybe they confuse consistency with correctness?They fed the same AI the same reference material and the same painting and got the same results - that's how computers wrk.

  • This probably means that someone else has the original. I wonder what they're thinking right now...

  • by votsalo ( 5723036 ) on Wednesday September 29, 2021 @02:50AM (#61843453)
    The article mentions that "the painting’s authorship has been disputed for decades." Scholars have argued that the style is more heavy-handed than any other works by Rubens. They point to Samson’s cropped toes as a clear sign that the piece wasn’t painted by the Flemish master. The scholars explain the reason for their opinion, but the algorithm doesn't. The algorithm just gives a 90% probability, that "every patch, every single square came out as fake", without explaining anything. So why is the algorithm more authoritative than people?
    • by Richard Kirk ( 535523 ) on Wednesday September 29, 2021 @04:49AM (#61843611)

      AI is a new tool. I am working on AI with image processing. Sometimes it works, sometimes it doesn't. So, do not say "This must be a fake, because Computers plus Science!'. However, it is very hard to quantify a general hunch that "It sorta doesn't look like most other Rubens stuff", particularly when other experts disagree. But if an AI can reproduce this feeling of a hunch from a scan of any small region, then it isn't just "the arm looks wrong" which may mean something was repositioned late in the process, but something generally wrong with every small bit of it.

      The Van Meegren fakes of Vermeer may have had his style, but X-Ray studies showed he used titanium white instead of lead and zinc white, and stuff like that. The AI evidence is much less clear-cut than that. Presumably all the materials are kosher with this one. But there is something rum about it.

      Disclaimer: I am not a Rubens fan. The picture is what it is. The real concern is the monster price tag because it is said to be by Rubens.

      • The Van Meegren fakes of Vermeer may have had his style, but X-Ray studies showed he used titanium white instead of lead and zinc white, and stuff like that.

        That is the kind of scientific evidence I would believe. Apart from anything else, it is basically chemistry, and not image processing.

        Image processing algorithms, designed to recognise and distinguish between different shapes, are notoriously unreliable. A face recognition algorithm that can't tell the difference between a black person and a Gorilla does not instil confidence in the technology.

        • "...he used titanium white instead of lead and zinc white"

          "That is the kind of scientific evidence I would believe."

          I believe its been shown that there are 2 different materials used for each white, but could he have possibly been out of titanium white and used an alternative due to supply issues? Even in the 21st century, we can't even seem to have a steady supply of toilet paper or convince 30% of the population to take a vaccine.

          • I believe its been shown that there are 2 different materials used for each white, but could he have possibly been out of titanium white and used an alternative due to supply issues?

            In a manner of speaking. The point of detecting forgeries by the materials used is the time when each material became usable is known. The original of this painting was painted in 1608. Titanium wasn't chemically isolated until 1791 and titanium dioxide for use in pigments wasn't mass produced until 1916. Titanium white requires knowing the element exists and separating it from its usual ores in a chlorine process which was unknown in the 1600s.

            Modern pigments literally didn't exist until modern chemist

            • by cusco ( 717999 )

              *Trivia Alert*

              The Romans also used lead to sweeten cheap wine for consumption by the commoners, since it was cheaper than sugar (an expensive luxury good until the industrialization of the industry in the 18th century). Wine was consumed from childhood to old age, the lead content in some of the corpses is truly astronomical.

          • but could he have possibly been out of titanium white and used an alternative due to supply issues?

            If you include "lack of a time machine" as part of "supply issues", then yes.

            Vermeer died in 1675. Titanium white was developed as a pigment in the early years of the 20th century, and requires considerable chemical processing to produce a white pigment from the ores. Most titanium-rich minerals are quite dark in colour (greys, near black, bright bronzy yellow), which would mean they'd be picked out in the p

    • So why is the algorithm more authoritative than people?

      Is this a trick question? Firstly no one has claimed the algorithm is authoritative. Secondly algorithms do show their working. Think of it as the difference between two scientists debating whether climate change is real, and an actual model showing a link to rising CO2.

      AI is just a tool in a toolbox.

    • by azcoyote ( 1101073 ) on Wednesday September 29, 2021 @05:50AM (#61843685)

      So why is the algorithm more authoritative than people?

      Haha, good point. For genuine experts the AI can really only be a tool for assisting more qualitative analysis, akin to a polygraph in a criminal investigation. The "91%" implies that there's still a 9% chance that the algorithm is wrong. But that's really only a numerical representation; it's not even fair to say, "If we had 100 such paintings, 9 of them might turn out to be authentic." If 100 such paintings existed, it would outweigh the input data set so much that the same percentage would no longer apply. So it's more like when Windows says it's "91%" done with a file transfer. Amazingly, that last 9% can sometimes take as long as the first 90%. The number is there more as an interpretive tool for the user ("I shouldn't cancel, it's almost done!") than a real, quantitative representation of the progress.

      The reason for having an output percentage is because in some way or another the AI must reduce the painting to quantifiable information, and the best way to correlate such quantifiable information back into the overarching qualitative analysis of the work is to use a probability, even one which is more representative (this 91%) than real (e.g. flipping a coin and getting heads will happen about 50% of the time).

      Still, your point is very valid precisely because the general public, led by sensational journalism, is likely to think that the AI alone is more definitive than the qualitative analysis of art experts. The flow of authority today is very peculiar. I always try to bring this to the fore when teaching people so they see how society grants authority. There is the "cult of the experts" on the one hand; we are a society that does not readily trust religion or government, but we trust people with letters after their name. Books often highly a physician's name proudly--even if he or she had little to do with the book's contents--because the idea of a physician sells. Even to sell a crackpot scheme or conspiracy theory, one needs "science" to back it up. Thus anti-vaxxers have their own "experts" who promote their anti-scientific views using their own brand of "science." Yes it's often superstitious, but it isn't usually on the authority of visions or revelations that they believe in their views, but rather on the authority of so-called experts.

      On the other hand, computers and AI have their own related authority. I received a mailer recently advertising real estate, and it shows an iPhone with glowing reviews on its screen. Having a phone there implies that the business is high-tech and modern, and filling its screen with reviews implies their authenticity. Of course it's just a mailer, but it works because the iPhone symbolizes authority. Many people today are more likely to trust the outrageous advice they discover via their phones than the common sense advice given to them by parents, teachers, priests, or other traditional authority figures.

      • the general public, led by sensational journalism, is likely to think that the AI alone is more definitive than the qualitative analysis of art experts

        In a case where art experts are split on something, I know that at least the AI won't be swayed by an emotional desire for one answer or the other.

        • by cusco ( 717999 )

          Or financial interest. Some of the "experts" are seen more as experts in inventing reasons why a work should be considered authentic rather than as actual critics interested in authenticity. Mercenaries of the art world, I suppose.

      • There will always be experts. You have to trust what experts tell you, because you don't have the means to find out things for yourself. But when there is a multiplicity of opinions, only some of which are based on knowledge, which expert do you trust? At present, there seems to be a mistrust of "official" experts, and more trust put in way out loony theories. Why is that? Perhaps it is a mistrust of all officialdom. There are plenty of reasons why people should mistrust officials. Politicians are mostly ly

    • So why is the algorithm more authoritative than people?

      I don't think they consider it authoritative - however they do consider it neutral. Most critics who feel one way or another either have a stake in the matter, or have strong personal feelings on the subject. For example, the pedigree of this painting is a bit shady, and experts know that pedigree, thus it would bias their opinion regardless of the qualities of the painting itself.

      This AI analyzes only the physical, factual characteristics of the painting. I presume that things like color choices, brush s

    • Is anyone accounting for the possibility that the painter was injured in some way while painting it? "Heavy handed" could be caused by numerous injuries. Who's to say the painter didn't say fuck it, I'm gonna go at this all as fuck heavy with the paint? This algorithm they're running these through is very interesting but I think they need to use it as a tool for suspicion and not an authoritative tool for determination. I'd be surprised if we can bring AI past that hump any time in the next 20 or 30 years.
    • This raises an interesting point: do people trust a computer algorithm more than human art experts? If so, we are in the shit. There is far too much trust put in "what the computer says". If all the AI was doing was looking for inconsistencies in the style of the painting, compared to Rubens' other works, then all I can say to that is that maybe Rubens was experimenting a bit. Artists do that, you know.

      The opacity of the algorithm verdict is also worrying. You can ask a human expert "why do you think this i

  • by Ecuador ( 740021 ) on Wednesday September 29, 2021 @02:57AM (#61843463) Homepage

    Reading the Guardian article, I don't see how anyone could have been "shocked" by the results, the evidence was pretty damning before. It's a painting that was lost in 1640 and suddenly appeared in 1929, authenticated by someone who was proven to authenticate fakes for money. Its colours and strokes different from other Rubens works (when we know it was painted reasonably early, but definitely not at the start of his career), but most damning, the painting appears on 2 copies that we know are from the original's era and the National Gallery's painting is CROPPED compared to those (one an engraving, the other a painting of the room where the original owner had it hanging).
    It seems that the National Gallery made a bad decision purchasing this for millions back in '83 and they have been doubling down on it since just to protect their investment, even though the evidence against it being a Rubens being pretty damning.

    • lost in 1640 and suddenly appeared in 1929

      Radiocarbon dating should be able to determine the difference between these two dates.

      There are ways to spoof radiocarbon dating, but nobody would have used those methods in 1929.

      • by jeremyp ( 130771 )

        My quick Google suggests C14 dating only works for samples more than about 300 years old. If that's the case, I doubt if it could provide a conclusive result for this paining and, even if it could, there's no reason why it couldn't date from around 1640 and still be a fake. It could be a copy made in the seventeenth century or it could be a modern forgery painted on seventeenth century canvas.

      • Forgers have been re-using old canvases for a lot longer than there has been C-14 dating for them to worry about. The size and twist of the fibres, the composition of the fibres, the pattern of the weaving have been known to vary since the first weaver put thread onto the first loom. Generally - and it is a generalisation - a painter with a significant output (it's a business, after all) finds a local weaver who produces a fabric that he likes, and buys from that source until forced to waste time finding an
    • That was a nice summary of what I would call proper forensic evidence. I am not sure this AI "evidence" really adds to that.

      I do have to ask a question, though. Is the fake actually any good as art?

  • If their database is large enough and the actual author made more paintings that are still around today, their AI should not only be able to tell who probably didn't paint it but also who probably did.

  • "it became the third most expensive artwork (PDF)" - Rubens released his work as PDFs?

  • that "experts" are fooled by fakes... Here's one that I find funny

    https://archive.archaeology.or... [archaeology.org]

    because even as a kid when I saw the pictures of those statues among the rest of the Etruscan items something was "off".

    And what about Piltdown Man?

  • Wait a minute ! In other news we learn that AI can produce art including paintings. Now AI says a painting is fake. Is there any conflict of interest here ? I think an AI figured out a way to make nice money here by creating new art and devaluating existing art.
  • I have some empty space on my lounge wall ...

  • by Bruce66423 ( 1678196 ) on Wednesday September 29, 2021 @07:04AM (#61843783)

    I am sceptical of the entire art market; the idea that a poor example of a very famous painter's work is worth far more than a solid work by someone less well known is deluded. The same applies in SF material; I made the mistake buying an early Robert Silverberg novel - and was wildly unimpressed. Similarly crawling through the detritus of Asimov's poor material is uninspiring.

    So it would be good to see galleries sticking to 'the good stuff' regardless of it origins, rather than fixating on the precise history. Or am I missing something?

    • Painting a copy of another painting is relatively easy compared to creating the original. It is already known that there is an original, so if it is a fake it would also be a copy.

    • by AmiMoJo ( 196126 )

      Art is valued more for the story behind it than the quality of the work. That's simple supply and demand, there is lots of good quality art and you can pay to have it created to your exact tastes, but the amount of centuries old art with and interesting history is quite limited.

  • I know we all like to informally think of a 0-1 normalized regression as a probability. But stop selling it in news and such, because people already have a hard time understanding statistics without media calling a probability a quantity that it is clearly not. Obtaining a probability density distribution requires much more than adjusting a certain well normalized function between 0 and 1. Commodity Neural Networs DO NOT spit out probabilities.
  • Without knowing the correlation of in authenticity probabilities for known authentic outliers saying something is true for all the patches is meaningless.

    Train a fucking pooling layer.

  • by WierdUncle ( 6807634 ) on Wednesday September 29, 2021 @09:31AM (#61844085)

    I can't resist this painting:

    Knickers [wikipedia.org]

    Rubens was of course a genuine artist, and not just taking the piss, like modern artists seem to like doing. Taking the piss might possibly be exemplified by this piece:

    Fountain [wikipedia.org]

    The trouble, it is now quite an old joke (1917), so I am not sure I would call it modern any more.

    Back on topic, there is a point about the intrinsic worth of a work of art, as art, versus its worth as an item in financial speculation. I would suggest that the fake Rubens has passed the intrinsic worth test for many years, but has pissed of the art dealing speculators, and I don't mind that at all.

  • Any significant museum owns a fake or ten.

    Sculptures in particular are known to be hard to verify. Paintings are better. Spectrometers can reveal what type of paint was used, radioactive dating can give the age of the organic materials used. Not so for stone. You can go to Italy, find the same quarry that Michelango used, carve up a fake and no one can say for sure.

    A good rule of thumb is the more info they have about a sculpture, the more likely it is to be real. I.E. "X bust, done by Mr. Y, on Z da

  • I remember learning in my college art history classes that it was common for artists to take on students that would produce paintings using their style and then still take credit for the work.

  • They bought the piece for 2.5 million pounds in 1980.

    They can make back ten times that amount selling a scan of the image as an NFT. They'll probably make yet more pointing out it's an original of a scam.

    Then it will be art again, instead of just a beautiful painting of great historical value. (/s)

  • investigators will not be able to tell why the AI flagged it. My point is, neurally-trained AI is basically a black box. With a 3 or 4GL at least you could "psychoanalyze" the source and determine the cause of rejection. Let's hope this kind of AI is never used in capital cases.

The relative importance of files depends on their cost in terms of the human effort needed to regenerate them. -- T.A. Dolotta

Working...