How to tell when that nicely phrased but possibly bullshit-laden piece of prose has been generated by artificial intelligence rather than ordinary human intelligence? One possible solution, proposed by seven companies active in the AI space, is digital watermarking.
iStock photo of water with watermarks
According to a fact sheet released on July 21, the Biden Administration has secured voluntary commitments from Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI “to help move toward safe, secure, and transparent development of AI technology” and to alert people to deepfakes.
“It's currently unclear how the watermark will work,” Ars Technica noted in an article about the development, “but it will likely be embedded in the content so that users can trace its origins to the AI tools used to generate it.”
You can read more about how the watermark might in fact work in Mark Liberman’s July 25 Language Log post, which, to be honest, strains my extremely modest technical comprehension. So let’s look instead at that word watermark, which branched away from its original aquatic meaning more than three centuries ago.
In design, a watermark is “a distinguishing mark or design impressed into a sheet of paper during manufacture, typically visible only when the sheet is held up to the light,” according to the OED, whose earliest citation for this sense is from 1708. This type of watermark was “probably so called because the watermark, being less opaque than the rest of the paper, had the appearance of having been produced by the action of water.”
Author’s note: Until I researched this post, I’d assumed that water was somehow involved in creating watermarks. I’d envisioned a die being pressed into a watery slurry as paper was being manufactured. But no!
Watermark became even less literal in the early 1990s, when it began to be modified by digital to signify “an item of code embedded in a digital image, video, or audio file in order to provide copyright information, typically being undetectable during normal use of the file.”
Originally, going back to the late 1500s, watermark had various senses related to actual water. One sense still in use is, per the OED, “The level at which a body of water, such as a river, well, lake, etc., normally stands, or to which the tide usually rises or falls; a line marking this. Also: the height of such a level above mean sea level.” We tend to see this usage in the form high-water mark or low-water mark, which can also be used metaphorically.
There are more than 60 live trademark registrations for Watermark, including those for an investment-management firm, a water-taxi service, a medical-testing facility, and a brewery. There’s also a Watermark condo development in San Francisco, which looks like the perfect place for me. Oh, that pool! But oy, those prices.
AI giveth watermarks, and it also taketh them away. If, say, one wished to remove the “iStock” watermarks from the photo at the top of this post, one could upload the image to WatermarkRemover.io, and hey presto—the watermarks disappear. No charge! This is not entirely ethical, but ethics remains a gray (or maybe grAI) area for AI.
Comments