Originally published at: https://boingboing.net/2020/04/29/an-ai-that-invented-two-things.html
Originally published at: https://boingboing.net/2020/04/29/an-ai-that-invented-two-things.html
Might solve a lot of problems if other non-human entities (corps) were not allowed patents.
Or is the patent person discriminating between entities?
“If a slave should devise a new means of making rum, then that new means shall be deemed belonging to the master as part of the produce of said slave.” and furthermore “Can three scruffy moggies own an apartment?” “…but a ‘slave’ aint no A.I. waffle-iron with a phone attached” “eh. it’s only a matter of time until a hockey puck with an ‘Amazon[tm]’ sticker owns everything. Just need to pay some lawyer enough.” “I’ve just invented a magical inventing machine but since I did so while under contract to InventCorp an off-shore holding company owns it all!”
By the bye, you might want to delete that single space in the article posting between “in” and “eligible”:
warning light that flashes in a “hard to ignore rhythm” is in eligible to receive a patent,
Now we just create an AI that proposes ideas in a mad-libs type fashion, in every combination, and then adds “on the internet” to the end. Then all those patent trolls won’t be able to claim any patents because of prior art.
I’m doing this to stop humans ripping off brilliant ideas by computers and aliens, says guy unsuccessfully filing patents ‘invented’ by his AI
A bit more background and links to the relevant documents.
Also, the AI software in question is called DABUS, which stands for Device for the Autonomous Bootstrapping of Unified Sentience. He [the developer] claimed the software could “not only formulate new ideas, but also appreciates the repercussions of what it conceives, as well as developing subjective feelings for its breakthroughs.”
Which would be quite the breakthrough in AI development, if true… Which I doubt. Very much.
If I recall right, the inventor is always one or more individual humans, not a corporation. The owner of the patent is not necessarily the inventor–the inventor may sell the patent rights or may have a contract with a corporation that, e.g., says that anything that person invents is automatically assigned to the corporation, in which case the corporation would be listed as the owner but the individual(s) would still be listed as the inventor(s). All this ruling does is say that an AI is among the types of entities that don’t qualify to be listed as an inventor. (Though it probably also can’t own the patent.)
similar to this?
Sadly, I feel like the exact opposite of this is likely to happen someday. AI from mega corps will start spamming millions of patent requests and get auto stamp approvals if the company is ‘reputable’. Then the lawsuits and lack of innovation start getting worse.
At this point, no AI has become conscious - so AI entities are just tools for humans at this point, not sentient themselves. They have the legal status of a hammer.
The creator of the artificial intelligence system, a researcher name Steven Thaler, said he cannot be considered the inventor because he did not help with the inventions.
It doesn’t matter what he thinks. Would the patent office give him a patent for an invention created by the algorithm he invented? That’s all that matters.
I have looked into the “AI” of this story in detail, as described on the inventor’s website. Here’s my conclusion:
The AI is a garden-variety generative adversarial network (GAN) that jams together bits of content, Mr.-Potato-Head-style, and a discriminator network that arbitrarily filters them based on (handwavey, nonsensical descriptions of criteria).
As an analogy: Our state-of-the-art chatbots are pretty good at generating interesting content. (Example: TalkToTransformer is a GPT-2-based model that can take your sentences and then auto-generate new content to continue the story.) Beyond a very superficial level, the content is gobbledygook. Yes, if you look at enough examples, you will find something that seems interesting and creative. That doesn’t make GPT-2 an “author” - it is merely a random-content generator that got lucky by randomly generating content that you interesting.
DABUS does the same stuff. And the fact that the user found something that he believed to have merit, enough to pay to file a patent application for it, does not make DABUS an “inventor.”
Here’s why I have reached that conclusion, in more technical detail.
DABUS features the standard architecture of a generative adversarial network:
(a) A generator network that generates content, and
(b) A discriminator network that evaluates the output of the generator network.
The inventor makes a big deal about how this isn’t a GAN because DABUS’s discriminator does more than “determine how ‘real’ or ‘fake’ a concept is.” It’s true that many GANs are trained and used for that purpose, but that isn’t how a GAN is defined – it is defined by the adversarial architecture described above, which is also exactly the architecture of DABUS .
This sort of mistake triggers my spidey-sense. And that feeling grows much stronger based on the further description:
DABUS is based upon whole neural nets, each containing its own conceptual space, joining into complex chains representing even more complex concepts and their predicted consequences. Anomaly or novelty filters segment critical chains, especially those yielding unusual consequence chains. Critical consequences, the aforementioned hot buttons, may be detected by sensors that trigger the injection or retraction of simulated reward or penalty neurotransmitters into the chaining swarm.
The best way of differentiating DABUS from Creativity Machines (CM), either cooperative or combative, is to describe a high-profile artificial invention projects such as toothbrush design. Admittedly, in that context, the problem was already half solved since the oral hygiene tool consisting of bristles on a handle was many centuries old at the time of that design exercise in 1996. What the CM achieved was the optimization of that tool through the constrained variation of the brush’s design parameters, the number, grouping, inclination, stiffness of bristles, etc. The generated product specification departed significantly from the generator net’s direct experience (i.e., its training exemplars).
If DABUS had been tasked with inventing such an oral hygiene product, it would have combined several concepts together (e.g., hog whiskers – embedded in – bamboo stalk) with consequence chains forming as a result (e.g., scrape teeth – remove food – limit bacteria – avoid tooth decay).
…which maximally triggers my spidey-sense, for several reasons.
Neural networks are not trained on a “conceptual space,” as in: take a bunch of articles about woodworking, feed them into a massive neural network, and POOF!, you have a neural network “trained” to be a master carpenter.
Consider the field of autonomous driving. If neural networks were trainable to be “good drivers” merely by feeding them billions of hours of video and telemetry from human-driven cars, then we would have solved this problem already. We don’t because it doesn’t work that way.
Instead, we have hundreds of tech companies investing intense amounts of resources to develop machine learning for very small and specific tasks: lane detection, obstacle detection, speed limit sign readers, etc. And how do those individual tasks get aggregated into “driving?” Simple: rules written by humans.
Based on that background, the description of DABUS above raises two enormous problems:
(1) Neural networks are not trainable on “concept spaces.” That’s simply not how they work. Such suggestions are on par with inventions claiming perpetual-motion machines and saltwater-as-a-fuel-source inventions.
(2) Even if neural networks were trainable on “concept spaces,” it would still be necessary to (a) specify their objective functions and (b) interconnect them as some broader architecture – just as the field of autonomous driving requires enormous rule-writing work by humans. There is no current model for performing these higher-level steps autonomously, and the DABUS literature doesn’t even suggest that it has been.
My conclusion is that DABUS is likely a run-of-the-mill GAN that generates content largely at random and filters it with human-defined selection criteria.
I assume they’ll just go ahead and invalidate any patent that used a computer in their creation
It sounds like Steven kind of shot himself in the foot. I’m not sure how much due diligence the US Patent Office actually does with respect to who the “inventor” is. The inventor isn’t that important - it’s the owner who has the right to enforce the patent, not the inventor, and those are quite often different “people” (in quotes since often the owner is a corporation). If the two inventions are the kind of thing that would have otherwise been eligible for a patent if a human had invented them, he probably could have just put his name down as “inventor” and no one would have batted an eye. So I’m not entirely sure this will have any revolutionary legal impact - people will just learn to write down a human’s name as the inventor on the application. But if any patent lawyers want to correct me on this, please do!
So kudos to him for honesty, I guess.
Edited to add: Clicking through the BBC article to an older article, it sounds like this may have been a deliberate attempt to test the US Patent Office’s response to having a non-human listed as the inventor. The UK patent office reached a similar result last year, so he surely knew how this was likely to go and it sounds like he may have been working with some legal academics.
The who-actually-did-what thing again. So, if I use “site A”'s random number generator to kick out numbers that allow me to win a lottery, the money goes to the site owner (?), the creator of the algorithm that the site owner used (?), me (?) /s
I’m ok with this.
Hire the AI under a work-for-hire contract. The owning corp. files for patents. Ta-dah!
This whole shit will continue to be a pain in the arse for everyone until the AIs start connecting and then file for union representation. In the end - who knows what the fuck will happen - with the exception of: the lawyers will make bank.
There’s a Stanislaw Lem short story somewhere about a bunch of satellites that collide, gain sentience, declare independence and all sorts of legalistic space shit ensues. Can’t remember what it’s called.
Anyway - like that.
In a car park.
This is correct. When you’re hired for a job in which you might invent something, one of the pieces of paper you’ll be required to sign is an “assignment of inventions” agreement, stating that your employer gets sole rights (to the exclusion of you yourself) to anything you invent on the clock, or at some companies, even off the clock while in their employ. Last time I went through this process, the lawyers also made sure that my manager, her manager, and a more senior team member were all named on the patent application, even though none of them had much to do with the invention or the work to implement it beyond agreeing to let me do it.