Intellectual property is the strongest it’s ever been, more or less worldwide.

Intellectual property is the strongest it’s ever been, more or less worldwide. Large, powerful organizations with enormous IP portfolios are active in the machine learning space: IBM has a huge number of active patents, and Google and Apple are both busy stockpiling them from third party sources. Even if learning models are legally determined unpatentable, that doesn’t keep people from attempting to (and often succeeding at) filing patents for them; the nonexistence of such patents doesn’t prevent any company with lawyers and brass balls from attempting to defend them (and any other nebulous or imaginary claims) in civil court.

Even if, somehow, these extremely powerful forces don’t manage to get their way in terms of ensuring learning models are protected by some form of strong IP (copyright for individual models or patent for novel formulations), and somehow the IP litigation system that has for decades been systematically favoring IP holders and ignoring strong fair use cases reverses tack, and somehow these companies forget that they could make a trade secret claim — in other words, even if somehow our dysfunctionally overpowered IP system suddenly started working properly — learning models are hardly the most common forms of potentially protected work, and they are years away from being capable of producing work of equivalent quality to most protected work. In other words, the end of protection for learning models is insignificant compared to the scale of IP.

Theft of learning models, of course, is both trivial and unprovable. A system intended to produce certain outputs for certain inputs can be trained on the same data or can be trained on API calls; as scale the result is the same, but the innards will be uncomparably different even for a very close match in behavior. Much like other behemoths of tech, the factor that would keep competitors out of the race with machine learning based API services is not the (public) concept or the (trivial and novel, mostly off-the-shelf/open-source) implementation but the cost of scaling to meet demand — anybody can write a facebook knock-off in a weekend but only facebook and a few others can afford the server cost to host facebook’s audience. Similarly, anybody can download tensor flow or torch, but few people can afford the cycles to train it on the entire google books corpus and add new books as they are released.

We don’t call facebook knock-offs (even very close ones, like those used for phishing) copyright infringment and consider them subject for suit, even though they definitely are using image assets against TOS; instead, we treat them as either legitimate attempts at competition doomed to failure or as cheap knock-offs indended to trick us. Likewise, trademark law is rarely applied directly against parasitic industries like that of mockbusters — the legal risk of loss of protection is low, and large film companies are mostly fine with allowing the parasites to continue preying on people with poor vision or damaged faculties of judgement who can’t distinguish between “Transformers” and “Transmorphers”; Universal is happy knowing that Asylum will never be able to compete with them head to head, and once the current generation of executives dies off and is replaced, they will treat internet piracy the same way.