IBM has won more patents over the last two decades than any other organization thanks to a prolific research apparatus that takes the credit for such notable inventions as the HDD, barcodes and the Watson supercomputer. But not all of that intellectual property originates from its product development efforts. A large portion is the fruit of a much less frequently mentioned aspect of the company’s competitive strategy that has reared its ugly head yet again.
Buried about halfway into the latest batch of its patent filings to have reached their publication date is an application seeking the rights for the use of machine learning in filtering program input, something that would hardly merit a second glance under normal circumstances. The specification is meant to be implemented as a server-side process that can enable the host machine to determine if ingesting a particular piece of data might cause technical issues and automatically decide whether or not to process it accordingly.
It’s difficult to fault IBM for trying to reduce the risk of program failures faced by its customers, which are typically large companies where even a minor outage can cause a significant loss of productivity for the employees who depend on the service. The problem lies in the fact that the specification doesn’t limit itself to any specific kind of failure. Understanding why that’s dangerous requires looking no further than the introductory summary of the filing, which lists unauthorized access prevention among the applicable use cases.
Practically every corporate security solution developed in the last few years employs some sort of behavioral analysis capability to distinguish malicious activity from everyday access requests. Much like a traditional antivirus is periodically updated with new virus signatures, such behavioral analytics engines continuously incorporate important new observations into their models to improve their accuracy. In other words, machine learning.
And if you go by the dictionary definition of failure as a state of not meeting a desirable goal, then even a registration form refusing a username because of a non-alphanumeric character would potentially constitute a violation of the patent should it be approved. The possibility of that happening is quite high given that IBM is framing the request in the context of the otherwise entirely legitimate goal of reducing server downtime.
A small consolation is the fact that retrospectively patenting a technology already in existence is grounds for dismissal even if the filing is accepted. That means that with enough time and lawyers, a company should that finds itself staring down IBM’s machine learning barrel should theoretically be able to avoid the fate of Twitter, which was forced to enter a cross-licensing agreement with Big Blue after violating a nearly 30-year-old advertising patent last year.