IBM – Testing Turing

Reply to this post

Removal of bias essential for commercialisation.

• IBM’s debating AI is not going to fool anyone, but it does represent a step forward from Jeopardy, reinforcing RFM’s position that in the enterprise, IBM has a pretty strong AI offering.
• However, ensuring that the opinions are created without bias will be crucial to ensuring that the AI can be used in a commercial setting.
• IBM’s latest Watson showcase is called IBM Debater upon which development was begun in 2012.
• The objective of the technology is not to build a debater but to build an AI that can scan millions of articles and online resources and put together an argument based on the facts and arguments that it finds.
• At this stage, the field of subjects is limited to around 100 and IBM estimates that IBM Debater is only be able to have a coherent debate around 40% of the time.
• However, in a controlled environment, where IBM picked the topic, IBM Debater was able to beat a seasoned student debater on the question of whether or not space exploration should be subsidised.
• IBM had time to prepare the AI beforehand and I suspect that if the human had deviated materially from the well known arguments on this topic, the AI would have got into great difficulty.
• Furthermore, it is worth taking into account that this AI is not a single entity but an amalgamation of a series of AIs each trained to do a specific job.
• This is a symptom of one of the big limitations of AI which is that algorithms can only perform one task and there is no ability to take learnings from one task and apply it to another.
• This is one of the fundamental limitations of AI today and one of the three goals that RFM looks for when assessing whether companies are really addressing AI or simply using statistics (see here).
• Hence, because IBM Debater has not made a significant advance against any of the three goals of AI (see here) or Moravec’s paradox, this is not a breakthrough in my opinion.
• However, what it does demonstrate is a unique and highly effective tool to make the process of background research simpler and easier.
• Critically, there is also the issue of bias, as the AI is creating opinions that are based on the views of its trainers as well material that humans have created.
• All humans demonstrate bias and removing this from the AI is going to be extremely difficult as AIs are only as good as the data set upon which they have been trained.
• To be useful, IBM debater needs to be entirely objective which is reasonably straight forward when it comes to undisputed facts but quickly runs into trouble when things are less straight forward.
• I think that the end result is an advance in language processing and understanding as well as a big advance in getting different AIs to work together to produce a meaningful result.
• Hence, I can see this becoming a very useful part of IBM’s offering to the enterprise, but it will have to be 100% certain that the AI is completely balanced and objective.
• In the grand scheme of things, IBM is still behind Google, Baidu and Yandex when it comes to AI in general but its expertise in the enterprise is working nicely towards cementing its position.