One of the questions, that I see baffling a lot of lay person is to understand how a rule based system would differ from a full blown ANI system.
In order to understand this we need to take a detour.
The detour is to take one real life use case that say a bank, uses to process home appraisals. Normally, when a home loan request is received by a bank, the bank would send an appraiser to the property sought to be bought and request a valuation. This valuation is performed by a trained person called appraiser. The appraiser will typically fill a form that consists of the details of home like, number of bedrooms, bathrooms, kitchen, quality of construction, the value of similar homes in the neighborhood and provide a value for such a home. This report is then filed back to the bank which will usually have two or three levels of approval (i.e. quality control) process. The reason for the approval is so that the any errors in the form or the final appraised value is caught before the loan is underwritten and sanctioned to the borrower.
For ages, the quality control process is often simplified by a rule based system. For example, if the appraiser has forgotten to fill a certain row (or for that matter filled something wrongly) for example, number of bedrooms, then the forms is rejected and sent back to the appraiser. Another QC optimization is done for higher value homes, which go through three level approval process instead of a two level process. This is familiar to most people as a rule based system.
During the QC process, the value of the home used for appraisal is usually cross referenced against Automated Valuation model (AVM), which has traditionally tended to be statistically derived value of the homes.
This will rule out obvious howlers. The AVMs are not very accurate but for the purpose of ruling out obvious errors, function well. Of late, the AVMs are moving towards a machine learnt model of deriving the home value, which improves the home value accuracy. In order to become better at it, the data of home sale values are fed into this system – state by state, by district, by ZIP/PIN and by neighborhood.
One such Machine Learning system will take the historical home value inputs and become better at accurately predicting the home value. A complex multi layered machine learning system will result in better AVMs within this decade. Such an ANI system will effectively replace their human counterparts.
So, in terms of evolution, a simple rules based system is initially and inaccurately competing with statistically derived models, which are bettered by machine learnt and ANI over a period of time.
However, this evolution comes with its own bias. For example, it has been seen that home values, as assessed by AVMs, in minority (i.e. underprivileged & black) neighborhoods in US, take longer to reflect the accurate values due to the inherent bias in data fed to train the ANI systems.
These data are a reflection of human biases that have built over a period of time. We can only presume that the lies in statistical data will not become encoded into machine intelligence.
AI-first world is being built by Systems that are "trained" by humans. It is imperative that extensive research is done to ensure unbiased and accurate data being fed. What do you think?