When can AI be Random ?
Sujit Karatparambil Madhavan
Experienced Developer | Designer | Java | Python | Data | Automation - BASH - PowerShell | Deployment | Azure | AWS | GitOps | K8S | Docker @ Capgemini
AI makes decision, does it make the same decision second time, as users have found it can make a random decision everytime and humans accept it. When can AI be Random ? - is the primary reason for bias or hallucinations. Bias happens when a question rephrased is asked the second time. Hallucinations happen when I press for a view, like the famous rock salt question - again a very specific view. You could have asked whether rock salt is a mineral and AI won't have hallucinated. But bring us to the topic When can AI be Random ? It is much more than bias or hallucinations, it is sort of a fickle minded results as someone might call. It could be random because the same epoch (test / train percentages) could result in different results with a different test data set / but with the same train data set, though test and train are again fixed in evaluating a model. An arbitrary assumption which doesn't seem to possible, or is it a limitation with When can AI be Random ?
When can AI be Random ? - primarily happens when you have limitations when you tested the scenario or benchmarked your AI program. You could have a LLM take up Medical Exams, but if you don't change the test data set and keep the train data set the same then you will never encounter When can AI be Random ? - When you are evaluating the AI Model or LLM.
When benchmarking - questions are being asked to the model, without test / train philosophy, but the problem when you conformed the AI Model you were testing the train philosophy which wasn't actually tested for a outlier - which could again be bias or hallucinations.
Summary : When can AI be Random ? - primarily a status that leads to AI not functioning partially or completely, when one says partially it might be a 98 % result, say accuracy of 99.09 %. Completely could be bias and hallucinations which are 0 % in terms.