Technologies like GPT have neither initiative nor contradictory, unless you try to insert this feature into their input, or prompt. I have seen ways that I find funny to get this result in other contexts, such as repeating several times, in the prompt, “don’t write nonsense (hallucinate), double-check the results several times”. There are also guardrails to force them towards certain types of behavior.
As for utility, let’s remember that at the moment Software as a Service (Saas) has more utility than AI (in this case GPT and others). SaaS are much more difficult to develop than agencies or consulting firms, such as Enterprise Statistics, even if they are more scalable. But obviously consulting firms have an edge over SaaS and absurdly even over GPT, even though they have trained it with the entire internet: experience and domain intuition allow them to burn many stages, shortening the time it takes for the customer to obtain utility.
As therefore you may have guessed, an AI-based data analyst, created via a custom GPT (CustomGPT) or HuggingFace assistants and who knows what else will come out in the future, give very low utility coupled with the risk that they will misread the numbers (hallucination). And as a result, for 2024, I do not recommend this approach to data analysis. In other contexts I have said that a non-text interface to start analyzing data, such as the free version of Tableau or Looker, have more speed and usefulness. Those who have given a few statistics exams in college may find an AI analyst more useful than the general population, for the rest, at best they are wasting time on things they could do as well on Excel, Tableau, etc., at worst they are producing damage in terms of results with strategic potential.
If you already use an artificial solution for data analysis and would like to understand what STATiCalmo would do instead, as a live statistical consultant, we can hear from you in a free call.