Sat, Jul 4

Here’s why generative AI will never replace humans

When I moved my blog from Blogspot to Substack last August, I needed help many times. On the Blogspot platform, there was (and still is) no help at all. You have to go online to a Google Group (Blogspot is owned by Google) and ask your question there. But Substack has a really great chatbot that springs to attention as soon as you ask for help. While it sometimes takes me two or three back-and-forths to get the chatbot to understand what I’m asking, 90% of the time it understands what I need and gives me a definite answer. Of course, the answer is sometimes, “We just don’t support that now, but I’ll submit a new feature request for you if you would like.” I always confirm I’d like that, but I know full well I’m unlikely to see the new feature implemented in my lifetime.

Recently, I wanted to see a post I had written in 2015. (Some background: I started the Blogspot blog in January 2013. I transferred all my Blogspot posts - over 1200 then - to Substack last August and stopped posting any new posts on Blogspot last fall. In fact, as you’ll see if you go to the link at the beginning of this post, I will soon take down the Blogspot blog because it’s become too popular among AI models eagerly seeking information in posts I wrote years ago. In fact, my Blogspot blog had 672,000 pageviews last month alone, while I guess I receive no more than 10-15,000 pageviews per month in Substack and in Energy Central, where I also put up my posts. I believe the Blogspot pageviews are almost entirely from LLM training)

When I entered “2015” in the Substack search bar, I wasn’t terribly surprised that, while I immediately was shown every post where “2015” appeared in the text, the search doesn’t look at the date of the post itself, and therefore couldn’t show me which posts I wrote in 2015 (or any other date range). I asked the chatbot if it was possible to search on a date range. It cheerfully pointed out to me (it’s always cheerful, of course) that the feature I want isn’t supported in the search bar, but it would be happy to submit a new feature request for me. Even though I knew this would be useless, I let it do that anyway. Then it closed the ticket.

After that, I received the usual email asking me to clarify what I wanted, so I did that – again, not expecting anything more to come of it. However, this time a human emailed me to point out that there’s a completely different way – not involving the search bar – that I can see all of my posts within any date range. It works fine, so now I can find posts from 2015 or any other time period.

The important point about this is that I don’t think the person who solved my problem is some sort of brilliant individual that Substack is lucky to employ. They just did what a human who’s trying to be helpful does naturally (or should, anyway): go back to the problem I was trying to solve and figure out how it could be solved in the context of the Substack platform. The chatbot had categorized my problem as narrowly related to the search bar (even though it was formulated more generally), but my problem was really that I wanted to see posts from a range of dates – and I didn’t know the platform well enough to realize that the search bar isn’t the only possible way to do that. My guess is the human figured out the answer to my problem within a few seconds, even though the chatbot would probably never have done so without being given additional prompts that a human wouldn’t need.

The moral of this story is that generative AI, which is simply an elaborate statistical algorithm for predicting the next word in a sentence, will never reach the point where it can satisfactorily answer a question, if the questioner didn’t ask their question in a way that the answer will be straightforward – specifically, if the questioner didn’t ask the question in the context of the correct response. I asked my question in the context of the search bar, so that was all the chatbot focused on in its answer. But the human understood the question in the right context: “Given that I’m a Substack user, is there a way for me to find all of my posts within a particular date range?”

Of course, since this was Substack’s chatbot, it could have been trained always to consider every question that appears to be about a particular function within Substack to be instead a question about the platform as a whole. That might have allowed it to solve my problem without requiring human intervention, but what if I’d really been asking whether blogging platforms in general allow searches by posting date? That would have required an even bigger context, which the chatbot probably wouldn’t have been trained for.

Human beings don’t know all the answers, but they should at least be able to reformulate the question in its proper context. Generative AI can’t even do that and probably never will. 

Tom Alrich’s Blog, too is a reader-supported publication. You can view new posts for two months after they come out by becoming a free subscriber. You can also access all of my 1300 existing posts dating back to 2013, as well as support my work, by becoming a paid subscriber for $30 for one year (and if you feel so inclined, you can donate more than that or become a founding subscriber for $100). Whether free or paid, please subscribe. 

If you would like to comment on what you have read here, I would love to hear from you. Please comment in my chat or email me at [email protected].