For some time now, I’ve been using various generative AI systems to replicate (or maybe I should call it shadow) the work I already do during normal work hours. It’s often done in response to a feature you saw someone demonstrating online. In some cases, both accurate information and reasonable references may be provided. Similar to a successful Google search. But once you get beyond narrow, simple facts, things quickly get confusing. When I urged Claude from Anthropic to provide a quote from me, he initially refused on grounds of “copyright”. I then encouraged readers to visit my Substack. I don’t use this newsletter platform, but I have frequently criticized it for courting and making money from viciously racist neo-Nazi groups.
I recently heard from an energy analysis friend that you can use Gemini or Claude to extract tabular data from charts. I’ve been using manual assistance tools to do this for years, but using drag and drop is a game-changer.
To test this, I created a graph of US power sector emissions by year from the data I already had and asked the AI tool to do the opposite: generate data based on the graph. I then compared the resulting table to the original table. Although the data were close, there were some differences. We created a graph from the new reverse engineered data and ran it again, repeating the process four times. After four runs, we observed an 8% variation in the emissions totals over the years, with an average variation of about 2% across all data points. This may not be a big deal once someone uses it, such as in a social media post. But what happens when everyone processes visual or tabular data like this? This phone game exacerbates errors in a completely untraceable and unauditable way.
