Data Is Beautiful

Here's a thought: instead of fighting this, make it a requirement to publish the prompts before making a speech. Speeches by politicians being low in information density is nothing new, and the usage of LLMs will undoubtedly make that worse, but it also means that they had to have written a terse description of the information they want to convey. If that were public, people could just read that and not waste time listening to speeches.

It would be ironic if the first use-case for LLMs that creates positive value for society involves ignoring its output, though.

I can accept the premise that LLMs are being used to write Commons speeches -- MPs are also people, I'm told -- but these graphs suggest that LLMs are overusing certain stock phrases which have existed in the business world and apparently in Commons speeches since at least 2007.

What puzzles me is why LLMs are more prone to using these particular phrases. Does this happen for all users of LLMs, or only when British MPs in particular are requesting a speech?

I'd be interested to know if the same trend for the same phrases can be found in the Canadian House of Commons, since although they also follow much of the same procedures, North American English should skew the frequencies of certain words. So if the same trend can be found, then that suggests that the common LLMs do lean towards certain phrases. But if the trend is not statistically significant in Canada, then perhaps British MPs issue different prompts than their Canadian counterparts.

What I'm saying is that I rise today to highlight additional avenues of intrigue, as MPs and citizens alike are navigating a world where AI supposedly streamlines daily activities. That certain trends may or may not exist underscores the gravity of this seemingly bustling industry that we call AI.

[just to be clear, that last paragraph is entirely in jest]

Given enough data, you can prove anything. Correlation doesn't imply causation folks!