I’m not sure if this would happen in the actuarial world, but I’m contemplating in the science world…
We have a lot of legacy code here in SAS (and in some cases in R), but many of our SAS programmers are retired. In other cases packages in R have been depreciated and the script no longer runs. I’d be tempted to run some of our legacy scripts through an AI to either translate them to a different language or update the script with active packages so that we can better use them today.
I can see someone doing something like this and not testing to see if the script is still doing what it should be doing and that it’s doing it properly.
And to be more explicit, I was contacted to talk about AI & professionalism for the Fall Hartford Actuarial meeting in November… so yes, I’d like examples.
I have the old examples of lawyers screwing up, but I’m trying to get some examples – like the vibe coding ones above – that are a bit more pertinent to some of the things we do.
I think an actuary would have to know exactly what goes on with the AI, lest it be a black box, and that it all has to be double-checked and documented, mainly by doing all the work over again manually.
In summary: don’t use AI for professional work.
Thank you for attending my TED talk.
Tough shit, you’re using AI if you’re using any Microsoft Office products now.
and any browsers
b/c everybody is injecting AI (or, rather, a bunch of algos you didn’t want) in your software whether you asked for it or not, which is a pain in the ass
The software you use has been black boxes for decades. You just have convinced yourself you know what it’s doing.
I remember someone bringing this up years ago when I was talking about spreadsheet best practices. He was correct - we’re generally testing that Excel is actually calculating the way you’re expecting it to, though there have been errors at various levels, whether it was an error in the software itself… or the chip.
We older folks remember this particular error:
That was a big brou-ha-ha at the time.
So saying: “Don’t use AI” is a bullshit statement.
I vaguely remember a chemistry professor I had (fall 87 or winter 88) mentioning he found an error in some commonly used software that happened when dividing by an “exact” power of 2.
FTR, I don’t really use ChatGPT for anything (or grok etc etc)
I keep trying to turn off various “AI features” in Office, because it’s annoying and hasn’t adding much of anything other than a propensity to crash the apps. I mainly use CoPilot for summarizing features, and that’s it. And even there, it often misses important points I knew are in the documents, so I still have to go over the full documents and edit the results from the generic AI.
I’ve been using Otter (https://otter.ai/) for a few years now, which has been extremely helpful in audio transcription. I’ve used it in conducting interviews for articles, for instance. I have other software from TechSmith (Camtasia, Audiate), that has made video & audio editing so much easier with new AI features.
I’ve found the special purpose tools as above to be useful. One can check the results directly. It doesn’t matter that it’s a “black box” - I can check the initial recording against the transcript, for instance, and fix any mis-transcriptions.
The “black box” problems are when we can’t check the results against sources or test the results in any meaningful way.
Assuming it wasn’t staged, I suspect it occurred due to someone pushing “vibe coding” to an extreme. They were letting it write, edit, and run hundreds of lines of code without checking it, in a poorly controlled environment, probably with a dubious LLM / prompt / temperature. And no pipeline to catch errors.
But sure, if you fed it a SAS program, and asked it to translate it, it probably would make some errors or hallucinations. And it’s possible some of those errors would change the basic logic of the code. At the very least you should read what it writes. And run a couple tests to make sure it produces what you expect.
Likewise, just updating R libraries should be easy enough (is there a way to automate that?) but there’s a big difference between an R programmer doing it and looking at the changes, and a person who can’t or won’t even look at the changes.
As a lowly entry level actuary I discovered a basic excel computation error if you entered it in just such a way, I think for this reason… basically a binary to base 10 conversion error. It only mattered because of how we were iteratively calculating each level of our HO value factor curve, but then when you checked what the value should be you’d be off by a couple hundreths.
Once I ran into a problem with Excel implicitly truncating numeric strings at 16 digits. It was such a rare case (who uses long numeric strings???) but it did eff up my match().
I think LLMs are worse, because the hallucinations are so common, and believable, and material. I wouldn’t say “don’t use them” but you -really- can’t afford to be ignorant of how they fail.
While with Excel you can go years without a problem.
The LLMs can be useful – but you have to use them in the right place (and edit them afterwards, if you plan on publishing anything.)
Recently, I wrote an article about using various AI tools to write articles, and used the tools to write the article (mainly, to edit down my usual verbose prose.) I put a note to the supposed editor(s) of the newsletter in a footnote… which the editor(s) did not remove after actually following the instructions I put in the note.
That was definitely human error.
I did not tell them anything about it, bc who cares about a footnote error.
It reminded me of the article I wrote about spreadsheet errors, in which the editors screwed up my title, misspelling it. They had printed copies at the SOA annual meeting, and were handing it out. I thought it was funny. (they later fixed the digital version)
There are cases where it is Excel screwing up (often, a screwed-up calculation tree - it still happens, but it’s ephemeral.) Usually, spreadsheet errors are due to the users.
The user error for generic LLMs is the expectations of what the output should be. And then not checking that output before handing it on to other people or as input to the next step of a process.
Its really good at summarising teams meetings so you dont need someone taking notes (or you yourself taking your own personal notes). It can also generate decent replies to coding problems (in VBA, R, Python) but you definitely need to review its responses. I don’t really use it at all for excel either (not sure how exactly it can help me). Some of my coworkers use it to create PP slide decks but that comes from a process (you can’t really create useable things from scratch).
Hard to quantify how much time it is saving me, but its probably around 10% at this point. Not terrible, but not great either.
Yes-- I think this is the main reason actuaries are going to be shielded for a while. AI can not see/think in 2 dimensions. I suppose you might translate a spreadsheet into a list of interacting formulas, and an AI might make sense of that, but it would have no training, and at least have a little difficulty making visually reasonable edits.
The rest the world is being spun upside down right now. Maybe AI will come for our jobs when it can rewrite our work in Python.
I don’t use LLMs for much at all, but I like to talk about them so…
I think LLMs are great at improving things. Like, you should always let an AI review a memo before submitting it. (This might be automatic right now? I see word has a big fat copilot icon now.) They are likely to catch spelling/grammar errors that you missed as well as provide excellent style suggestions.
Yes, it can invent errors and make bad suggestions.
However, you don’t need to blindly accept its suggestions. You can review the suggestions, one by one. If you’re talking about a 10 page document, and it finds 10 errors, it just takes a minute validate each one.
There is basically no risk at all in using an AI in this way. It can only make things better.
The risk is that AI always beckons you to be lazy.
Why proofread the document yourself if AI can do it?
Why review the errors one by one? Just let AI auto-make the changes?
Why write all those paragraphs, just come up with an outline, and let AI write them?
Why write an outline-- just brainstorm some ideas, and let AI write that.
Why come up with ideas-- AI is good at that right?
etc.
And the depressing truth is LLMs can be helpful at every step, to a degree. But each additional part of a process you unload to an LLM will have trade-offs.
“Vibe coding” is basically the same issue. AI is most obviously wonderful when it fixes your code, by pinpointing an error that would have taken you 10 hours to find and resolve. It is also great at “autocomplete”, and good at writing code… and that naturally makes you want to use it for everything, but the more you use it the more careful you need to be. And instead people tend to be less and less careful. And some people start to think, why make the effort? Why check it? Why read it? Why learn it at all?
I have heard actuaries ask LLMs about regulations. And that makes me cringe. It can absolutely hallucinate regs. Or simply be out of date. It’s probably fine – even good-- if it provides a link taking you to the current website, but you need to validate that, and read the website yourself.
In general, I think AI search can be useful. The other day I had googled a tech issue and the LLM took me not just to the right link, but the correct paragraph within the link.
Another time, I was looking for a certain book, and I couldn’t google it because I only had some vague common words to describe it. But an LLM was like “oh you’re looking for ____ ____ ____”
Oh, and another use was I fed it some information about my cat’s condition. And it provided much more useful information than the vet. That is something I could have googled though.
More generally, I think LLMs should be used to get second opinions on medical matters.
I understand why society is uncomfortable with letting a bot diagnose you, but doctors make a lot of mistakes in real life.
My nephew is in an international math competition and is training. He and my mom got stuck on a problem and I don’t remember enough geometry to solve it and may not have learned the math needed to do it anyway.
I tried getting AI to solve it and it apparently got the answer wrong.