[XenophileJKO]

The crazy thing is the level of effort to say, "have a sub agent validate all references and figures" is so low. I'm paraphrasing, but you don't need much more than that. It would have prevented 99% of the face palms.

I use this regularly for my personal financial research system. Even flagship models make mistakes. Though currently the issue is usually the model using a figure from and older report. Cross-check reduces that dramatically.

[watwut]

Thinking that such prompt will cause the report to be factual is root issue. No it wont, no it is not enough.

[modzu]

dont be so sure they didnt. they can go back and forth hallucinating with each other

[figassis]

This is where the absolutism of let agents to 100% of the work fails. You get adversarial agents pulling all reverences into a table, they might miss some, so run this a few times.

Then have another set of agents, with skills like web browsing (to verify that links actually exist, maybe that references and abstracts actually match, etc), have one engineer (or agent) write a small script to help with this (just make sure you test it, and a bit).

So your work is not verified until your references table is 90% green checkmarks, maybe with uncertainty figures.

A human can then verify the ones with under 90% certainty.

This alone gets you a long way there. Does not costs the millions they're being paid.

It's quite interesting that these companies marketed themselves as them best of the best in excellence, accept no mistakes. I can imagine the countless keynotes and books about this. Or the sales pitches.

Has always been a lie, they just understood how to hide it. Today they don't, and it's embarrassing.

[e12e]

> A human can then verify the ones with under 90% certainty.

How about the author actually reads the finished report a couple of times and checks all the references?

It really is the lowest bar - even lower maybe than running a spell check.

[SpicyLemonZest]

The hallucinations here (https://gptzero.me/news/investigations-kpmg/) would have passed a cursory reference check. It's easy to see when it's laid out in a table that "BNP Paribas. AI Integration: Transforming Financial Journeys. The Banking Scene, 2025." is a false citation, because the title doesn't quite match and it wrongly attributes BNP Paribas authorship to an article written about BNP Paribas by some random Belgian guy doing business as "The Banking Scene". It'd be a lot harder to see when you're skimming through browser tab 9 of 45 and see all the key words match up.

[e12e]

I'm not talking about a reference check by someone other than the author. You'd not put a reference in in the first place, that you hadn't read, since you couldn't formulate the text that relates to the reference?

Ed: thanks for the link - I hadn't seen that yet.

[palmotea]

> How about the author actually reads the finished report a couple of times and checks all the references?

But then you wouldn't be embracing the new agentic ways of working!

[danaris]

How about the author actually, y'know

authors

the report?

[figassis]

Even then, AI can still be helpful. My point is if you're going to use it, use it right, and not with the intent to deceive.

[iugtmkbdfil834]

Eh.. without going into too many details, having seen some face palms at work, I realized that the anecdotes may be closer to a pattern than I would like to believe, which prompted me to start basic howtos available company-wide.

I kinda get it, without experience and trying, how are they to know ( unless they are already 'into it')? After all, corporate training is laughable at best.