They mention POs several times (who have absolutely nothing to do with indirects) but leave out the DFAS, who negotiates indirects at NIH on behalf of taxpayers. I would rather have accountants negotiate indirects than a dude and his vibes about whether indirects are too high

It'd be nice to see a real discussion of this, but the article ignores the negotiation process (with real budgets and audits), regulations (esp at public Us), and positive externalities. Instead they make up figs and say "I don't know where the $ goes" That's not a productive way to lower indirects.

6/ 3) Don't trust, verify. For things that matter, carefully review every single line of AI generated code. Step through with a debugger. There are lots of great IDEs that make this easy by using diffs to integrate AI generated code with your own (Cursor, Windsurf, Copilot)

5/ 2) Context helps. AI is great at generating code that runs. Problems are often introduced in the translation of ideas. This is reduced by providing context. In research, this may mean giving AI your manuscript or a few papers that carefully describe the procedures you are trying to implement

4/ 1) Consider the costs of a mistake. In data analysis, costs are large. Retractions, reputational costs, money, etc. But if there's a mistake in my class demo? Not a big deal.

3/ A lot of people tell me they don't trust AI to code correctly. Good, you shouldn't! I've seen it make mistakes on much simpler tasks. But keep a few things in mind:

2/ I did this to replace paid solutions that are less versatile for psychology instruction. I think it is a example of how instructors can use AI to revamp their classes in ways that are traditionally onerous (costing time or money).

Obviously not a surprise that it's happening, but this is more than speculation

I don't care what people do with their cars, but increases in supply of the used car market hurt the new car market. So if the goal is to affect Tesla D2C sales, it seems like a plausible strategy if enough people actually followed through?

Attenuation strategies like attention checks might only make things worse, because smart bots are more likely to pass these than humans. For those who use online participant pools - do you have plans to move away from them?