I wrote the guidance for our department on the use of Co-pilot when it was integrated into Teams for this exact reason. And don't feel bad, it's a very widely held misconception.

You can't use LLM's like ChatGPT, Grok, P.ai or Co-pilot for this kind of thing. It's not what they're designed for. They're designed to produce realistic sounding bodies of text. You can get accurate answers from them, and if the training data is pretty uniform and robust in the outputs you're expecting to get it will probably spit those out reliably. However, if the outputs are not uniform or robust it will still produce a confident, coherent sounding piece of text based on what you ask it, but the odds of it being correct are much, much smaller. It's great for automatic the writing of code, because generally that's verifiable. It's not there to "tell you the truth".

The reason I asked for your sources on this is because I was interested, not because I immediately disagreed with your assessment of the particular piece of legislation.

When I did actually look for proper studies (and I'll be honest, it was a quick one, I'm not that interested) it was a pretty mixed bag. But there was a masters thesis which looked into the price changes in the most scrapped cars under the scheme and found the price change compared to the background control was about $13. I'd be dubious about putting much weight on that particular study, but it does seem mildly indicative of something much more considerable going on in the background (which is unsurprising, given the global economic collapse at the time). This fits the eye test too, when you look at used car prices in the US, where it's been a pretty steady increase until covid happened https://www.economist.com/graphic-detail/2021/07/14/the-used-car-market-sent-american-inflation-soaring-in-june