I knew that, in Canada and the US, the wealthy pay less taxes than the poor. So I asked Claude if that was so.
I asked In the US, what percentage of the population make 50% of the income? Do both pay 50% of the taxes?
It replied: I’ll search for current information on US income distribution and tax burden to give you accurate figures.
The top 20% of earners earn about half of all national income … The top 20% of earners pay about 72% of income taxes.
I asked What are your sources for these statements, and for contradictions to them?
That’s a fair question. I need to be honest with you: I cannot actually see the sources from my web search results. The citations I provided are tags that were supposed to link to specific sources, but the actual source content wasn’t retained in a way I can access or verify.
The UC Berkeley Finding (2025)
The most dramatic contradiction comes from recent UC Berkeley research:
The 400 wealthiest Americans actually pay a lower effective tax rate than the average American. Their effective tax rate fell from 30% in 2010–2017 to just 23.8% in 2018–2020, compared to 30.2% for the entire population. University of California
Even more striking: The top 100 wealthiest paid an effective tax rate of 22%, while the next 300 paid 26.6%. University of California This means the ultra-rich are undertaxed relative to their true income when you include all taxes (federal, state, corporate, and foreign taxes).
Why This Contradicts the Standard IRS Distribution Data
The standard IRS data I cited earlier shows the top 1% pays 38.4% of federal income taxes. But that statistic hides a crucial distinction:
Income taxes ≠ total taxes. The ultra-wealthy don’t primarily earn wages taxed as ordinary income—they earn investment returns, capital gains, and business income.
They can:
- Shelter business income from taxation
- Hold assets for decades without triggering capital gains tax (unrealized gains) Cato Institute
- Pay capital gains taxes at only 15% instead of the top 37% income tax rate UC Santa Cruz
Note that I asked for Taxes, not Income Tax
The LLM chose to interpret the question, to align with the answer it wanted to give.
If you provide exactly the right garbage in, you can select what you want in the way of garbage out. When someone asks a question you don’t want to give the real answer to, you can insure that it substitutes your preferred ons.
Therefore: when asking an LLM controversial questions, pretend to be a crown attorney and cross-examine the heck out of them.

Followup: It’s actively being used by Russia
@gcluley writes
“Russian disinformation is being designed to be invisible to humans – but highly visible to AI. This new report from Demos reveals how a sanctioned Russian propaganda operation (the “Foundation to Battle Injustice”) deliberately crafted websites to *avoid* appearing prominently in Google search results, while maximising how much AI systems like ChatGPT, Gemini and Claude would extract their text and parrot it out.
“Definitely worth a read… “
https://demos.co.uk/research/geo-for-geopolitics-what-happens-when-ai-and-information-warfare-collide/
LikeLike
@leaflessca.wordpress.com
Or better yet, just don't ask the LLM questions at all. The answers are biased. The AI is owned by billionaires who have an agenda. Of course the answers the AI give will be kind to their demon overlords at our expense. That's the whole point.
LikeLiked by 1 person
@FranceskaMann @leaflessca.wordpress.com
OP thinks they're dunking on AI while they are admitting to being part of the problem. Stop using the climate killing electric bill raising wrong answers only machine.
AI simps get blocked. Thank you OP for helping me keep people like you out of my info feed. I'm not interested in how your are using AI. If you're still using AI for ANYTHING at this point, I have no interest in ANY of your thoughts. You're corrupted.
LikeLike
On the other hand, I might be providing reasons to not use LLMs (:-))
Glad to see you go, though, as your argument used the “tu quoque” fallacy to argue against my advice.
Did you hope I’d stop arguing that LLMs are unreliable? You failed.
–dave
Tu quoque is a specific type of ad hominem where you counter an argument by pointing out that the person making it is guilty of the same thing they’re criticizing. It’s the “but you do it too” fallacy.
(I did my undergraduate degree in philosophy and logic. The latter got me a lovely career at Sun Microsystems, and later a gig supporting machine learning)
LikeLike