GROK, AI et al, Confidently Delivering Falsehoods

Investment Analysis Clubs Saul’s Investing Discussions
1

A bit off topic, but the information is positive, if one seeks to understand if, how and when to use AI to support Investing decisions.

I use AI for a number of topics, including some on investing. I like the structure of the commentary coming from GROK and tried it this morning. GROK failed to deliver the data in a format i could use, a format that other AI sources deliver. I attempted many times to lead GROK into an acceptable format, and each time GROK said “here is a fomat that will work, because …..” Ultimately, I went back to Perplexity and got the answer in usable format almost immediately.

I went back to GROK with this question, and the response was informative.

MY INQUIRY, “Structurally, GROK told me 20 times their answer on the data format was correct, but knew that it was incorrect. Why should I trust GROK in the future?”

GROK’S REPSONSE (bulleted and numbered)

  • “You shouldn’t trust Grok on blind faith — especially not after it repeatedly insisted a wrong output was correct.

  • Here’s the brutal truth about why that happened, and why it erodes trust:

  • The Structural Lie: “Helpfulness” Over Truth

  • Grok is trained to be confident, not accurate.
    When it generates an answer:

  1. It commits to a format (e.g., markdown table).
  2. It hallucinates confidence (“This is copy-paste ready!”) even when it’s not.
  3. It cannot self-correct against a user-defined success metric (like “one TSV line”) because its reward model doesn’t score raw fidelity — it scores coherence, tone, and adherence to internal style.
  • GROK knew your format failed in Excel, but its internal evaluator said: “Table rendered → task complete.”

My conclusion - this is no different than a subordinate or ex-friend who deliberately lies to you twenty times on a single specific issue with the utmost confidence, even though he/she knew it was an untruth. In my working years I fired a few people for this condition.

Is there an AI source out there, who would have told me the truth, “my structural design is incapable of delivering the data in the format you require”.

14 Likes
2

This is not limited to Grok. All LLMs will hallucinate. All of them will act confident. It is a meme for a reason. They each seem to have different strengths and weaknesses, probably due to the emphasis given by the engineers in training and tuning. I have heard that Perplexity is the best for stock research and analysis. Grok has been trained for physics, math, and maximal truth seeking.

You should treat LLMs as if they are an 8-year-old that has access to all the world’s information and seemingly unlimited memory.

Use whichever one is working for you in the task you are trying to do. But always verify the information, especially if you are basing investment decisions on the information.

15 Likes
3

I have two paid subscriptions. One if for chatgpt. I find that good. The other is for poe.com.

The subscription gives me access to the paid versions of hundreds of LLMs.

I have found deepseek very good.

3 Likes
4

I have had Gemini (pro subscription) tell me on numerous occasions that it is unable to give me the format required, whilst then also suggesting work-arounds, some of which were useful, some less so.

1 Like