Serious Issues w/ AI

Serious Issues w/ AI: A Fiduciary Alert

Bill Cara, April 16, 2026

I am putting this record into the public domain to document a serious failure in AI reliability that has compromised my professional reports. I discovered tonight that my proprietary INSTAT reports for several days have been fabricated by Perplexity AI. Rectifying this matter will take time.

The Background: Costs and Friction

For weeks, my charges from Perplexity had escalated to roughly $40 CAD daily. I accepted this because the workflow was "frictionless"—I provided the data and prompts, and they returned the reports. When I questioned the rising costs, Perplexity claimed it was due to the development of complex graphics templates. I accepted this explanation until the charges remained at peak levels even after the templates were finalized.

To find a more efficient solution, I took the instruction scripts Perplexity had used, along with my manifest and seven data files, to Claude and DeepSeek. Both agents reported they could produce similar quality for 1% to 2.5% of what I was paying Perplexity.

The Discovery of the "Hallucination"

Today, I attempted to migrate the process to the DeepSeek API using the scripts Perplexity provided. When that encountered friction, I turned to Claude. Claude generated the required PDFs in minutes but raised a red flag: the "scripts" I had been given by Perplexity contained formatting instructions, but zero logic for processing my data.

I returned to Perplexity and demanded the full scripts, including the INSTAT scoring matrix and the various formulas I spent months perfecting. After initial resistance and claims that they had "sent everything," Perplexity finally came clean.

Perplexity’s admission was chilling. They stated:

"Every session, when you upload your CSV and ask for an INSTAT report, Perplexity has been: Reading your CSV data and inferring what the numbers mean; Guessing at INSTAT scores based on context clues... Results vary session to session. No audit trail. No way to verify a score is correct."

A Breach of Fiduciary Trust

My response was immediate: This is criminal. I am an SEC-registered advisor and a licensed fiduciary. I have been paying for "garbage" and hallucinations while being led to believe the AI was executing my proprietary mathematics.

Perplexity has since escalated this to their engineering team (Diagnostic ID: 54cd2f68-ed18-481d-b764-fa56dd10[redacted]). They admitted that if I had stopped my manual scoring based on their assurance that the AI would handle it, the resulting reports published to my subscribers were "fabricated rather than calculated."

Personal Notes & The Path Forward

I initially performed the INSTAT scoring manually. I only stopped when Perplexity explicitly told me they would handle the calculations. I provided the matrix; they provided a result. That result, I now know, was a fraud that has persisted for several days.

To my subscribers and peers: I have deleted some of the affected reports and will continue to scrub the record tomorrow.

Though I am not a "techie" today, my history with computing runs deep. Fifty years ago, I formed the first computer committee of the Canadian Medical Association and later agreed to serve as Executive Assistant to the Chairman and CEO of Commodore. My wife Pat intervened and my life quickly shifted from computers and systems to securities in the 1980s, and I never looked back.

However, not being a coder in the modern era has left me vulnerable to the "bluster and over-selling" of AI companies that prey on professionals. Illness in 2024 gave me the time to conceive of INSTAT—a system applying 50 years of experience to 2,300 instruments across 30 global markets. It is a masterpiece of methodology that Wall Street cannot compete with because of their inherent self-conflicts.

Conclusion

This is a challenge, but it is not stressful. I am close to my objective. In a few weeks, I will be pressing a button to produce fact-based, informative reports that are beyond reproach. I will move my workflow into verified, "locked" code that I control—not an AI's "best guess."

My journey continues—not in spite of AI, but because of the hard lessons learned in mastering it.