We Pitted Two AIs Against Each Other to Solve a Wall Street-Level Puzzle. The Results Were Not What I Expected.

The First Attempt: GPT-5 and the “Illusion of Thinking”

When I first tested GPT-5, it looked confident. But the moment the ledger got messy, the logic fell apart. It was a big infinite loop and while it was fast, it kept making the same mistake, over and over and over.  This is what Apple’s The Illusion of Thinking study describes: large reasoning models (LRMs) seem capable until tasks become too complex.^[1]

Apples research outlines three regimes:

1. Low complexity: conventional models often outperform LRMs.
2. Medium complexity: LRMs provide an advantage.
3. High complexity: performance breaks down for both.

My ledger project quickly pushed us into the high-complexity regime. Along the way, GPT-5 demonstrated several well-known failure modes:

• Confabulation & hallucination — inventing rules that didn’t exist.
• Inference & reasoning errors — drawing incorrect connections.
• Contextual drift — forgetting prior corrections.
• Overthinking & analysis paralysis — exploring irrelevant alternatives.
• Confirmation bias — sticking to an incorrect early assumption.

The Turning Point: Two AIs, Two Roles

At that point, I stopped asking GPT-5 to “figure it out” and just tried to get it to follow instructions. I distilled the logic that GPT-5 and I had developed, then handed this structured playbook to Gemini.

Could Gemini execute the instructions consistently and solve the problem? GPT-5 had me in an infinite loop and the project was on the brink.

The Gemini Difference: Methodical over Memorizing

Gemini thrived where GPT-5 struggled. Its strength aligned with the “medium-complexity advantage” described in Apple’s research.^[1]

Rather than improvising, Gemini followed the established rules step by step.  It was methodical:

1. Precise discrepancy analysis: traced numerical gaps to specific transaction types (e.g., “Legal Fees,” “Misc Credit”).
2. Rule permanence: corrections were retained in a growing Master Instruction Block.
3. No drift: context was preserved, avoiding contradictions.

AI is not taking our job anytime soon, but it is a good collaboration partner. It functioned like a junior analyst who’s not totally wet behind the ears.

Why AI Works Better as a Collaborator

The problems I saw in practice are echoed in research. Multiple studies come to the same conclusion:

AI today is most effective as a collaborator, not a replacement.

Taken together, these studies highlight a central truth: AI often suffers from hallucination, logical errors, and bias. But paired with human oversight, it becomes a powerful tool for consistency, scale, and rule-based execution.

The Takeaway: Scope over Scale

At Intro XL, the lesson is clear: success comes NOT from deploying the largest models, but from scoping tightly, codifying rules, and treating AI as a disciplined collaborator.

• Humans: provide judgment, context, and course corrections.
• AI: provides memory, consistency, and scalability.

That balance turned a messy ledger into a somewhat reliable Borrower Profitability Scorecard. AI’s true power is unlocked not when it tries to “think” like us, but when it works with us. It’s a tool and not an end all be all.