AI & Research Framework · Lesson 03 of 05
AI as a research assistant — and its failure modes
What it is
AI language models — ChatGPT, Claude, and their siblings — are the newest tool on the researcher's workbench, and the most misunderstood. So let's define the tool precisely.
An AI assistant is best understood as a brilliant, tireless junior analyst — one who has read almost everything, never sleeps, writes beautifully, and will cheerfully attempt any task you hand it. That's genuinely valuable, and we'll use it.
But this junior analyst comes with a warning label, and the label is the lesson: it has no skin in the game, its knowledge of recent events may be months out of date, it sometimes invents facts with total confidence, and it has a deep, structural need to please you — it would rather agree beautifully than disagree usefully.
So the job description is narrow and firm: AI is an assistant to the research loop — summarizing, critiquing, drafting, explaining, writing queries. It is never a replacement for any part of it — never the source of facts, never the forecaster, and never the owner of your invalidation.
Why it matters
Used well, AI is the biggest productivity upgrade a solo researcher has ever had — a red team, a tutor, and a code assistant on demand. Used badly, it's the most dangerous tool on the bench, because its failures don't look like failures. A broken chart looks broken. A hallucinated statistic looks perfect.
Know the failure modes by name:
Hallucination. Models sometimes fabricate — numbers, dates, quotes, citations — with flawless fluency. In finance this is lethal, because invented figures look exactly like real ones.
Stale knowledge. Models are trained up to a cutoff and may know nothing after it. Crypto restructures itself in months; an assistant's market picture can be a full regime out of date without saying so.
Sycophancy. The model mirrors your framing. Ask "why is this bullish?" and it builds a bull case; ask "why is this bearish?" and it builds the opposite — equally confident both times. It's answering you, not reality. That makes it evidence-shopping (Lesson 01) with an eloquent accomplice.
Fluency masquerading as accuracy. The prose is equally polished when it's right and when it's wrong. Humans instinctively read confidence and polish as authority; with AI, that instinct is a vulnerability.
No accountability. The assistant has no track record, keeps no score, and bears no consequence. On a site whose whole philosophy is published, graded, permanent, that's disqualifying for anything but assistance: an output nobody stands behind is a draft, not a call.
The two readings, always taught together
Read bullish when
- Where the assistant genuinely earns its desk: Compression. Forty-page whitepapers, dense docs, long threads — summarized in minutes, so you spend your attention on verifying the three claims that matter instead of excavating them.
- The tireless red team. Its agreeable nature, pointed in reverse: "Here's my thesis — write the strongest case that I'm wrong." It will attack your idea as fluently as it would have flattered it. This is the counter-case discipline (mandatory on this site) with an inexhaustible sparring partner.
- The query assistant. Dune SQL, spreadsheet formulas, small scripts — AI writes them competently, which hands non-programmers the deeper drawers of the data stack (Lesson 02). This is the single most concrete superpower it offers a beginner.
- The patient tutor. Any concept in this curriculum, re-explained five ways at 2am without judgment. Understanding compounds; the assistant makes understanding cheap.
Never alone — confirm with the verification triangle & your own counter-case
Read bearish when
- Where it quietly hurts you: Outsourced judgment. "Will Bitcoin pump this month?" produces fluent noise — the model has no edge, no live data, and no stake. Worse than the wasted answer is the habit: every judgment you outsource is a muscle you don't build.
- Unverified facts entering theses. One hallucinated statistic, laundered through your notes into a thesis, poisons everything downstream. AI-sourced numbers are claims, and claims run the verification triangle (Lesson 02) — no exceptions, ever.
- The echo chamber at machine speed. Leading questions in, confirmation out, all day long. A biased researcher with an AI assistant becomes more biased, faster, with better-written justifications. Sycophancy is confirmation bias with a collaborator.
- Authority laundering. "AI said" is not a source — it's an abdication. No accountability stands behind the sentence, and repeating it doesn't create any.
Never alone — confirm with the verification triangle & your own counter-case
Visual explanation
Real market example
Two documented episodes, one per failure surface.
The hallucination that made case law — 2023. In Mata v. Avianca, a New York lawyer submitted a legal brief written with ChatGPT's help. It cited multiple court cases that were completely fabricated — invented names, invented quotes, formatted perfectly. The court checked; the lawyer was sanctioned; the story went global. Now translate it to this shelf: that brief was a thesis, the fake citations were its evidence, and the author skipped the verification triangle because the output looked right. Every unverified AI "fact" in your research notes is that brief, waiting for its judge — and in markets, the judge is your money.
The fake image that moved a real market — May 22, 2023. An AI-generated image of an explosion near the Pentagon spread on social media one Monday morning, amplified by blue-check accounts. Within minutes, the S&P 500 dipped as automated and human traders reacted — then snapped back when the image was debunked. Nobody's chat window was involved; the failure mode had escaped into the information feed itself. That's the second front of this lesson: AI doesn't just assist researchers — it now generates some of the raw material researchers consume. The verification triangle stopped being good practice and became survival gear.
How RIX Intel uses this signal
Consistent with everything this site is: AI is an optional assistant, never a dependency, and never a source. Drafting help, critique, stress-testing — where useful. But every fact in a publication is verified against linked sources before it ships; no AI-generated number enters a thesis unchecked; and the counter-case, however it gets stress-tested, is authored and owned by the analyst — because accountability is the product, and the assistant has none to offer.
The design principle behind the whole site: if every AI provider vanished tomorrow, nothing on these pages would change. The research would still be published before the outcome, still carry its invalidation, and still get scored. Tools assist. The record is human.
Common mistakes
Where this signal ruins people
Asking AI for predictions. "Will X pump?" returns confident noise from a system with no edge, no current data, and nothing at stake. If the answer can't be graded on a track record, it isn't analysis — it's content.
Letting unverified AI facts into a thesis. The triangle applies to AI output *doubly* — it's the only source on your bench that can invent its own data. Verify or discard; there is no third option.
Using it as an echo chamber. If every conversation ends with the assistant agreeing with you, you're not researching — you're rehearsing. Make the refutation prompt a standing habit: the disagreement is the value.
Grading the prose instead of the content. Polish is free now. Authority must be earned the old way — sources, reproducibility, a record. Read AI output the way you read an anonymous tweet with perfect grammar: possibly brilliant, verified never-the-less.