How to Evaluate AI-Generated Stock Reports: A Checklist for Skeptical Investors

Why AI stock reports need evaluating

AI-generated text has one property that makes it uniquely dangerous in finance: it sounds equally confident whether it is right or wrong. A fabricated revenue figure and a correctly cited one arrive in the same fluent, assured sentence. Tone, polish, and specificity — the cues people normally use to judge a human analyst — carry almost no signal when the writer is a language model. A report can be beautifully organized and still be built on numbers that trace to nothing.

That is why evaluation has to shift from how the report reads to what the report lets you check. Every check in this list is structural: it asks whether the report contains the machinery of verification — sources, dates, scenarios, invalidation conditions — rather than asking you to judge whether the prose feels trustworthy. Structural checks work on AI output because they are the one thing fluency cannot fake. (For the broader question of what AI does well and badly in this domain, see AI stock research; this page is the standalone checklist.)

The one-question shortcut: can you check it?

The entire checklist compresses into a single test: is every material claim traceable to something you could verify yourself? A revenue figure should point to a filing. A margin trend should point to reported quarters. A price should carry the date it was observed. A thesis should name the events that would break it.

If the answer is no — if the report asserts things you would have to take on faith — then whatever else it is, it is not research. It is a narrative. Narratives can be entertaining and even correct, but you have no way to know which, and in a your-money context that difference is the whole game. Every section below is this one question applied to a specific part of the report.

Check the sources

What to look for: claim-level citations — material numbers and facts that each carry their own source, pointing to SEC filings, company disclosures, or dated reports. The test is granular: pick three specific numbers in the report (a revenue figure, a share count, a margin) and see whether each one individually traces to a named, dated document.

Where to look: at the claims themselves, not just the end of the document. A source list appended at the bottom is weaker than per-claim sourcing, because it tells you the report consulted sources without telling you which claim rests on which source — the gap where fabricated numbers hide. If the report offers only a bibliography, spot-check it: open one cited filing and confirm a specific number actually appears there.

What failing looks like: precise-sounding figures with no source at all (“revenue grew 34% last quarter” — says who, for which quarter?); citations to vague authorities (“analysts estimate”, “reports suggest”); or secondary commentary cited where a primary source exists. Primary sources outrank commentary: a 10-K outranks an article about the 10-K, and an earnings transcript outranks a summary of it. A report that consistently cites commentary when filings are freely available is choosing convenience over verifiability.

Check the as-of date

What to look for: an explicit analysis date — “as of” — and, separately, price-as-of context: the price the analysis was built on and when that price was observed. These are two different freshness claims, and a rigorous report states both.

Where to look: the top of the report, near the summary or the price data. It should not require detective work. If you find yourself inferring the report’s age from which quarter it discusses, the report has already failed this check.

What failing looks like: no date anywhere; a date on the document but not on the price; or valuation math built on a price that has since moved materially. This is the most common silent failure in AI-generated analysis, because nothing about the prose changes when the data goes stale — a report describing last month’s price sounds exactly as current as one describing this morning’s. Stale data is not a flaw in itself; undated data is, because it presents old information as current. And note the honest limitation this check implies for every tool, Monsaic included: a report is a snapshot. It can be complete and sourced on its analysis date and still be overtaken by events the following week. The date is what lets you know to check.

Check for falsifiability

What to look for: kill criteria — the specific, checkable conditions that would prove the thesis wrong, stated as part of the analysis rather than retrofitted after the fact. Good kill criteria name observable events with thresholds. From Monsaic’s example analysis of NVIDIA (NVDA): hyperscaler AI capex plans rolling over for two consecutive quarters; gross margin structurally falling below 68% without a mix-transition explanation; competitor accelerators taking enough share to flatten data center growth. Each one is something a reader could watch for and recognize when it happens.

Where to look: the thesis or risk sections. The kill criteria should be specific to this company’s thesis — you should not be able to paste them under a different ticker unchanged.

What failing looks like: no invalidation conditions at all; or conditions so vague they can never be triggered (“if fundamentals deteriorate”, “if sentiment shifts”). A thesis that nothing could invalidate is not a strong thesis — it is a story engineered to survive any outcome, which means it makes no real claim about the future. This check is also the fastest one to run: ask “what would prove this wrong?” and see whether the report already answered.

Check the valuation framing

What to look for: valuation scenarios — bull, base, bear, and tail cases — each with a price target, a probability, and stated assumptions: what must be true for that scenario, and what breaks it. The probabilities should sum sensibly, and the assumptions should be specific enough to argue with.

Where to look: the valuation section. The assumptions matter more than the targets — a scenario is only as useful as the “what must be true” behind it. Note that a good bear case is not just pessimism; it has its own breaking condition. In the NVDA example analysis, the bear scenario breaks if demand stays sold out and earnings revisions keep moving up — a specific, watchable condition, not a mood.

What failing looks like: one confident price target with no range, no probability, and no stated assumptions. A single number is false precision — it implies a certainty about an inherently uncertain future that no analysis, human or AI, actually has. Ranges without probabilities are only slightly better: “somewhere between $50 and $500” is honest but useless. The discipline is scenarios with probabilities with assumptions, so that when reality diverges you can see which assumption failed.

Check both sides

What to look for: a genuine bull case and a genuine bear case, each argued on its merits, plus a risks section written for this specific company. In a well-structured report the two sides get comparable seriousness — the bear case should read like it was written by someone trying to win the argument, not by someone ticking a box.

Where to look: compare the depth of the upside and downside sections directly. Then read the risks section and apply the transplant test: could these risk paragraphs be pasted into a report on a different company without anyone noticing? “Macroeconomic uncertainty”, “competitive pressures”, and “regulatory risk” fit every ticker on the exchange, which means they inform you about none of them. Company-specific risk names concrete exposures — a customer concentration, a margin dependency, a pending proceeding.

What failing looks like: a report that is all bull case with a token risks paragraph (it has an agenda), all bear case with no steelmanned upside (same problem, inverted), or boilerplate risks that betray the analysis never engaged with the actual company. One-sidedness is sometimes a blind spot rather than an agenda — but from the reader’s side the effect is identical: you are seeing half the picture.

Red flags worth walking away from

Any one of these is disqualifying on its own — not a point deduction, a reason to close the tab:

• Guarantees or return promises. “Guaranteed upside”, “will double”, “can’t lose”. No honest analysis of an uncertain future contains these words.
• Superlatives without criteria. “The best AI stock”, “unmatched moat” — with no stated basis for the ranking. Superlatives are conclusions wearing a costume.
• No sources. Material numbers that trace to nothing. Unverifiable is the same as unverified.
• No date. No “as of”, no price context. You cannot know what the report knew or when it knew it.
• No invalidation conditions. Nothing stated that could prove the thesis wrong. A story, not a claim.
• Advice framing. “You should buy”, “add this to your portfolio now”. Research describes a company and a thesis; it does not know your situation, and anything telling you what you should do while knowing nothing about you is marketing.

How Monsaic holds itself to this checklist

The checklist above is not adjacent to how Monsaic works — it is the structure Monsaic enforces on every report. Each report carries claim-level citations with its source list, an explicit analysis date with price-as-of context, four valuation scenarios (bull, base, bear, tail) each with a price target, probability, what must be true, and what breaks it, and kill criteria defined before any verdict is assigned. Forensic grades on management and governance, legal and regulatory exposure, dilution, and revenue quality — plus a narrative-versus-substance score — put the “check both sides” discipline into fixed sections rather than leaving it to mood.

The full fifteen-section structure and the reasoning behind it are on the methodology page, and who builds Monsaic — and why the design assumes you shouldn’t take his word for it — is on the about page. The honest limits apply here too: a Monsaic report depends on the quality of its sources, reflects its analysis date rather than this moment, and is not personalized to any reader. Which is exactly why the invitation stands — take this checklist and run it against a Monsaic report. If a claim doesn’t trace, a date is missing, or a kill criterion is vague, the checklist worked.

FAQ

Run the checklist on a real report

A checklist only helps once it’s applied. Open a covered stock’s excerpt and test it: is the verdict falsifiable, is the key risk stated, is there a condition that would break the thesis? Monsaic publishes the excerpt so it can be graded.

Browse covered stocksSee pricing

Keep reading

• AI stock research, done rigorously — the Monsaic pillar
• The Monsaic methodology
• How AI stock research works
• How Monsaic checks claims
• Generate your first Monsaic