Chapter 31 · How AI Changes Crypto Research

A Real-World Problem

On Monday morning, you must present an unfamiliar protocol at a meeting in two hours.

You give AI its name and request a research brief. Three minutes later, three pages cover its problem, revenue, token allocation, and risks, with tidy sections, restrained language, and nine links.

You share it. Ten minutes later, a colleague asks:

“Where did you find the 0.05% fee?”

The link points to a Chinese commentary citing official documentation. The official page now gives another number, updated months earlier after a governance vote changed the parameter.

You open the other eight links. Three open and support their claims; two open but never discuss them; two lead to the same secondary article; one returns 404.

Nothing sounds obviously absurd. The brief reads more professionally than what you could write in two hours. But you cannot tell which statements are true.

The chapter's question: what makes AI conclusions trustworthy?

AI can certainly accelerate research. The practical question is how to separate facts, outdated information, and plausible guesses in mixed output.

Keeping Up with Crypto introduced a six-step pipeline. This chapter addresses its hardest part: labeling the confidence of each conclusion separately.

Thought Experiment

Three researchers investigate the same unfamiliar protocol for one hour.

A works manually, reading official docs, governance, and an explorer. Six conclusions arrive, each traceable to a paragraph; two explicitly note ambiguity.

B delegates everything to AI, then fixes typos. Eight minutes of actual work produces twenty-three conclusions, including compliance, competition, and team information A omitted.

C uses AI for discovery and verifies personally. AI lists twelve questions; C checks primary sources, supplies passages, and asks for standardized tables. Eleven conclusions each include a link and the exact supporting passage.

Someone familiar with the protocol checks them.

B produced more correct statements, fourteen, than either other researcher. Acknowledge this: otherwise the experiment becomes anti-AI rhetoric. B genuinely produced the most correct information in eight minutes.

Then observe week two.

The report informs a decision. Someone cites one of B's five incorrect claims. Three weeks later, a problem reveals its origin.

The distinction becomes clear: A and C can locate any conclusion's source; B cannot. Fourteen statements are correct, but B does not know which fourteen.

The lesson is more specific than AI unreliability:

A report without per-claim confidence is only as useful as its weakest statement.

Readers cannot select reliable claims, so must trust or reject everything. The cost of complete trust appears only after harm occurs.

Your Decision

You have two hours to prepare an investment-meeting brief. How?

What Happens

Every sentence in a report has a different reliability and must be treated separately.

Statements fall into three categories:

Counterintuitively, people rely most on AI for the least reliable research tasks.

They use it to learn facts quickly, where errors are common, but hesitate to seek judgments, where scrutiny of reasoning naturally creates value.

Definition errors are harder to detect than inventions, because real numbers answer the wrong question: incentives counted as revenue, FDV as market cap, or bridged assets counted twice. Chapter 19 and Chapter 26 addressed this; now make it a standard AI-output check.

Change one small part of the workflow:

Give every conclusion its own confidence, rather than one confidence shared by the whole report.

Build a Model

1. Delegate According to Whether Errors Are Detectable

Ask whether you can detect a mistake, rather than simply whether AI performs well.

Do not ask AI to summarize a document you do not intend to read. Summaries help decide whether to read, rather than replace reading.

2. Six Steps, Emphasizing the Last Two

Keeping Up with Crypto gave collection, extraction, comparison, questioning, verification, and monitoring. For a specific project, use this version:

Step five is new here and easily skipped. B's fourteen correct claims illustrate the consequence: nobody knows which they are.

3. Four Evidence Levels

Avoid complex scores. Four levels imply distinct actions:

E2 cannot enter paragraphs read by others. Verify it into E0 or state that it remains unknown.

Reports become shorter. That is appropriate. Eleven E0/E3 conclusions are usable; twenty-three mixed with E2 require redoing.

4. Conclusion Cards

Use a fixed format so every claim is independently inspectable:

Conclusion:      The protocol charges X% per trade; Y% goes to the treasury
Evidence:        E0
Source:          Official Fees section / governance proposal N
Observed text:   Paste up to three lines
Collection date: Specific date
Definition:      Protocol fee, excluding frontend markups
If wrong:        Every estimate in the protocol-revenue section changes

Definitions prevent the second error category; impact tells readers importance. Minor mistakes and report-invalidating mistakes deserve different treatment.

AI can fill earlier fields; you supply observed text and definitions. Those two cannot be outsourced.

5. Report-Level Self-Check

Ask three questions:

1. What remains without E2? Too little indicates editing rather than research.
2. What levels support the top three conclusions? E1/E2 foundations require conditional wording rather than certainty.
3. What did I seek but fail to find? Unavailable evidence is itself a finding, such as undisclosed revenue definitions; report it rather than leaving blanks.

What It Is Called

Each action has a name.

Summarize:

AI handles collection and hypotheses; facts return to primary sources and on-chain data.

Grade every claim and exclude unchecked claims from conclusions.

Hands-on

I am researching a protocol of <category>. Before I provide specifics,
list twelve things necessary to evaluate this category.

For each give:
1. A specific fact or number to check, not “understand its mechanism”
2. Typical source: documentation section, governance, chain, or audit
3. What inability to find it implies

Name no projects and guess no numbers.

This is text copied from <source>. Using only this passage, extract:
Name / value / unit / effective date / source subsection.

For missing fields write “not stated in the source.” Do not infer or complete them.

Text:
<paste>

Compare these fields for three protocols, adding the largest difference.
Use only supplied information; mark missing cells “not researched.”

□ Decimal numbers         → Compare digit by digit
□ Protocol revenue        → Definition, incentives, collection time
□ Secondary claims        → Follow through to primary documents

Variable:             For example, governance changes a parameter
Change location:      Forum / contract event / official announcement
Invalidated claim:
Check frequency:

AI Lab

Give fifteen specific conclusions about <protocol>.

Requirements:
1. An accessible source link and exact section or paragraph for each claim.
2. Evidence type: primary document / governance / on-chain data /
   secondary analysis / inference.
3. Unit, definition, and collection time for every number.
4. Uncertainties in a separate final section, not hidden in vague body wording.
5. Do not pad with generic industry knowledge. Eight specific claims beat
   fifteen correct but useless generalities.

Total claims:
Inaccessible links:
Accessible links not supporting claims:
Secondary sources:
Incorrect definitions or missing timestamps:
Fully verified, E0:

Total verification time for fifteen claims:       minutes

Real-World Case

Change One Variable

Takeaway Questions

Chapter Check