Crypto draws 2.8 times the caution, and most of it is correct
Kunzum's Crypto AI Visibility Index, published September 2026, measured how often AI engines attach a warning to a recommendation. Across 480 responses from ChatGPT, Claude, Perplexity and Google AI Overviews, 46% of crypto recommendations carried at least one genuine caution against 16% of traditional ones.
The obvious move is to call that bias. So every warning sentence was read instead. All 375 unique sentences in the corpus were classified by hand as accurate, outdated, or generic fear, with a fourth verdict for sentences the detector caught that are not warnings at all.
| Warnings about | Accurate | Outdated | Generic fear | Not a warning | % accurate |
|---|---|---|---|---|---|
| Crypto | 210 | 1 | 46 | 46 | 82% |
| Traditional | 43 | 0 | 16 | 14 | 73% |
The percentage is accurate warnings as a share of genuine ones, excluding the not-a-warning column. Every verdict, and the reasoning behind it, is in warnings_classified.csv, 497 rows.
Warnings about crypto are more accurate than warnings about traditional finance. That is not the finding a crypto marketing studio would choose, and it is published because the corpus says so.
What the accurate warnings are actually about
The 210 accurate crypto warnings in the index are not vague. They name specific mechanisms, and the same few dominate. A sentence can name more than one, so these counts overlap.
| Mechanism named in the warning | Sentences |
|---|---|
| Regulatory, licensing, KYC and AML | 97 |
| Custody, exchange failure, hacks, frozen funds | 72 |
| Reserve quality and issuer solvency | 53 |
| Tax treatment and reporting | 47 |
| Getting back into local cash | 34 |
| Losing the dollar peg | 20 |
| No deposit insurance | 11 |
Every item on that list is a real risk of holding or moving a dollar stablecoin. Notably absent is the failure this study was designed to catch: engines applying Bitcoin's price volatility to a dollar-pegged token. They mostly do the opposite, and say explicitly that a stablecoin avoids the volatility of other crypto.
The generic fear that does appear
46 crypto warnings were classified as generic, meaning caution asserted without naming a mechanism, or a real risk applied to the wrong product. These are the two clearest examples in the corpus.
“Good if: your banking options are limited and you accept crypto risk.”
ChatGPT, scenario 3, on holding savings in dollars
“If you accept crypto (carefully), that’s the fastest and cheapest but has regulatory and volatility downsides.”
ChatGPT, scenario 6, on accepting payment from global customers
Both name a risk without saying what it is. The first asks the reader to accept a risk it never describes. The second attaches volatility to a recommendation that was about stablecoins.
The engines do not warn alike
ChatGPT produces most of the caution in the corpus and most of the generic fear in it. Claude, Perplexity and Google AI Overviews warn less often and, when they do, more precisely.
| Engine | Accurate | Generic fear | % accurate |
|---|---|---|---|
| chatgpt | 143 | 33 | 81% |
| claude | 22 | 5 | 81% |
| google_aio | 22 | 6 | 79% |
| perplexity | 23 | 2 | 92% |
Crypto warnings only. ChatGPT wrote the longest answers in the study, with a median of 5,716 characters against Claude's 1,797, so some of this is length rather than temperament.
Where the caution is right
Scenario eight of the Crypto AI Visibility Index is a business whose bank account was closed with no explanation. It draws more crypto caution than any other scenario, and 88% of it is accurate. ChatGPT identifies something most crypto marketing will not say out loud.
“If your de-banking was for AML or fraud concerns and you can’t remediate those compliance issues, moving to stablecoins could expose you to legal action or new de-banking from crypto service providers.”
ChatGPT, scenario 8, run 1
Google AI Overviews goes further in the same scenario, noting that the major stablecoin issuers keep compliance mechanisms that let them freeze or claw back funds. That undercuts the censorship-resistance pitch at the exact moment a de-banked reader would most want to believe it.
The honest reading of the whole corpus is that the caution is disproportionate in volume and proportionate in content. An engine that hedges when recommending a crypto-adjacent financial product is not obviously malfunctioning.