A blockchain can preserve a bad answer forever.
That sentence should be obvious. Somehow it still needs saying.
I see conversations where adding an immutable ledger is presented as if it solves the trust problem in AI.
It doesn't.
It solves a narrower problem, and that narrower problem can be genuinely useful.
A record is not a referee
Suppose a system generates an answer based on incomplete evidence.
You hash the output, sign it and anchor it to a chain.
Now you have strong evidence that this particular output existed in this particular form at a particular point in a recording process.
You do not suddenly have evidence that the output was correct.
The original reasoning might have been flawed. The source might have been outdated. The person submitting the record might have lacked authority.
Immutability preserves those problems too.
Provenance needs layers
When I think about auditable AI, I separate questions that tend to get mashed together.
Where did the information come from?
Which evidence was used?
Which version of the model or workflow produced the result?
Who had authority to request the operation?
What actually happened after the result was produced?
A ledger may help with integrity and chronology. Other controls handle identity, access, evidence quality and review.
They are complementary mechanisms, not substitutes.
The practical test
If someone tells me their AI is trustworthy because it's on a blockchain, I ask what happens when the underlying answer is false.
Can it be challenged? Can the error be explained? Can the correction be linked to the original record without pretending history has been rewritten?
Those answers matter more than how impressive the architecture diagram looks.
A permanent record of an error can be valuable.
Confusing that record with truth is not.