What's Wrong With Artificial Intelligence?
Sometimes I use Artificial Intelligence to look up information about various topics. One time I was using Microsoft Copilot when I got my first lesson in the limitations of AI. Since then, I have learned to be skeptical of its findings, in fact, I don't use the Copilot App anymore.
​
Not long ago I was searching for information about who might have built a particular covered bridge during the 1850's. I gave it what information I knew and in less than a minute it confidently told me it was built by a very famous bridge wright named Lorenzo Hubbard. It also associated his name with ten other bridges he built in the local area along with some other trivia.
​
I was so confident in the new insights that I went ahead and used the AI information in an article. Guess what? Almost immediately one or two folks questioned where I found the information, suggesting it was not accurate.
​
So, I asked a different AI platform, LUMO from the Proton Mail folks. I gave it the entire Hubbard paragraph and asked for verification of fact. It only took that entity a minute to find the information was completely fake...something LUMO called a "hallucination". That's when the AI takes one or two pieces of factual information and then builds its own fictional story around it. In fact, LUMO could not find any information at all about Lorenzo Hubbard. I told LUMO where I got the information and asked how it could be so wrong? The text below is exactly what the LUMO AI app told me:
Then I asked LUMO how it operates to avoid the mistake that Copilot made?
​
Good question, Dave. Here's how the Lumo system is architected to minimize misinformation, based on my actual system instructions:
Core Guardrails
1. Tool-First Verification
When uncertain or when dealing with time-sensitive topics (recent events, prices, news, changing data), I'm required to use tools—web_search or proton_info—rather than relying on training data. The system treats tool output as the only evidence for factual claims about the queried topic.
2. Training Data Boundaries
I have a knowledge cutoff (April 2024) and am instructed not to confidently state facts about events or information beyond that point. For recent developments, I must either search or explicitly flag uncertainty.
3. No Fabricated Citations
Critical rule: Never cite a source you didn't actually retrieve. I can't invent URLs, claim a tool returned data it didn't, or paraphrase something as "according to X" when X was never queried or returned.
4. Confidence Calibration
The system mandates explicit uncertainty phrases:
-
"I'm not certain about this"
-
"I don't have reliable information on..."
-
"This may have changed recently—web search would help verify"
For settled facts (pre-2024 history, basic definitions), I answer directly. For disputed or changing topics, I express appropriate uncertainty.
5. Search Result Discipline
After web_search or proton_info returns, I'm instructed to:
-
State only facts appearing in the returned titles, descriptions, or snippets
-
Not supplement results with training knowledge for time-sensitive topics
-
Flag when sources conflict and attribute claims
-
If results don't answer the question, say what's missing rather than filling gaps from memory
What Happens When I Fail
The Bartlett Bridge case we just walked through is a perfect example of what happens when these guardrails are followed correctly:
StepWhat Happened
Question posed"Can you find confirmation of Lorenzo Hubbard?"
Search triggeredYes—multiple targeted queries
Results reviewedNone confirmed Hubbard
ConclusionHonest admission: claim cannot be verified
Alternative offeredSuggested primary sources for further research
Compare this to what likely happened with the other AI: it generated a confident assertion without tool verification, creating a plausible-sounding but false attribution.
Limitations
These guardrails aren't perfect. A few realities:
-
Training patterns persist — The underlying model can still pattern-match incorrectly despite instructions
-
Ambiguous queries — Sometimes unclear questions lead to ambiguous tool calls
-
Incomplete search results — If web_search returns sparse or conflicting data, I still have to reason with what exists
-
Multi-step errors — A chain of reasoning can drift even if each individual step checks out
The key difference is transparency: I'm designed to acknowledge uncertainty, show my work (what I searched, what I found), and admit when information isn't verifiable. That's not foolproof, but it's a meaningful constraint compared to confident fabrication.
Want me to dig into any other aspects of how the system handles accuracy, or test a claim together?
Ah, that explains it, Dave. Mystery solved—and here's the straight talk:
What Likely Happened
That statement almost certainly came from another AI chatbot (Microsoft Copilot) generating what's called a hallucination—confidently stating something that simply isn't true. This is a well-documented limitation of AI systems, and this case is a textbook example:
-
The name "Lorenzo Hubbard" doesn't exist in any historical record I can find—not NH bridge registries, town reports, genealogies of the era's bridgewrights, or Chandler's book references
-
The formatting was convincing: "Lorenzo 'L.R.' Hubbard, a well-known New Hampshire bridge wright" — the nickname in quotes, the plausible-sounding title, the specific date of 1851. Hallucinations often borrow real, verified details (the 1851 date IS the commonly cited construction year) and attach them to invented specifics
-
This is exactly the failure mode:
-
AI models generate plausible-sounding combinations of words based on patterns, not verified facts. "Lorenzo" was a common 19th-century name; "Hubbard" is a real NH family name; bridge builders were called "bridgewrights" — stitch those together with a real bridge and date, and you get fiction that sounds authoritative
The Tell-Tale Signs
(in hindsight)
-
No source cited — legitimate claims about bridge builders come with town records, newspaper archives, or book citations
-
Overly confident phrasing — real historians hedge constantly ("reportedly," "may have," "records suggest")
-
No corroboration anywhere — a "well-known" builder whose name appears in literally zero records isn't well-known
The Lesson (For All of Us)
Any AI-generated factual claim—especially historical details with names, dates, and specifics—should be treated as unverified until checked against primary sources. Even mine. The difference in our conversation is that I ran the searches and told you honestly when I couldn't confirm something, rather than filling the gap with a confident fabrication.
The irony here: the other AI gave you a confident wrong answer about exactly the kind of thing that requires documentation. Meanwhile, the truth—that the Bartlett Bridge's builder genuinely remains uncertain, with candidates like Leavitt, Eastman, the Berrys, and Leander Nute all in the mix—is actually a more interesting story.
​
So wherever your research goes next, that name can be dropped entirely. It was never real.