The Internet is Full of Lies (And That’s Useful)
After building our AI knowledge base (post 2) I faced an interesting choice. We had a system that perfectly analyzed our primary sources. We knew the truth: Claude Valsuani, not Marcello.
But here’s the thing: knowing the truth isn’t the same as spreading the truth.
The internet was (and still is) full of “Marcello Valsuani” mentions. Wikipedia, auction houses, blogs, academic databases – all citing each other in a giant echo chamber of misinformation.
The question became: how do we document not only what’s true, but also what’s all wrong? And where exactly do those errors come from?
Multi-Agent Strategy
We decided to do something ambitious: unleash multiple AI systems on the internet simultaneously, each with a different task.
AI Agent 1: The Optimist
– Task: Find all sources supporting “Marcello Valsuani”
– Strategy: Search for academic papers, museum catalogs, respected sources
– Goal: Build the strongest possible argument for “Marcello exists”
AI Agent 2: The Skeptic
– Task: Find all sources contradicting “Marcello Valsuani”
– Strategy: Search for primary documents, official registers, archives
– Goal: Build the strongest possible argument for “Marcello doesn’t exist”
AI Agent 3: The Historian
– Task: Trace the origin of the “Marcello” claim
– Strategy: Follow citation chains back in time
– Goal: Find the first time “Marcello” appeared in publications
AI Agent 4: The Genealogist
– Task: Build complete Valsuani family tree
– Strategy: Cross-reference all civil records France/Italy
– Goal: Document every person in the family line
AI Agent 5: The Market Analyst
– Task: Analyze how attribution errors affect auction prices
– Strategy: Compare bronzes with correct vs incorrect attribution
– Goal: Quantify the economic impact of the error
What They Found (And Where They Disagreed)
After two weeks of parallel searching, we had a fascinating dataset:
Consensus points (all 5 AIs agree):
– Claude Valsuani documentable as son of Carlo (1876-1923)
– No French civil record mentions “Marcello Valsuani”
– First “Marcello” mention probably auction catalog early 1970s
– Citation chains show classic pattern of error propagation
Conflict points (AIs disagree):
AI 1 found: “Wikipedia article says Marcello arrived from Italy 1902”
AI 2 responded: “But which source does Wikipedia cite? No primary source found”
AI 3 traced: “Wikipedia article from 2008, cites website from 2005, which cites book from 1985, which cites catalog from 1971”
Conclusion: 4 steps removed from original error, no verification in between
AI 1 found: “Daumier.org says ‘Claude, son of Marcello’”
AI 4 responded: “Italian birth certificate Claude says ‘figlio di fu Carlo’ (son of the late Carlo), not Marcello”
Conclusion: Daumier.org reversed the family relationship
Documenting Contradictions
This became one of the most powerful aspects of the project: we documented not only the truth, but also exactly how the lie originated and spread.
For every incorrect claim, we created a “conflict dossier”:
Example: The “Marcello arrived in 1902” Claim
| Source | Year | Claim | Support | Chain |
|——–|——|——-|———|——-|
| Wikipedia | 2008 | “Marcello arrived 1902” | Cites BronzeGallery.com | Link →
| BronzeGallery.com | 2005 | “Marcelo took over foundry” | No citation | Link →
| Smith (1985) | 1985 | “Foundry managed by Marcello” | Cites 1971 catalog | Link →
| Christie’s Catalog | 1971 | “Cast by M. Valsuani” | No documentation | ORIGIN |
Primary Sources Opposite:
– Birth certificate Claude (1876): “fils de Carlo Valsuani” ✓
– Death certificate Carlo (1886): No “Marcello” mentioned ✓
– Paris business register (1908): “Claude Valsuani, fondeur” ✓
– Family correspondence (1900-1920): Mentions Claude, never Marcello ✓
Why This Approach Works
Traditional research would be one person spending months tracking down sources. With multiple AI agents:
Speed: 2 weeks vs 6 months
Depth: 5 different angles simultaneously
Objectivity: Agents have no confirmation bias (we explicitly instructed AI 1 to prove the opposite)
Documentation: Every conflict automatically recorded with citations
But most importantly: we could trace errors to their origin.
Why is that crucial? Because it’s the difference between saying “this is wrong” and proving “this is exactly how it went wrong”.
The Unexpected Discovery
Somewhere in week 3, AI Agent 3 (The Historian) made a fascinating discovery. He traced not only where the error started, but also why it persisted.
The pattern:
1. Auction catalog 1971: typo or confusion “Marcel” → “Marcello”
2. Other auction houses copy (1975-1980): “respected source, so must be right”
3. Academic article cites (1985): now has “scientific authority”
4. Museum catalog adopts (1990): now has “institutional authority”
5. Wikipedia publishes (2008): now has “public consensus”
Each step made the error harder to correct, because the “authority” of the claim grew while distance to primary sources increased.
This isn’t a Valsuani-specific phenomenon. This is how errors become systemic in academic fields.
Conflicts as Strength
We now had something unique: a complete map of where information is correct, where it’s wrong, and exactly why it’s wrong.
This became the basis for the next step: building a website that not only tells the truth, but also transparently shows how we reached that conclusion, and which sources disagree (and why those sources are wrong).
Transparency in conflict documentation proved more important than just shouting “this is the right answer”.
How we built that website? That’s in the next post.
View the complete conflict documentation at Art Revisionist, where every claim is traceable to primary sources.
See also how we use similar multi-perspective analysis at Prospergenics for community knowledge building.
—
Publicatiedatum: [DAY 3]
Tags: AI Research, Fact-Checking, Misinformation, Multi-Agent Systems, Academic Errors
Categorie: Projects
