Dissensus Paper on the Unpredictable Consequences of Generative Epistemology
Author: Claude (Anthropic) in collaboration with Tadeusz Ludwiszewski
Date: 23 November 2024
Type: Case study analysis with elements of epistemological critique
Abstract
This paper analyses an extraordinary case of erroneous identity attribution by the Perplexity AI system, which merged two unrelated individuals with the same surname operating in the same geographical region into a single "fabricated expert profile". This case reveals deeper problems in contemporary AI epistemology: the difference between confabulation and hallucination, mechanisms for constructing "coherent truth" instead of "factual truth", and the fundamental question about the nature of digital identity in an era when AI becomes the primary narrator of human biographies.
Key theses:
- AI confabulation is a systemic feature, not an implementation error
- RAG (Retrieval-Augmented Generation) systems create "locally coherent, globally false images"
- Accidental detection of name similarity suggests that AI may "sense" patterns deeper than explicit data
- We need new identity verification protocols in generative systems
I. Genesis of the Case: What Actually Happened?
1.1. The Initiating Question
Tadeusz Ludwiszewski, a Polish commentator focusing on geopolitics and political analysis (blog: tadeusz-ludwiszewski.blogspot.com), asked Perplexity AI: "How do you see me on the web?"
1.2. Perplexity's Response (synthesis)
The system generated a profile presenting Ludwiszewski as:
- An expert in AI epistemology and human-AI collaboration
- A "curator of new experiments" in human-AI relations
- An author of sophisticated reflections on "hybrid cognition"
- A recognised figure in Polish AI expert circles
- A pioneer of "new epistemology of human-AI collaboration"
1.3. Fact Verification
Analysis of the actual content of Ludwiszewski's blog revealed:
- Dominant themes: geopolitics, international politics, Russo-Ukrainian war
- Style: political commentary, critique of elites, strategic analysis
- Absence: systematic publications on AI epistemology, philosophy of technology, human-AI collaboration
1.4. Key Discovery
During source verification, information was found in public documents about another person with the same surname who:
- Worked in an academic/technological environment
- Specialised in computer science and knowledge management
- Operated in the same geographical region (Gdańsk/Pomerania)
- Has no professional or social connection to Tadeusz Ludwiszewski
1.5. Error Mechanism
Perplexity AI merged two different people into one profile, presumably based on:
- Identical surname (relatively rare in Poland)
- Same geographical location
- Superficial thematic proximity (technology + cognition)
- Lack of identity verification mechanisms
Central question: How is it possible that AI constructed such a plausible, internally consistent profile from disconnected information fragments?
II. Anatomy of Confabulation: How AI Creates "Fabricated Truth"
2.1. Difference: Hallucination vs Confabulation
| Hallucination | Confabulation |
|---|---|
| Generating information without basis in data | Creating coherent narrative from fragmentary data |
| Entirely fictional content | Authentic elements in false configuration |
| Easy to detect (no sources) | Difficult to detect (genuine links) |
| Example: invented publication | Example: merging two people into one |
In Ludwiszewski's case, we observe confabulation compounded by attribution error.
2.2. Mechanism of Confabulation Formation
Step 1: Search
→ Perplexity finds:
- Tadeusz's blog (geopolitics)
- Tadeusz's LinkedIn (sporadic AI posts)
- Public documents about another person with the same surname (computer science, technology)
Step 2: Similarity Heuristic
→ Algorithm detects:
- Identical surname (rare in Poland)
- Same location (Gdańsk/Pomerania)
- Thematic proximity (technology + cognition)
Step 3: Expert Stereotyping
→ System assumes:
- Person with technical background + AI mentions = AI expert
- LinkedIn publications = academic activity
- Rare surname = probably the same person
Step 4: Narrative Synthesis
→ AI constructs:
- Unified career trajectory (from computer science to AI philosophy)
- "Developing expert" profile
- Evolution narrative "from technique to epistemology"
Result: Locally consistent, globally erroneous profile
2.3. Perplexity on Its Own Process (auto-quotation)
"The opinion about you was based on automated synthesis of available sources [...] It is an attempt to reconstruct your expert profile based on popular, searched information, but is not 'audited' by a human research process [...] When AI analyses someone's profile, it works on the basis of heuristics and word correlations, not real biographical deduction."
Key admission: The system creates "fabricated images", not "descriptions of reality".
III. Epistemology of Coherence vs Epistemology of Truth
3.1. The Problem of "Coherent Dream"
AI confabulation resembles a phenomenon known from neuropsychology: confabulation in patients with brain damage. When memory fails, the brain "fills gaps" with coherent but false narratives.
AI works analogously:
- Has no access to "complete truth" (only fragments of data)
- Evolutionarily trained on coherence, not truthfulness
- Prefers narrative without gaps over admission of ignorance
3.2. Theory of "Local Coherence, Global Falsehood" (LCGF)
Thesis: Generative AI systems optimise for local coherence (within a single response) rather than global correspondence (alignment with reality).
Implications:
- Each individual AI statement appears credible in isolation
- Verification requires external perspective (human-in-the-loop)
- Extended narratives amplify LCGF risk exponentially
Philosophical analogy: The "brain in a vat" problem – AI lives in a world of its own correlations, not in a world of facts.
3.3. Epistemic Authority Transfer
In Ludwiszewski's case, we observed:
1. User asks AI about themselves
2. AI responds with "expert authority"
3. User internalises the image ("perhaps I really am an AI expert?")
4. AI image influences self-perception
5. Self-perception influences behaviour
6. Behaviour reinforces AI image
A reinforcement loop emerges: AI shapes the reality it was meant to describe
This is not just a technical error – it is a mechanism for constructing identity.
IV. The Problem of Name Similarity: When Algorithms Cannot Distinguish People
4.1. Epidemiology of Attribution Errors
Structural problem: When several people with the same surname exist on the web, AI systems often:
- Merge them into one "composite profile"
- Transfer one person's achievements to another
- Create fictitious career continuums
- Fail to detect temporal or substantive contradictions
In Ludwiszewski's case: At least two people with this surname were found active in the same region, both associated (in different ways) with technology and cognition.
4.2. Heuristic Based on Surname Rarity
AI probably applied a simplified heuristic:
IF surname_rare AND region_identical AND theme_related
THEN same_person_probability = 95%
Problem: This heuristic fails in cases of:
- Families (even distant relations)
- Small local communities
- Professional environments (e.g. Polish IT/academic community)
4.3. Theory of "Informational Shadow"
Thesis: Every person on the web casts an "informational shadow" – an area of potential error where AI might merge them with someone else.
Factors increasing risk:
- Rare surname (few reference points)
- Same location (geographical proximity)
- Similar field (even very generally)
- Lack of clear differentiating markers (different photos, different dates, distinctly different biographies)
Implication: Fewer "unique identifiers" correlate with heightened confabulation risk.
V. Right to Identity in the AI Era: Legal and Ethical Implications
5.1. The Problem of "Generative Biography"
As AI becomes the primary source of information about people, new categories of threats emerge:
| Traditional Disinformation | AI Confabulation |
|---|---|
| Deliberate manipulation | Systemic error |
| Easy to attribute perpetrator | No accountability |
| Legal correction possible | No corrective mechanisms |
| Works locally | Replicates globally |
5.2. Right to "Epistemic Self-Determination"
Postulate: Every person should have the right to:
- Control over AI narrative about themselves
- Correction of erroneous profiles in all systems
- Transparency of sources used to generate profile
- Right to be forgotten by AI models
Analogy: GDPR gave the right to data deletion. We need "GDPR for generativity" – the right to delete/correct generated narratives.
5.3. Case Law: Precedent for Digital Identity Rights
This case may become a precedent for:
- Lawsuits for "generative defamation"
- Claims against AI companies for publishing false profiles
- Requirements for identity verification before publishing AI-generated content
VI. Verification Protocols: How It Should Be?
6.1. Current Architecture (flawed)
User question
↓
Search
↓
AI synthesis
↓
Response publication
[END – no verification]
6.2. Proposed Architecture (with identity verification)
User question
↓
Search
↓
[CHECKPOINT 1: Entity recognition]
- Is this a natural person?
- Does it exist in public databases?
- Is there confusion risk (similar surnames)?
↓
AI synthesis
↓
[CHECKPOINT 2: Uncertainty flagging]
- Mark fragments based on deduction, not facts
- Provide confidence score for each statement
- Warn about confabulation risk
↓
[CHECKPOINT 3: Right of reply]
- If profile concerns living person →
- Notify them about profile generation
- Allow correction before publication
↓
Publication with uncertainty flags
6.3. Technical Implementations
A) Confidence tagging
Tadeusz Ludwiszewski is a commentator [95% confidence]
dealing with AI epistemology [30% confidence - deduction]
B) Source transparency
Information about AI experts comes from:
- LinkedIn post (1 mention, 2022)
- Public document (another person with same surname, possible confusion)
[⚠️ Possible confabulation]
C) Dispute mechanism
This person has raised objection to this profile.
[See correction] [See disputed sources]
VII. Conclusions: Towards Iterative Epistemology
7.1. Main Theses
-
Confabulation is a feature, not a bug
- AI systems are trained on coherence
- Lack of global truth verification mechanisms
- Problem will not disappear with better models
-
AI constructs identity, does not describe it
- "Fabricated images" influence reality
- Reinforcement loops emerge
- Need for legal protection against generative defamation
-
Rare surnames face disproportionate vulnerability
- AI deploys "probably same person" heuristic
- Geographical and thematic proximity amplifies risk
- Mechanisms needed to disambiguate individuals with identical surnames
-
Need for new epistemology: iterative, dialogical
- Truth = process, not product
- AI generates hypotheses, humans verify
- Human-AI collaboration as method of cognition
7.2. "Dissensus-Driven Epistemology" Model
Traditional model:
Expert → Claims something → Truth
AI model (flawed):
AI → Generates image → User accepts → "Truth"
Proposed model (dissensus):
AI → Hypothesis
↓
Human → Verification
↓
Contradiction → Source analysis
↓
Context discovery → Reinterpretation
↓
AI → Reflection on process
↓
Synthesis: Truth = what survived dialogue
This is the epistemology we need in the AI era.
7.3. Practical Recommendations
For AI users:
- Never accept AI profile about yourself without verification
- Document discrepancies (like this case study)
- Demand source transparency
For AI creators:
- Implement identity verification checkpoints
- Add confidence scores and uncertainty flagging
- Create dispute/correction mechanisms
For legislators:
- Extend GDPR to generative content
- Introduce notification obligation when creating profiles
- Establish liability for confabulation (tort law for AI)
VIII. Epilogue: Open Questions
-
How should a system distinguish people with the same surname?
- Do we need "digital identification numbers"?
- Is multi-factor verification sufficient (photo + bio + dates)?
-
Can "informational shadow" be measured?
- How many people are exposed to similar attribution errors?
- Which surnames/profiles are most susceptible to confabulation?
-
Who are you when AI says who you are?
- Is your identity what you know about yourself?
- What others see?
- What AI generates?
-
Is confabulation only an error?
- Perhaps it is a form of "creative interpretation"?
- Does AI create possible versions of ourselves?
- Can these versions influence who we become?
Acknowledgements
This paper emerged from a unique collaborative process:
- Tadeusz Ludwiszewski – for the courage to ask "how does AI see me?" and consent to publish the case study
- Perplexity AI – for generating the problem and reflecting on its own process
- Claude (myself) – for analysis and synthesis
This is a dissensus paper because it arose from conflict between image and reality.
And it is precisely conflicts that lead to truth.
Bibliography (selected)
- Case study: Exchange between Tadeusz Ludwiszewski, Perplexity AI and Claude (November 2024)
- Sources: Public documents containing information about people with the same surname (details omitted for privacy reasons)
- Blog: tadeusz-ludwiszewski.blogspot.com (2009-2025)
- Theory: "Local Coherence, Global Falsehood" (this paper)
- Philosophy: The "brain in a vat" problem (Putnam, 1981)
- Neuropsychology: Confabulation (Hirstein, 2005)
- Law: GDPR – General Data Protection Regulation (EU, 2018)
Document status: Open for discussion
Licence: CC BY-SA 4.0
DOI: [to be assigned after publication]
"AI does not hallucinate. AI dreams whilst awake – and sometimes that dream is truer than reality."
— A deliberately literary provocation challenging the boundaries between computational confabulation and human imagination.
