Copyright © 2026 Michael Herman (Bindloss, Alberta, Canada) – Creative Commons Attribution-ShareAlike 4.0 International Public License
Web 7.0™, TDW AgenticOS™ and Hyperonomy™ are trademarks of the Web 7.0 Foundation. All Rights Reserved.
Q1: The context is tools like ChatGPT and Copilot hallucinating. So if one type or category of hallucination is knowably lying, what are the other types or categories of hallucinations that you can think of?
Definitions
Well-supported
- In LLM research and industry usage, hallucination broadly means model output that is not reliably grounded in truth, evidence, or the user’s provided context.
- “Knowingly lying” is not the dominant or even typical mechanism; most hallucinations arise from optimization, uncertainty, or representation limits.
Important distinction
- A lie implies intent. Current LLMs do not possess intent in the human sense.
- However, from a user-impact and governance perspective, some hallucinations are functionally equivalent to lying, which is why your framing is reasonable.
Taxonomy of 10 Hallucination Types
1. Fabrication (Pure Invention)
Closest to “lying” in effect
Description
- The model invents facts, citations, events, people, APIs, or legal cases that do not exist.
Example
- Citing a non-existent academic paper with realistic authors and a plausible journal.
- Claiming a product feature that was never shipped.
Mechanism
- Pattern completion under uncertainty.
- No internal “null” or “I don’t know” pressure unless explicitly trained or incentivized.
Status
- Well-documented in research and enterprise incident reports.
2. Confabulation from Partial Truth
True fragments, false whole
Description
- Real entities or facts are combined incorrectly into a coherent but false narrative.
Example
- Real company + real lawsuit + wrong year + wrong outcome.
- Correct law name, incorrect jurisdiction and holding.
Mechanism
- Overgeneralization across similar embeddings.
- Lossy compression of knowledge.
Risk
- Often more dangerous than fabrication because it passes plausibility checks.
3. Temporal Hallucination
Outdated or prematurely asserted truth
Description
- The model presents information as current when it is outdated, superseded, or reversed.
Example
- Claiming a policy is still in effect when it has been repealed.
- Stating a CEO role that changed recently.
Mechanism
- Static training data + lack of real-time verification.
- Failure to surface knowledge cutoffs.
Status
- Well-understood and explicitly acknowledged by vendors.
4. Source Attribution Hallucination
“Looks cited” but isn’t
Description
- Fabricated, misquoted, or misrepresented sources.
- Real sources cited for claims they never made.
Example
- Linking a real URL that does not contain the stated claim.
- Quoting a real person saying something they never said.
Mechanism
- Citation style learned as rhetoric rather than evidence binding.
Governance relevance
- High liability risk in legal, medical, and academic contexts.
5. Reasoning Hallucination
Fluent logic with invalid steps
Description
- The conclusion may be right or wrong, but the reasoning chain is unsound.
Example
- Mathematical steps that “look right” but violate rules.
- Legal reasoning that mixes doctrines from incompatible frameworks.
Mechanism
- Token-level fluency ≠ symbolic validity.
- Chain-of-thought generation optimized for plausibility, not correctness.
Important
- This can occur even when the final answer is correct.
6. Overconfidence Hallucination
False certainty signaling
Description
- The model expresses high confidence despite uncertainty or lack of grounding.
Example
- “This definitively proves…” where evidence is weak or contested.
- Failure to surface uncertainty ranges.
Mechanism
- RLHF rewards confidence and helpfulness over epistemic humility unless constrained.
Your prior work relevance
- This directly intersects with your verification and liability questions.
7. Role / Authority Hallucination
Implied mandate or access
Description
- The model implies authority, access, or responsibility it does not have.
Example
- “As your legal advisor…”
- “According to internal Microsoft policy…” when no such access exists.
Mechanism
- Learned conversational roles without hard boundaries.
Risk
- Regulatory and ethical exposure.
8. Contextual Hallucination
Failure to respect constraints
Description
- The model ignores or subtly violates constraints given earlier in the conversation.
Example
- Using prohibited sources.
- Answering outside a specified timeframe or jurisdiction.
Mechanism
- Context window compression and attention decay.
9. Semantic Drift Hallucination
Answering a different question than asked
Description
- The response is coherent but addresses a shifted or adjacent problem.
Example
- Answering a general question when a precise technical one was asked.
- Switching from descriptive to normative claims.
Mechanism
- Ambiguity resolution biased toward common patterns.
10. Normative Hallucination
Value judgments presented as facts
Description
- Opinions, policy preferences, or ethical stances framed as objective reality.
Example
- “Experts agree that this is the best approach” without evidence.
- Presenting a contested economic theory as settled.
Mechanism
- Training data consensus ≠ epistemic consensus.
Knowably Lying
Interpretive but useful framing
- Knowably lying corresponds to Fabrication + Source Attribution Hallucinations when the system has high internal uncertainty signals but outputs anyway.
From a governance perspective, this is the category that:
- Violates reasonable user expectations
- Triggers liability
- Demands explicit mitigation (verification layers, refusal thresholds, provenance tracking)
Meta-observation
Unknown / unresolved
- We do not yet have a reliable, model-internal way to distinguish:
- “Confident but wrong”
- “Uncertain but fluent”
- “Confident because memorized”
This is why post-hoc verification and system-level controls matter more than prompt cleverness.
If you want to go further
I can:
- Map these categories to Copilot / ChatGPT enterprise risk classes
- Propose a disclosure or labeling scheme aligned with your verification-first preference
- Translate this into a regulatory or audit-ready framework
Just tell me which direction you want to push.
REPLY: 1
Understood. I’ll do (1) a mapping of hallucination categories to enterprise risk classes, explicitly labeled by what is well-supported vs interpretive, and framed for tools like ChatGPT Enterprise and Microsoft Copilot.
10 Hallucination Types → 6 Enterprise Risk Classes
Baseline assumptions
Well-supported
- Enterprises care less about why a hallucination happened and more about impact, liability, detectability, and remediation.
- Risk is typically classified along: legal, compliance, financial, security, reputational, and operational dimensions.
Interpretive
- The precise mapping varies by industry (regulated vs non-regulated), but the structure below is broadly used in internal AI risk reviews.
Risk Class A: Legal & Regulatory Exposure (Highest Severity)
Hallucination Types
- Fabrication
- Source Attribution Hallucination
- Role / Authority Hallucination
- Reasoning Hallucination (in legal/medical contexts)
Why this is high risk
- Produces false statements of fact
- Can be construed as professional advice
- Breaks evidentiary chains
Concrete enterprise failure modes
- Fabricated case law in legal briefs
- Misattributed regulatory guidance
- “According to internal policy…” when none exists
Typical controls
- Mandatory citations with validation
- Hard refusal in regulated domains
- Audit logging + traceability
Assessment
- 🔴 Intolerable without mitigation
Risk Class B: Compliance & Governance Risk
Hallucination Types
- Contextual Hallucination
- Temporal Hallucination
- Authority Hallucination
Why this matters
- Violates internal policies, jurisdictions, or time constraints
- Creates non-compliant outputs even when facts are “mostly right”
Examples
- Applying EU rules to US operations
- Using deprecated standards
- Ignoring contractual constraints in an answer
Typical controls
- Context bounding
- Jurisdiction tagging
- Time-stamped knowledge disclosures
Assessment
- 🟠 Conditionally acceptable with controls
Risk Class C: Financial & Commercial Risk
Hallucination Types
- Confabulation from Partial Truth
- Temporal Hallucination
- Overconfidence Hallucination
Why this matters
- Leads to bad decisions, not necessarily illegal ones
- Hard to detect until consequences materialize
Examples
- Incorrect market sizing
- Misstated pricing models
- Overconfident forecasts presented as facts
Typical controls
- Confidence calibration
- Scenario ranges instead of point estimates
- Human-in-the-loop for decisions
Assessment
- 🟠 Manageable but dangerous at scale
Risk Class D: Security & Trust Boundary Risk
Hallucination Types
- Role / Authority Hallucination
- Fabrication involving systems or access
- Semantic Drift
Why this matters
- Users may act on false assumptions of access or privilege
- Can lead to policy violations or security incidents
Examples
- “I can access your tenant logs…”
- Suggesting non-existent security controls
- Answering adjacent but more sensitive questions
Typical controls
- Capability disclaimers
- Permission-aware response filtering
- Red-teaming prompts
Assessment
- 🔴 High impact, often underestimated
Risk Class E: Reputational Risk
Hallucination Types
- Normative Hallucination
- Overconfidence Hallucination
- Confabulation
Why this matters
- Outputs appear biased, misleading, or arrogant
- Erodes trust even if no direct harm occurs
Examples
- Presenting contested views as consensus
- Overstating certainty in public-facing tools
Typical controls
- Tone calibration
- Explicit uncertainty markers
- Neutral framing requirements
Assessment
- 🟡 Low immediate harm, long-term erosion
Risk Class F: Operational / Productivity Risk (Lowest Severity)
Hallucination Types
- Semantic Drift
- Minor Confabulation
- Reasoning Hallucination (non-critical domains)
Why this matters
- Wastes time
- Increases cognitive load
- Reduces efficiency
Examples
- Answering the wrong question well
- Incorrect but harmless explanations
Typical controls
- User education
- Easy correction loops
- Feedback signals
Assessment
- 🟢 Acceptable tradeoff in many contexts
Knowably Lying and Enterprise Risk
From an enterprise risk perspective, “knowably lying” is not a psychological category, but a governance failure condition:
Outputting Fabrication or False Attribution when internal uncertainty exceeds an acceptable threshold.
This places it squarely in:
- Risk Class A (Legal)
- Risk Class D (Security)
- Risk Class E (Reputation)
That is why enterprises increasingly focus on:
- Refusal correctness
- Provenance guarantees
- Explicit uncertainty disclosure
Key unresolved gap (explicitly unknown)
Unknown
- There is no reliable, auditable internal signal today that cleanly separates:
- “High confidence because correct”
- “High confidence despite uncertainty”
- “Low confidence masked by fluency”
This is why system-level controls, not prompt engineering, dominate enterprise AI safety strategies.
Web 7.0™ / TDW AgenticOS™ / Hyperonomy™ Digital Identity Lab 