"Hallucination" is the polite industry word for confabulation — the moment an LLM produces a confident, well-formed, plausible-sounding statement that has no factual basis in its training data and no source in the user's prompt. In wealth, hallucinations are not abstract. They show up as an invented cost basis on a tax-loss harvest, a fabricated account number on a trade authorization, the wrong RMD age inside a Reg BI memo, a made-up FINRA rule citation, plausibly-shaped IRC section numbers with wrong subparagraphs, a Social Security PIA figure that doesn't add up, and trust language the document never contained. The cost of an unrecognized hallucination is the cost of the recommendation built on it. This lesson is the field guide — sixteen specific hallucination patterns advisors actually see in 2026, why each pattern emerges from the mechanics in L1 Ch1.2, and the 30-second visual check for each before it reaches a client or a file.

Why Hallucinations Are Structural, Not Occasional

The temptation, after a few months of using AI without obvious incident, is to conclude that hallucinations are rare events caused by edge-case prompts. This conclusion is wrong, and the wrongness has cost advisors meaningful enforcement and reputational exposure since 2024. Hallucinations are structural to how next-token prediction works. The model samples one token at a time from a learned probability distribution; when the local pattern strongly suggests the shape of an answer ("a section number," "a basis figure," "a rule citation"), the model produces the shape — even when the underlying training data never contained the specific fact at that probability. The output looks confident because the model has no internal signal distinguishing "I learned this from data" from "I am generating this from pattern." The user sees fluency and treats it as accuracy. The accuracy is, in many cases, decoupled from the fluency.

The 2026 wealth-specific consequence is that hallucinations cluster in predictable categories. Citation-shaped answers — IRC sections, FINRA rules, court cases, regulatory citations — are high-hallucination because the local pattern strongly suggests the shape. Numeric answers — cost basis, account numbers, RMD ages, contribution limits, PIA figures — are high-hallucination when the model is interpolating from context rather than reading from a verified source. Quote-shaped answers — "the trust says..." or "the IPS states..." — are high-hallucination when the source document was truncated, malformed, or never actually given to the model. Each cluster has a 30-second visual check; teaching the team the checks is the operational fix. The Cardinal Rule in L1 Ch2.3 is the structured verification protocol that makes the checks routine.

The Sixteen Hallucination Patterns Advisors Actually See

Pattern 1 — The Fabricated IRC Section

Example: "Backdoor Roth pro-rata is governed by IRC §408(d)(6)." Reality: §408(d)(6) covers IRA transfers incident to divorce. The pro-rata aggregation rule lives at IRC §408(d)(2) read with §72(e)(8) and is reported on Form 8606. 30-second check: any IRC section citation in an AI output goes through the IRS-published code text — if the cited subparagraph doesn't match the text or the topic, the citation is hallucinated.

Pattern 2 — The Fabricated FINRA Rule

Example: "FINRA Rule 2210(d)(4)(B) requires testimonial disclosure." Reality: the subsection may not exist, or may exist but not say what the model claims. 30-second check: open the FINRA Rulebook (finra.org/rules-guidance/rulebooks), find the rule, verify the subsection text. If the cited subsection does not match the model's claim, the citation is hallucinated.

Pattern 3 — The Wrong RMD Age

Example: "RMDs begin at age 70½." Reality: SECURE 2.0 set the RMD beginning age at 73 (current) stepping up to 75 in 2033; SECURE 1.0 had used 72. The 70½ figure has not been operative since SECURE 1.0 (2019 enactment, applicable to those reaching 70½ after 2019). 30-second check: RMD age claims default to IRS Publication 590-B / IRC §401(a)(9); any deviation from 73 (current) or 75 (2033+) is hallucinated.

Pattern 4 — The Invented Account Number

Example: "Client's IRA account 8475-021-3398 will receive the conversion." Reality: the model never had access to the actual account number; it generated a plausibly-shaped digit string. 30-second check: any account number in AI output that wasn't in the prompt or the verified source-system feed is hallucinated. Never trust an AI-generated account number on a trade authorization, custodian form, or client document.

Pattern 5 — The Fabricated Cost Basis

Example: "The XYZ Corp lot acquired 2018-03-14 has basis $42,317.81." Reality: the model has no access to the custodian's basis record and generated a plausible figure. 30-second check: all cost basis figures in AI output must originate from the custodian's basis system (Schwab, Fidelity, Pershing, BNY Mellon, TD/Schwab-Advisor), the firm's portfolio accounting system (Orion, Tamarac, Black Diamond, Addepar), or a verified client-supplied document. AI-interpolated basis is dangerous because tax-loss harvest decisions depend on it.

Pattern 6 — The Wrong Social Security PIA

Example: "The client's PIA at FRA is $3,427." Reality: the model has no access to the client's SSA earnings record and either invented the figure or extrapolated from a generic estimate. 30-second check: all PIA figures in AI output must originate from the client's actual SSA statement (downloadable from ssa.gov/myaccount), a verified prior-year benefits letter, or a verified planning-software computation that itself used verified inputs.

Pattern 7 — The Stale Contribution Limit

Example: "The 2026 IRA contribution limit is $7,000." Reality: annual limits change with COLA; the model with a 2024 training cutoff may not have the 2026 figure. 30-second check: all contribution limits in AI output verify against the current-year IRS-published limits (IR-2025-XXX news release, Publication 590-A, the irs.gov/retirement-plans pages). The 2026 limits, the 2026 Social Security maximum taxable wage base, the 2026 IRMAA thresholds — verify before client use.

Pattern 8 — The Hallucinated Trust Language

Example: "Per Article 4 of the trust, the trustee may distribute principal for the beneficiary's HEMS." Reality: the model may be confabulating language consistent with standard trust patterns even if the actual document says something different. 30-second check: any AI quotation of trust, will, POA, or healthcare-directive language goes through the source document. If the cited Article, Section, or paragraph doesn't contain the quoted language, the quote is hallucinated.

Pattern 9 — The Roth Five-Year Clock Confusion

Example: "The client's Roth conversion this year starts a new five-year clock for the contribution." Reality: there are two five-year clocks — the contribution-account five-year rule under IRC §408A(d)(2)(B) (one clock for the oldest Roth IRA), and a separate conversion five-year rule per conversion year for under-59½ withdrawals to avoid the 10% additional tax. The model often blends them. 30-second check: any AI statement on Roth five-year clocks goes through IRS Publication 590-B and the firm's tested decision tree.

Pattern 10 — The SECURE 2.0 Mid-Stream Rule Shift

Example: "Inherited IRA RMDs for non-EDB beneficiaries must be taken each year of the 10-year window." Reality: the IRS guidance evolved post-SECURE 2.0; final regulations clarify the annual RMD requirement during the 10-year period in many cases, but specifics depend on whether the decedent had begun RMDs and the beneficiary's status. The model may state an outdated or incomplete version. 30-second check: any inherited-IRA / 10-year rule statement goes through current IRS guidance (Notice 2024-XX series, final regs) and the firm's tested decision tree.

Pattern 11 — The SECURE 2.0 529-to-Roth Provision Misstated

Example: "The client can roll the entire $50,000 of unused 529 funds to a Roth IRA." Reality: SECURE 2.0 §126 permits limited rollovers — $35,000 lifetime cap, 15-year-account-age requirement, annual Roth contribution limits apply, beneficiary of the 529 must be the Roth owner, and other conditions. The model often understates conditions. 30-second check: any 529-to-Roth statement goes through SECURE 2.0 §126 and the firm's tested decision tree.

Pattern 12 — The IRMAA Bracket Misalignment

Example: "MAGI above $103,000 (single) triggers the first IRMAA tier." Reality: IRMAA thresholds adjust annually and the cited number may reflect a prior year. 30-second check: verify the current-year IRMAA brackets against the SSA / CMS published tables before using in any client recommendation.

Pattern 13 — The AMT Crossover Miscalc on ISOs

Example: "Exercising 5,000 ISO shares with a $20 spread will not trigger AMT." Reality: the AMT bargain element is included in AMTI and depends on the client's full AMT picture. AI estimates of AMT impact without the full AMT calculation are unreliable. 30-second check: all ISO exercise / AMT projections go through the firm's tax-projection software (or the CPA) with verified inputs; AI is for surfacing the question, not answering it.

Pattern 14 — The QSBS Cap Confusion

Example: "QSBS Section 1202 excludes up to $10 million." Reality: post-OBBBA (July 2025), §1202 operates under a dual regime. Stock issued on or before July 4, 2025 retains the legacy $10M / 10x lifetime cap with a five-year hold for 100% exclusion. Stock issued after July 4, 2025 has tiered exclusions (50% at 3-year, 75% at 4-year, 100% at 5-year), a raised $15M / 10x lifetime cap, and a raised $75M gross-asset threshold (from $50M). 30-second check: any §1202 statement explicitly classifies the stock under the applicable regime and the per-tranche issuance date.

Pattern 15 — The Fabricated Court Case or Settlement

Example: "The 2023 Smith v. Custodian case established that..." Reality: the case may not exist or may not say what the model claims. 30-second check: all case citations verify through a primary source (court website, official reporter, Westlaw, Lexis); no case citation goes to a client or a file without verification.

Pattern 16 — Arithmetic Errors in Multi-Step Math

Example: "$96,000 conversion at 24% federal + 5% state + 3.8% NIIT = $31,200 tax." Reality: a chain of multiplications can drift. The model may compound rounding or miscarry a step. 30-second check: any multi-step arithmetic gets verified with a calculator or the firm's projection software. AI is for surfacing the structure of the calculation, not for delivering the final number into the client memo.

Why Each Pattern Emerges (Mechanics Recap)

The mechanics from L1 Ch1.2 explain each pattern:

• Tokenization fragility drives the fabricated IRC section, fabricated FINRA rule, fabricated court case, and invented account number — technical strings reconstruct from local pattern, not from verified retrieval.
• Training cutoff drives stale contribution limits, stale IRMAA thresholds, outdated SECURE 2.0 / OBBBA / final-reg statements — the model didn't see the most recent figures or guidance.
• Stochastic sampling drives run-to-run variability in the multi-step arithmetic and PIA estimation; the same prompt may give a different (still wrong) number tomorrow.
• Context-window truncation drives hallucinated trust language and hallucinated quotes — when the source document was silently truncated, the model may answer about parts it never saw.
• Local-pattern completion drives the fabricated cost basis, invented account number, plausibly-shaped IRC subparagraphs — the model produces the shape the local pattern suggests.

The 30-Second Visual Check System

The operational form of hallucination defense is a habit, not a technology. Every AI-touched artifact gets three checks before it leaves the screen:

1. Citation check (10 seconds). Any rule, statute, case, regulation, or publication citation: scan the AI output for citations; open the primary source for any cite you didn't recognize on sight; verify the subparagraph text matches the claim. If unverified, strike or correct.
2. Numeric check (15 seconds). Any number that came from the AI rather than a verified source: account numbers go through the custodian feed; cost basis goes through the portfolio accounting system; PIA goes through the SSA statement; contribution limits / IRMAA thresholds / RMD ages go through the IRS / SSA current-year publication. If interpolated, strike or correct from source.
3. Quote check (5 seconds). Any quoted language from a client document (trust, will, POA, healthcare directive, IPS): match the quote to the source document; if the source document was uploaded and may have been truncated, verify both the quote text and the section reference.

Thirty seconds per artifact, applied to every artifact, prevents most of the wealth-specific hallucination exposure that has driven 2024-2026 enforcement and reputational matters. The Cardinal Rule in L1 Ch2.3 formalizes this into a three-tier verification protocol (source-system, regulatory, client-fit) at workflow scale.

Reg BI and the Care Obligation — Hallucination Implication

Under §240.15l-1(a)(2)(ii), the registered person must exercise reasonable diligence, care, and skill to understand the recommendation. A recommendation built on a hallucinated cost basis, a hallucinated rule citation, or hallucinated trust language fails the diligence test. The 2025-2026 FINRA AWC pattern on inadequate Reg BI rollover documentation is the warning; the structural cause of similar matters going forward will be hallucinations in AI-drafted memos that registered persons signed without verification. The 30-second visual check system is the operational defense; the Cardinal Rule (L1 Ch2.3) is the formalized protocol; the firm's WSPs (L4 Ch3) are the supervisory architecture.

Key Takeaways

• Hallucinations are structural to next-token prediction, not occasional. The model samples one token at a time and produces plausibly-shaped outputs without an internal "I know this" vs "I'm guessing this" signal.
• Sixteen recurring patterns in wealth: fabricated IRC section, fabricated FINRA rule, wrong RMD age, invented account number, fabricated cost basis, wrong SS PIA, stale contribution limits, hallucinated trust language, Roth five-year clock confusion, SECURE 2.0 mid-stream rule shifts, 529-to-Roth misstatements, IRMAA bracket misalignment, AMT crossover miscalc, QSBS Section 1202 cap confusion (post-OBBBA dual regime), fabricated court cases, multi-step arithmetic drift.
• The IRC §408(d)(6) miscitation is the canonical 2026 example — §408(d)(6) governs IRA transfers incident to divorce; the backdoor-Roth pro-rata rule lives at §408(d)(2) read with §72(e)(8), reported on Form 8606.
• The QSBS Section 1202 dual regime post-OBBBA (July 2025) — legacy $10M / 10x for stock issued on or before July 4, 2025; tiered exclusions (50%/75%/100% at 3/4/5-year hold) with $15M / 10x and $75M gross-asset threshold for stock issued after.
• The 30-second visual check system: citation check (10s), numeric check (15s), quote check (5s) — applied to every AI-touched artifact before it leaves the screen.
• Reg BI Care Obligation under §240.15l-1(a)(2)(ii) is the regulatory anchor — a recommendation built on a hallucination fails diligence; verification is non-negotiable.
• The Cardinal Rule (L1 Ch2.3 next lesson) formalizes the 30-second checks into the practice's standard verification protocol at workflow scale.