AI: Can It Be Trusted? – A CPA, Esq. and Expert Witness View (2025-2026)
Should we trust AI? For accounting, legal, and expert witness work – mostly NO without verification
As a CPA, CVA, CFE, Esq. who has been CFO on 6 public boards and expert witness in 150+ cases, I test AI weekly for forensic accounting. I also teach attorneys who have been sanctioned for using AI to make up fake cases.
Short answer: AI is wrong up to 90% of the time for expert work if you don’t verify. Here is when to trust it and when it will cost you your license.
Pros of Trusting AI – Where it actually helps
1. Efficiency: AI can process vast amounts of data quickly
For my forensic exams with 18,000 bank transactions, AI coding saves 20 hours. It categorizes “AMZN” as Amazon, “SQ” as Square. Then I manually verify 20% sample. Without AI, coding takes 40 hours. With AI + verification, 15 hours.
2. Pattern recognition: AI can identify patterns humans miss
Benford’s Law analysis, anomaly detection – AI finds transactions that don’t fit pattern: e.g., 12 payments to same vendor on Sundays, or round numbers like $5,000.00 that are rare in real business. I use AI to flag, then I investigate.
3. Consistency: Reduces human error on repetitive tasks
Data entry from bank statements to Excel – OCR + AI extracts dates/amounts with 98% accuracy vs 95% manual. But 2% error is still 360 errors in 18,000 transactions – must verify totals tie to bank statement.
Cons of Trusting AI – Why experts get excluded
1. Bias: AI perpetuates biases in training data
If AI trained on Big 4 audit data, it may flag all small business owner personal expenses as fraud, when some are legitimate per operating agreement. AI doesn’t read operating agreement Section 5.2 allowing $10k personal auto.
2. Lack of transparency: Black box problem under FRE 702 / Sargon
Under new FRE 702 (Dec 2023 amendment), proponent must prove by preponderance that testimony is product of reliable principles and methods reliably applied. If you say “AI said damages are $1M” but can’t explain methodology, you are excluded. Judge will ask: “What data did AI rely on? What model? What’s error rate?” If you can’t answer, excluded.
3. Dependence on data quality: Garbage in, garbage out
AI is only as good as data. In my retail store fraud case, seller gave me P&L claiming $300k revenue. If I fed that P&L to AI, AI would calculate valuation at $600k. But bank, POS, and sales tax showed $72k revenue. AI can’t detect fake P&L unless you give it all three sources. I do three-way match: bank vs POS vs tax – AI alone won’t.
4. AI can hallucinate – The biggest risk for attorneys and CPAs
In AI, particularly in large language models, “hallucination” refers to when a model generates information or outputs that aren’t based on any actual input data or facts. This can result in false, nonsensical, or unrelated content.
For instance, if you ask a language model to summarize a news article and it provides information not present in the article, that’s a hallucination.
In legal and accounting, hallucination = fake citations, fake accounting standards.
Other Uses of AI and Chunking Issues
Chunking in AI – Why it destroys forensic analysis
Chunking is a technique used to break down complex information into smaller, more manageable units called “chunks.” These chunks can be phrases, sentences, or even individual words.
Chunking helps AI models in several ways:
- Improved processing efficiency
- Better understanding of dependencies
Chunking can be a problem in forensic accounting:
Loss of context: When chunks are too small, model loses important contextual information. Example: Bank statement line says “Check 1234 – $7,000 – John Doe.” AI chunks into “Check 1234” and “$7,000” and “John Doe” separately, loses that John Doe is owner taking personal distribution, not vendor payment. I saw AI code $84k personal Amex as “office supplies” because it chunked vendor name and missed memo.
If chunks are too large, they might not capture local dependencies – e.g., 12 months bank statements as one chunk, AI misses that $7k Amex payment happens every month on 15th – pattern of fraud.
Overlapping or ambiguous chunks: When chunks overlap or have ambiguous boundaries, it can lead to confusion. Bank description “TRANSFER TO 5678” – is that transfer to account ending 5678 or transfer of $5,678? AI confused.
Chunking bias: If chunking based on flawed assumptions, perpetuates biases.
Personal Scenarios Where I Have Seen Major AI Issues – Real Cases
1. Medical AI program invented body parts – obviously huge problem
I tested medical AI that invented a “third kidney” in radiology report. If CPA AI invents revenue, same problem. I now never let AI generate final numbers – only flag.
2. CPA AI – Where there are similar numbers AI cannot differentiate between 2 different items
Case: Company had two loans: Loan A $100k at 10% to partner, Loan B $100k at 5% to company. AI saw both $100k and merged them, calculated interest wrong, understated partner loan by $5k/year. Over 5 years, $25k error. In litigation, opposing expert caught it, I had to recalculate. Now I require AI to use unique IDs: Loan A vs Loan B, not just amounts.
3. Broker input for expert work – AI botched paystub analysis
I had broker give me paystubs for lost earnings case. AI calculated lost wages using wrong assumptions: used gross pay instead of net pay after taxes (California requires net for PI per CACI), used 40 hours instead of actual 32 hours average, and calculated on wrong numbers (used W-2 Box 1 which includes 401k, not actual wages). AI analysis was off by 40%. I redid manually using timecards and paystubs, got correct lost earnings. AI did not approach problem correctly.
4. Legal – Client warned he would be heavily sanctioned if he used AI again, since it made up fake cases in his filings. Many lawyers have been fined tens of thousands of dollars and risk malpractice and worse. AI is wrong up to 90% of the time
This is epidemic in 2023-2026:
- Mata v. Avianca (SDNY 2023): Attorney used ChatGPT, cited 6 fake cases. Fined $5,000, judge wrote opinion “AI hallucinations are here to stay”
- California attorneys 2024: 3 attorneys sanctioned $10k-$25k for fake AI citations in family law and PI
- My client: Was warned by judge he would be heavily sanctioned if used AI again, since it made up fake cases in his filing. He used AI to draft motion, AI invented CCP section that doesn’t exist. Opposing counsel caught it. Judge threatened $10k sanctions and State Bar referral.
Why AI invents cases: AI predicts next word, not retrieves law. If you ask “cases where CPA breached fiduciary duty,” AI may invent “Smith v. Jones, 2023 WL 123456” that sounds plausible but doesn’t exist. Westlaw/Lexis search shows no such case.
Rule for attorneys: Never cite case without pulling it on Westlaw and reading it. I tell hiring attorneys: If you used AI to draft complaint, send me complaint and I will Shepardize all cases before we file expert report citing them.
When to Trust AI – My CPA/Esq. Rules for Expert Witness Work
1. Well-defined tasks with clear objectives and high-quality data
- OK: Use AI to OCR bank statements, extract dates/amounts, code transactions by vendor name – THEN manually verify totals and sample 20%
- NOT OK: Use AI to calculate damages, lost profits, or business valuation alone
2. Human oversight: AI systems should be designed with human oversight and review processes and outputs
My workflow:
- AI does first pass – coding, flagging anomalies
- I review 100% of flagged transactions and 20% random sample
- I recalculate damages manually in Excel – AI does NOT calculate final number
- I cite actual documents (Bates stamped) not AI summary
3. Never let AI write expert report or declaration
Expert report must be your opinion under oath (CCP 2034.260). If AI wrote it, you can’t testify you hold that opinion. Sargon requires you personally applied methodology.
4. Disclose AI use if court requires
Some judges now require disclosure: “Was AI used to draft this?” Central District of California requires disclosure. I disclose: “AI used for data coding, all damages calculated manually and verified.”
AI Tools I Actually Use and Trust (Partial List)
- OCR: Adobe Acrobat Pro OCR for bank statements – 98% accurate, I verify totals
- Transaction coding: Custom GPT trained on my coding rules – 85% accurate, I correct 15%
- Benford’s Law: ACL, IDEA software – reliable, mathematically sound
- Document review: Everlaw for large doc review – but I still read key docs
Tools I don’t trust for expert work:
- ChatGPT, Claude, Gemini for calculating damages or writing report – hallucinates
- Any AI that says it can do “business valuation in 5 minutes” – violates NACVA standards
- AI that claims “IRS approved” 409A – no such approval, must be qualified appraiser
Bottom line: AI is like an intern, not an expert. Interns can code data, flag issues, but you – CPA, CVA, Esq. – must verify, calculate, and opine. If you rely on AI for final opinion, you risk exclusion under FRE 702, Sargon, sanctions, and malpractice.
If you are attorney needing forensic accounting that survives Sargon and new FRE 702, or you need to challenge opposing expert who used AI, contact hpaccounting.com/contact. I test AI so you don’t get sanctioned.
Sources:
- New Scientist: AI hallucinations getting worse and here to stay
- Medium: 500 billion hallucinations how LLMs failing in production