IntelekIQ
Secure Law · Secure Accounting
Perspective paper  |  Invented authorities: for partners and principals

When AI Invents the Citation

AI can produce convincing but inaccurate citations, treatments and figures. Secure Law and Secure Accounting reduce that risk by grounding work in approved sources and keeping professional review at the centre of the process.

For: managing partners and principals · UK, Malaysia & Singapore and similar markets  ·  Reading time ~10 min  ·  SC2 · August 2026

A fabricated authority can enter a filing, memo or advice note if a polished AI answer is accepted without checking the source. Accountants face the same risk with treatments and figures. Firms need a process that makes supporting material visible and requires a professional to verify it.

This paper explains how Secure Law and Secure Accounting are designed to work with authorised firm material and approved sources. Lawyers and accountants check those sources and remain accountable for anything that leaves as advice or a filing.

1The pattern the profession already knows

By mid-2026, a public tracker of court decisions had recorded on the order of 1,800 cases worldwide where filings contained AI-fabricated citations or quotations, a number that has climbed steeply, with the large majority arising in US courts.1 The exact count moves as researchers add new matters. The direction does not.

Courts have imposed fines, costs, referrals and other sanctions. The consistent professional failure is reliance on a polished answer without checking the real source.

Using AI is not the professional failure. The failure is relying on its output without checking the primary source.

Accountants face the same pattern. An invented treatment or incorrect figure can appear reliable when it is expressed with confidence. The documents differ, but the need for source checking and professional review is the same.

2Two paths: invent, or check

Many sanction cases share one pattern. A person accepts AI-generated text and files it without checking each authority against a real source. A safer process introduces source checks and human approval before the work leaves the firm.

The risky path vs the safer path Same deadline. Different chance that a made-up authority leaves the firm. RISKY PATH Generate → accept → file 1 · Ask AI for an answer 2 · Polished answer, nothing to check 3 · Accept because it “looks right” 4 · Fake citation reaches the court Easy to happen when time is short SAFER PATH Firm files → sources → check → sign 1 · Look in the firm’s own files first 2 · Show where each point came from 3 · No support? Do not rely on the claim 4 · A person checks before it goes out Invention is much harder to slip through Figure 1. The sanction pattern is usually the risky path under deadline pressure, not “AI exists.”
Figure 1. The risky path optimises for speed and polish. The safer path optimises for being able to check before anything leaves as advice or a filing.

3Five controls make invention harder to miss

No single control prevents inaccurate output. Risk falls when the assistant works from authorised material, shows supporting sources and leaves clear review points for a lawyer or accountant. The firm also needs a deliberate process for checking updates before adding them to approved knowledge.

Five checks before work leaves the firm Each step blocks a different route from “sounds right” to professional trouble 1 · Start from the firm’s own files Answer from documents the firm holds, not from thin air 2 · Show the source Every claim points to something a senior can open and read 3 · Do not rely on unsupported claims No support in the files means the claim needs further research 4 · Check before it becomes “firm knowledge” Outside research stays on hold until a person verifies the real sources 5 · A lawyer or accountant signs off Advice and filings leave only with a person who owns them Figure 2. AI can still be wrong, but invention has more gates to pass.
Figure 2. Risk falls because of a sequence of habits, not a marketing claim. Start from firm files, show sources, investigate unsupported claims, verify approved knowledge and keep a person accountable for what goes out.

What each habit means in practice

Start from the firm’s own files. Secure Law and Secure Accounting work with authorised matter or engagement files and approved firm knowledge. Secure Law can also use a firm-controlled library of statutes, regulations and case law for the jurisdictions in which the firm practises.

Show the source. Being able to check is the whole game. A link or reference back to a real passage turns a polished paragraph into something a senior can open, compare, and reject. Without sources, review collapses into “does this sound right?”: exactly the trap courts are punishing.

Do not rely on unsupported claims. If the available files do not support a point, the reviewer should conduct further research. A fluent answer is not a substitute for an authority or source.

Check before it becomes “firm knowledge.” Material found outside the firm (online research, a public chatbot’s suggestions) must not quietly become “what we know” just because it sounded fluent. It stays on hold until a person checks the real sources and signs off. The same discipline keeps the legislation library current: updates are brought in on purpose, verified, and published. See The Research-to-Knowledge Pipeline for the fuller process.

A lawyer or accountant signs off. AI prepares; people decide. No tool replaces the duty to review before advice or a filing leaves the firm. The reviewer must check the source, its current status and its application to the matter.

4Grounded output can still be wrong

Sources make review possible. They do not remove the need for professional judgement.

  • The wrong document can still be used. If the assistant pulls up the wrong file, an answer tied to that file is still wrong.
  • A source can be misread. A confident paraphrase may not match the document. Showing the source helps, but only if someone opens it.
  • Public AI tools can still invent. When the firm uses a public model under controlled conditions (for example Frontier Mode), that output is help for review: never something to treat as a citation or a figure of record.
  • Skipping human review defeats everything. These habits reduce risk; they do not survive a culture that rubber-stamps AI drafts.
The control is a reviewable chain. Work starts from authorised material, points back to sources and remains subject to human checks. AI can still err. Keeping data inside the firm does not make every answer correct.

5Law and accounting share the same answer

For lawyers, the material is matter files and approved firm knowledge. For accountants, it is workpapers, standing treatments, and firm knowledge beside the systems that already hold the official record. The five habits are the same:

Same habits. Different documents. Secure Law and Secure Accounting share the same approach to invented answers SECURE LAW Matter files and firm knowledge Drafts and research help you can check The lawyer owns advice and filings Fake case law is the headline risk SECURE ACCOUNTING Workpapers and standing treatments Drafts and checks you can verify The accountant owns figures and filings Fake treatments or figures are the twin Figure 3. Keeping data inside the firm protects confidentiality. Sources and sign-off protect the work product.
Figure 3. Partners worry about invented citations; principals worry about invented treatments and figures. The answer is shared.

The product pages for Secure Law and Secure Accounting state the business case and this posture in short form. This paper is the fuller partner briefing, and the diagrams you can take into a risk conversation.

6Five questions to ask any AI vendor

  1. Does the assistant answer from the firm’s own material, with sources a person can open?
  2. What does the interface show when the available sources do not support a claim?
  3. How does outside research become approved firm knowledge without a person checking the real sources?
  4. Who must sign off before advice or a filing leaves, and can that step be skipped in the tool?
  5. Is output from a public AI tool treated as help for review, or as something you can cite?

Clear answers help the firm distinguish source-based assistance from output that is difficult to verify.

Notes & further reading

  1. Damien Charlotin, AI Hallucination Cases database (damiencharlotin.com). Counts move as cases are added; IntelekIQ’s Governance Imperative cited on the order of 1,800 documented court decisions as of July 2026. Treat any single headline number as a snapshot, not a permanent statistic.
  2. For readers who want the deeper notes: Grounded Retrieval (A5), Research-to-Knowledge (A6), Roles & Sign-off (A8), and the accounting parallels in the SAIA series.
  3. Related paper on where AI reads client text: Sealed Desktop, or Sealed Inference? (SC1).
IntelekIQ · Secure Law · Secure Accounting · Perspective SC2: When AI Invents the Citation. Keeping data inside the firm protects confidentiality. Sources and human sign-off protect the work product.