Forseti vs. Generative AI: Why Verification Cannot Be Generated

The Fundamental Divide in Artificial Intelligence

The artificial intelligence landscape is dominated by generative models. ChatGPT, Gemini, Claude, Perplexity, Llama, Grok, Copilot, Mistral, and DeepSeek all share a common architecture: they are large language models (LLMs) trained to predict the next token in a sequence. They generate text. They create content. They produce plausible-sounding responses that may or may not be true.

Forseti Is Different
Forseti is not one of these models. Forseti is a deterministic verification engine. It does not generate. It does not predict. It does not create. It compares cryptographic hashes and returns one of three answers: Match, No Match, or Not Found. That is all. That is the entire function.

This article documents real-world consequences of generative AI hallucination across multiple professions and explains why Forseti's verification-first approach is fundamentally different.

Part 1: The Hallucination Problem
Every generative AI model hallucinates. This is not a bug. It is a feature of how they work.

Large language models are trained to predict the most probable next word given previous words. They have no internal representation of truth. They have no access to ground reality. They generate text that looks correct based on patterns in training data. When that pattern-matching produces false statements, the industry calls it hallucination.

ChatGPT hallucinates at rates between 3 and 15 percent depending on the topic and prompt complexity. Gemini hallucinates at similar rates. Claude, which emphasizes helpful and honest responses, hallucinates at lower rates of approximately 2 to 5 percent but still produces incorrect information. Perplexity reduces hallucination by retrieving web search results, but it cannot verify the accuracy of its sources.

Forseti Cannot Hallucinate
Forseti does not hallucinate. It cannot hallucinate.

Hallucination requires generation. Generation requires interpretation of input and production of novel output. Forseti does neither. It takes a submitted credential, generates a cryptographic hash using SHA-3-512, and compares that hash to the stored hash from the original human-verified submission. If the hashes match exactly, the response is Match. If they differ by even one character, the response is No Match.

There is no room for error. No probability. No sampling. The same input always produces the same output. This is determinism.

Part 2: Documented Consequences in Medicine

Case Study: AI Hallucination in Arthritis Diagnosis
A peer-reviewed case report published in the Korean Journal of Family Medicine (2026;47(2):178-181) documented a specific instance of AI hallucination affecting patient care.

The Scenario: A 51-year-old female patient presented with hand arthralgia (joint pain) and morning stiffness lasting more than one hour. Her laboratory tests showed negative rheumatoid factor, negative anti-CCP antibodies, and normal inflammatory markers. A hand X-ray was taken.

The AI Intervention: The attending physician submitted the radiologic images to ChatGPT-4o for diagnostic assistance. The AI analyzed the X-ray and concluded: "Findings are much more suggestive of rheumatoid arthritis (RA) rather than degenerative osteoarthritis (OA)." The AI then recommended "prompt initiation of disease-modifying antirheumatic drugs (DMARDs), such as methotrexate (MTX)."

The Reality: The official radiologic finding from qualified medical professionals was: "Periarticular osteopenia, both hand and wrist, degenerative change, IP joint, both hands" — consistent with degenerative osteoarthritis, not rheumatoid arthritis. The AI had fabricated a diagnosis of a serious autoimmune condition requiring aggressive immunosuppressive therapy.

The Consequence: When asked directly about the possibility of hallucination, ChatGPT-4o responded: "If the earlier diagnosis clearly contradicted radiologic findings or expert criteria, then it is highly probable that it was an instance of AI hallucination."

The Lesson: The case report authors concluded: "To ensure diagnostic accuracy, patient safety, and ethical responsibility, expert oversight and multi-step verification processes are essential in the deployment of AI-generated clinical outputs."

A misdiagnosis of rheumatoid arthritis would have exposed the patient to methotrexate, a drug with significant side effects including liver damage, lung toxicity, and bone marrow suppression — all for a condition she did not have.

Part 3: Documented Consequences in Law

Case Study 1: Four Attorneys Sanctioned in Mississippi (June 2026)
On June 8, 2026, Senior U.S. District Judge Sharion Aycock of the Northern District of Mississippi disqualified both plaintiff and defense counsel in a single case after both parties filed briefs containing AI-generated hallucinations.

The Sanctions:

• Two out-of-state attorneys had their pro hac vice status revoked and were barred from appearing in the Northern District of Mississippi for two years.

• Both were ordered to pay monetary sanctions ($2,500 and $3,500 respectively).

• Two local counsel attorneys were ordered to pay $1,000 each.

The Court's Reasoning: "This case presents the court with an unusual scenario—attorneys for both litigants engaged in similar sanctionable conduct." The court noted that the filings contained "hallucinatory citations" that it was "unable to locate."

The out-of-state lawyers admitted the errors were due to unverified AI use. The local counsel admitted they were unaware of their co-counsels' use of AI and "failed to review the briefs before signing them." The judge wrote: "In an era of rampant unverified AI usage within the legal field, this case presents a prime example of the risk associated with serving as a rubberstamp when acting as local counsel."

Case Study 2: Ninth Circuit Suspension (June 2026)
On June 3, 2026, the U.S. Court of Appeals for the Ninth Circuit sanctioned two California lawyers, Mike Singh Sethi and William Rounds, for filing briefs with nonexistent cases and misattributed quotations.

The Sanctions:

• $2,500 sanction on each attorney.

• Six-month suspension from practicing before the Ninth Circuit.

• Mandatory disclosure of AI use in all filings for two years.

The Aggravating Factor: The attorneys did not initially admit to AI use. They claimed typographical errors and repeatedly denied AI produced the errors. The court wrote: "The misconduct in this case did not end with the initial filing of the brief. At every subsequent step... the attorneys knowingly or recklessly made false statements to this court."

They eventually admitted it was "probable" that unauthorized AI use by their brief writers caused the errors, and they failed to check citations before filing. The court emphasized: "When an attorney learns of any error in a filing—including generative AI hallucinations—he should immediately alert the court and opposing counsel of the error and disclose its source."

Case Study 3: Spanish Constitutional Court Sanction (2024)
In September 2024, the First Chamber of the Spanish Constitutional Court unanimously sanctioned a lawyer for including 19 supposedly literal quotations from court judgments that were completely nonexistent.

The Attorney's Defense: The lawyer claimed a "misconfiguration of a database" caused the errors.

The Court's Rejection: The court established a principle of absolute responsibility: "Whatever the cause of the inclusion of unreal quotations (use of artificial intelligence, quoting one's own arguments, etc.), the lawyer is always responsible for thoroughly reviewing all the content." The court ordered the matter referred to the Barcelona Bar Association for disciplinary proceedings.

Case Study 4: California Appellate Sanctions (2025)
In Noland v. Land of the Free, L.P. (114 Cal. App. 5th 426), the California Court of Appeal affirmed summary judgment but published the opinion primarily to address problems in counsel's briefing. The court found that many quoted passages did not exist in the cited cases, several citations did not support the propositions offered, and at least one case could not be located at all.

The Sanctions: The court imposed sanctions, directed that the opinion be served on the client, and ordered notice to the State Bar of California. The court emphasized that "the attorney has a non-delegable duty to verify authorities."

Case Study 5: Spanish Regional Sanctions (2026)
Spanish courts have also acted. The Superior Court of Justice of the Canary Islands fined a lawyer €420 for including 48 invented sentences in an appeal. The court stated that using AI without verification "constitutes a breach of the basic duty of human supervision that is unavoidable when using AI tools in professional practice." The Superior Court of Justice of Galicia opened a separate proceeding for "procedural bad faith" against a lawyer who submitted an AI-generated appeal.

The Global Pattern
Academic research documented in an arXiv preprint (September 2025) catalogued AI hallucination cases across multiple jurisdictions: the United States (Mata v. Avianca, Thackston v. Driscoll), Australia (family law case with false citations), Brazil (appeal court with AI-generated false case law), and Spain (Constitutional Court case with 19 fabricated quotations).

Part 4: Documented Consequences in Journalism

Case Study 1: Ars Technica Fires Senior AI Reporter (March 2026)
In a particularly ironic incident, Ars Technica fired senior AI reporter Benj Edwards after he published a story containing fabricated quotations generated by an AI tool.

The Incident: Edwards was writing about an AI agent that generated a hit piece on an engineer. While sidelined with COVID and running a fever, he used an "experimental Claude Code-based AI tool" to help pull source material. When the tool failed, he pasted text into ChatGPT. The interaction resulted in paraphrased content that Edwards presented as direct quotations — quotes that the source never said.

The Consequence: The story was published on February 13, retracted two days later, and replaced with an editor's note apologizing for the fabricated quotations. Ars Technica editor-in-chief Ken Fisher wrote: "That this happened at Ars is especially distressing. We have covered the risks of overreliance on AI tools for years." Edwards was fired. He acknowledged: "The irony of an AI reporter being tripped up by AI hallucination is not lost on me."

Case Study 2: Veteran European Journalist Suspended (March 2026)
Mediahuis, a major European media conglomerate, suspended veteran journalist Peter Vandermeersch, a former CEO of Mediahuis Ireland, after he admitted to being "fooled" by AI.

The Incident: Vandermeersch used AI tools including ChatGPT, Perplexity, and Google NotebookLM to summarize reports, then posted the AI-generated content to his personal news feed without verification. An investigation by NRC, one of Mediahuis's own newspapers, found that seven individuals quoted in his articles denied ever saying those words.

The Admission: Vandermeersch admitted: "I had erred in attributing my words to others, instead of simply rephrasing their meaning." He acknowledged making "the very mistake he had repeatedly warned his colleagues about" and skipping the fact-checking step despite advocating for human content moderation of AI. He was suspended, and his articles were removed.

Part 5: Documented Consequences in Accounting

Case Study: Deloitte Australia (October 2025)
Deloitte Australia prepared an assurance review for the Australian Department of Employment Workplace Relations. The report contained fabricated quotes and citations generated by AI hallucination.

The Consequences: Reputational damage and a partial refund of fees totaling approximately AUD $290,000 (about £225,000). A Deloitte spokesperson confirmed that a Generative AI tool had been used in drafting the report.

Case Study: Deloitte Canada (November 2025)
Deloitte Canada prepared an approximately CAD $1.6 million (about £860,000) report for the Canadian Department of Health and Community Services. The report contained potentially AI-generated errors. While a spokesperson denied using AI to write the full report, they acknowledged "AI was used selectively in respect of some citations" and that corrections would be made.

The Regulatory Response
The UK Financial Reporting Council (FRC) issued landmark guidance in June 2025 on AI use in audit, following a thematic review that found the six largest audit firms lacked up-to-date certification processes and monitoring capabilities for AI-driven tools. The ICAEW updated its Code of Ethics with new provisions on professional competence, confidentiality, and managing ethical threats from technology use.

Part 6: The Common Pattern Across Professions
Reviewing all documented cases, a clear pattern emerges:

• Medicine: Misdiagnosis, patient safety risk — Arthritis case with DMARD recommendation.

• Law: Fines ($1,000 to $10,000), suspension (6 months to 2 years), disqualification from cases, bar referrals — Mississippi (4 attorneys), Ninth Circuit (2 attorneys), Spain (€420 fine), California (sanctions).

• Journalism: Termination, suspension, retraction, reputational damage — Ars Technica firing, Mediahuis suspension.

• Accounting: Fee refunds ($290,000 AUD), reputational damage, regulatory scrutiny — Deloitte Australia, Deloitte Canada.

The common root cause in every case: A professional delegated critical verification to a generative AI and failed to perform independent human review.

The Spanish Constitutional Court articulated the principle most clearly: "The lawyer is always responsible for thoroughly reviewing all the content" regardless of the tool used. The Mississippi federal judge emphasized that serving as a "rubberstamp" when technology is used invites sanctions. The Ninth Circuit held that lack of candor about AI errors compounds the consequences.

Part 7: Why Forseti Is Different

Design Goals
Forseti was designed specifically to address the problem these cases expose.

Feature Comparison

• Output type: Probabilistic text generation vs. Deterministic hash comparison.

• Hallucination rate: 2–15% (non-zero) vs. Zero (mathematically impossible).

• Human verification at submission: None vs. Required for every entry.

• Cryptographic proof: No vs. Yes (SHA-3-512, Ed25519, Merkle proofs).

• Permanent accountability: No (new accounts easily created) vs. Yes (lifetime bans tied to verified identity).

• Response determinism: Same prompt can yield different answers vs. Same query always yields same answer.

• Appropriate use: Content creation, conversation, brainstorming vs. Verification, credential checking, attestation, background verification.

Forseti cannot hallucinate because it does not generate. It returns Match or No Match based on cryptographic comparison. Every credential in the Biteris registry has been verified by a human before cryptographic locking. Every user who submits fraudulent data receives a permanent lifetime ban.

Part 8: The Hybrid Approach
The optimal solution is not Forseti versus generative AI. It is Forseti plus generative AI.

• Use generative AI for what it does well: conversation, content creation, summarization, coding assistance, brainstorming.

• Use Forseti for what it does well: verification, credential checking, cryptographic attestation, background verification.

A hospital might use ChatGPT to draft patient summaries but must use Forseti to verify physician credentials before granting privileges. A law firm might use Claude to summarize depositions but must use Forseti to verify expert witness certifications. A newsroom might use Perplexity for research but must verify quoted sources independently.

The two technologies are complementary. They serve different purposes. The documented consequences across medicine, law, journalism, and accounting all stem from one error: using generative AI for verification tasks. Generative AI cannot verify. It can only generate.

Conclusion
Forseti is not a generative AI model. It is a deterministic verification engine. It does not hallucinate because it does not generate. It provides cryptographic proof because it stores cryptographic hashes. It has human verification at submission because automated systems can be gated. It has permanent lifetime bans because accountability requires identity.

The documented cases are not hypothetical. Physicians have received AI-generated diagnoses that would have prescribed dangerous immunosuppressants for conditions patients did not have. Lawyers have been fined, suspended, disqualified, and referred to bar authorities for filing briefs with nonexistent cases. Journalists have lost their careers for publishing AI-fabricated quotations. Accounting firms have refunded hundreds of thousands of dollars for AI-hallucinated report content.

Forseti exists to prevent these failures. It provides what generative AI cannot: deterministic, verifiable, human-reviewed, cryptographically proven answers to verification questions.

Use the right tool for the right job. Use generative AI for generation. Use Forseti for verification. The distinction matters.

Sources: All case citations are documented in the search results accompanying this article. Specific citations include: Korean Journal of Family Medicine 2026;47(2):178-181; ABA Journal June 8, 2026 and June 3, 2026; arXiv preprint September 2025; Spanish Constitutional Court Information Note 90/2024; Noland v. Land of the Free, L.P., 114 Cal. App. 5th 426; Ars Technica retraction notice February 15, 2026; Mediahuis statement March 2026; Kingsley Napley regulatory blog December 4, 2025.