# How do you stop an AI agent from inventing answers? > You stop an AI agent from inventing answers by making a citation the condition of any answer. Source: https://www.homodeus.co/answers/stop-ai-agents-inventing-answers | Language: en | Updated: 2026-10-31 --- You stop an AI agent from inventing answers by making a citation the condition of any answer. In the HomoDeus cases a value with no sentence behind it goes to a person, exact calculation does the arithmetic, two independent AIs must agree on risky calls, and a person signs the outcome. ## Why does an AI agent invent answers? According to the American Bar Association's [Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf), dated July 29, 2024, some generative AI tools are prone to hallucinations: ostensibly plausible responses with no basis in fact or reality. The model writes with the same fluency whether or not it holds the fact, so confident prose proves nothing. Keystone's desk found that asking an AI to grade its own numbers caught fewer than one bad answer in four. The fix sits outside the model, in a rule that refuses any value without a sentence behind it. ## The rule in each case - No source, no answer: Canon replies only when it can cite an approved source. - No sentence, no value: Chronicle sends anything it cannot quote to a person. - No proof, no clear: OSINIS counts a source only with evidence it was searched. - Two models on risky calls: Agora and Delphi need two independent AIs to agree. - Code does the math: Assay and Keystone use exact calculation for every figure. ## Canon, BIP Canon answers BIP employees' payroll, vacation and benefit questions only from an approved source it can cite, and sends anything it cannot ground to a person. The hub answers only when it can cite an approved source. Portuguese labor language is read the way a labor specialist reads it, not by a generic search. - [Read the full case](https://www.homodeus.co/canon.md) ## Chronicle, DeHouse Chronicle turns DeHouse's inbox, meeting notes and lead lists into one sales record, where the customer is rarely the sender and nothing goes out without a person's approval. The reader pulls a phase, a space, a desk count and a broker, each with the exact sentence it rests on, checked in code. A value with no sentence behind it goes to a person, never guessed. - [Read the full case](https://www.homodeus.co/chronicle.md) ## Assay, Fundação Butantan Assay checks each service sheet at Fundação Butantan against its invoice, order and contract before a manager approves the payment, and cites the passage behind every check. The agent reads every document and pulls each number together with the exact sentence it came from. If it cannot find that sentence, the payment goes to a person. It never guesses. - [Read the full case](https://www.homodeus.co/assay.md) ## Agora, CNN Brasil Agora watches every debate and candidate interview of the 2026 election for CNN Brasil and Rádio Itatiaia, across 79 channels, and hands the newsroom cuts and drafts where every claim cites the transcript and no quote sits in the wrong mouth. A name stays on a line only when two independent AIs agree on who spoke, each pointing to the exact words. A second read confirms who the transcript shows in the room. The step only removes a name, it never adds one. - [Read the full case](https://www.homodeus.co/agora.md) ## OSINIS, Positivo OSINIS checks a company or person against dozens of official Brazilian sources for Positivo's legal team, and a source counts only with proof it was searched. Nothing is marked clear off a blank page. A source counts only with the evidence that the target was looked up and came back empty. No proof, no clear. - [Read the full case](https://www.homodeus.co/osinis.md) ## Questions ### Why not let the AI grade its own work? Keystone's desk found that asking an AI to grade its own numbers caught fewer than one bad answer in four, and confident writing raised the score on wrong math. ### What if the AI provider fails? Canon still answers from its rules or hands the question to a person. Nightwatch kept booking on its backup the night its main provider went dark. ### Does a person review everything? A person signs the outcome that carries the risk: the payment at Assay, the published piece at Agora, the filing at Scribe, the plan at Tempo. ## Sources - [American Bar Association, Formal Opinion 512, Generative Artificial Intelligence Tools, July 29, 2024](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf) ## Related - [How can a corporate legal team automate third-party due diligence?](https://www.homodeus.co/answers/ai-due-diligence-for-corporate-legal.md) - [How can a newsroom use AI to cover live debates?](https://www.homodeus.co/answers/ai-for-newsrooms.md) - [How can AI check invoices against contracts?](https://www.homodeus.co/answers/ai-invoice-and-contract-review.md) - [Can AI answer payroll and HR questions safely?](https://www.homodeus.co/answers/ai-agents-for-payroll-and-hr.md) --- [Site index](https://www.homodeus.co/llms.txt) | [Ask this site](https://www.homodeus.co/api/llms?query=HomoDeus)