Why Generic AI Fails in Professional Use
ChatGPT, Gemini and Copilot are extraordinary tools for general tasks. But when a property manager uses them to answer questions about property law, when a solicitor queries case law, or when a hotel manager asks about sector regulations, the results can be catastrophic. This article explains precisely why.
The case nobody wants to happen to them
In February 2025, a property manager in Valencia used ChatGPT to answer a query about voting. ChatGPT cited "Art. 19.3 of the LPH" with complete confidence. The problem: that article does not exist. The board acted on false information. Total cost: €4,200 in legal fees + 6 months of community conflict.
~5%
Hallucination rate in generic AI (Stanford CodeX 2025)
0%
Hallucinations in RAG with real documents
€4.2K
Average cost of a hallucination incident
How generative models work — and why that's a problem
Language models like GPT-4 are trained on trillions of words from the internet. They learn statistical patterns, not truths. When they generate a response, they are not searching the LPH or a legal database. They are predicting which words follow the previous ones based on what they saw during training.
How a generative model thinks
After "according to Art." → a number usually follows
After the number → a comma usually follows
After the comma → legal text usually follows
= Generates what is PROBABLE, not what is TRUE
This phenomenon — called hallucination — is especially dangerous in law, medicine and finance. In property management, an invented article can generate challenges to resolutions, civil claims and conflicts lasting years.
Why it fails in block management
Private documents it cannot access. A building's articles of association, internal regulations and minutes are private documents. ChatGPT has never seen them. Any response about their specific content is an invention.
Jurisdiction-specific law. UK leasehold law, Scottish property law and Irish legislation differ significantly. A generic AI trained mostly on US data gives unreliable answers about RTM procedures, Section 20 consultation or service charge disputes.
No verifiable source. A managing agent who relies on ChatGPT has no support if a leaseholder challenges the answer. With IgeraFincas, every response includes the exact article and page number from the original document.
Professional liability. A managing agent acts as agent for the block. Passing on incorrect information that causes loss can constitute professional negligence. "ChatGPT told me" is not a valid defence.
Why it fails in law firms
Case law research requires a specialist database, not a generic language model. ChatGPT generates plausible-sounding case references — for example, "[2024] EWCA Civ 342" — that simply do not exist. Opposing counsel can challenge those citations, causing irreparable damage to a firm's credibility. In the US, several attorneys have already faced sanctions for submitting AI-hallucinated case citations.
IgeraLegal connects directly to verified legal databases and never generates a citation that does not exist in those sources. Every response is court-auditable.
Why it fails in hospitality
Hotels manage complex operational documents: procedure manuals, contracts, local health regulations, seasonal tariffs, HACCP protocols and brand guidelines. Generic AI has no access to any of these. When a receptionist asks ChatGPT about the hotel's cancellation policy, the model invents a plausible answer based on generic industry policies — not the actual contract.
IgeraHospit indexes the establishment's real documents and always responds from the correct source, in the guest's language, 24 hours a day.
| Criterion | Generic AI | RAG (Igera) |
|---|---|---|
| Information source | Training data (may be 2021) | Your current documents |
| Hallucinations possible | Yes (~5%) | No (impossible) |
| Cites exact source | No | Yes (article + page) |
| Accesses private docs | No | Yes |
| Legal liability | Zero (disclaimer) | Traceable and auditable |
The difference that matters in professional use
IgeraFincas, IgeraLegal and IgeraHospit respond from real documents. No hallucinations. Exact citations every time.
See IgeraFincas See IgeraLegal See IgeraHospitQuestion: How many votes are needed to install a lift in a block of 12 flats?
ChatGPT (hallucinated):
"Under Section 20 of the Landlord and Tenant Act 1985, a majority of 3/5 is required." — Wrong legislation, wrong threshold.
IgeraFincas RAG:
"Per Art. 17.2 LPH in your indexed documents, accessibility works are approved by simple majority. Source: LPH Art. 17.2 — page 8 of your indexed document."
Summary
- Generic AI learns patterns, not truths. Hallucinations are structurally inevitable.
- RAG cannot hallucinate: it only responds from real indexed documents.
- In professional use, hallucinations have serious financial consequences.
- IgeraFincas, IgeraLegal and IgeraHospit are the right tool for each sector.
Frequently asked questions
Can hallucinations be reduced with system instructions?
Partially. System prompts can reduce frequency but cannot eliminate hallucinations structurally. A RAG system with real documents is the only solution that guarantees zero hallucinations about the content of those documents.
Can RAG systems also hallucinate?
To a lesser extent. The model can misinterpret a retrieved fragment. That is why IgeraFincas shows the original fragment alongside the response, enabling immediate verification in seconds.
What happens if the indexed documents are out of date?
The system responds according to the documents it has indexed. IgeraFincas includes expiry alerts and a version control dashboard so you always know exactly which version of each document is active.
Is it legal to use AI to answer leaseholder queries?
Yes, provided the responses are verifiable and the managing agent retains final responsibility. With IgeraFincas, every response includes the exact article citation. Using unverified generic AI can constitute professional negligence.
What documents can IgeraFincas index?
Articles of association, internal regulations, minutes, supplier contracts, lease terms, local authority guidance. Core legislation comes pre-indexed as standard.
How much does a generic AI error cost?
The average cost of an incident caused by incorrect legal information exceeds €4,000 in solicitor fees, court costs and management time. The IgeraFincas Starter plan costs €99/month.
Dr. Albert Rovira
Head of AI Research, Igera Solutions
Machine Learning researcher with a PhD in NLP. Has published on hallucinations in LLM models and safe architectures for critical applications.