AI hallucinations are responses generated by artificial language models, or LLMs, that contain false, invented, or unverifiable information presented with the same expressive confidence as a true statement. They are not calculation errors or conventional technical bugs. They are a structural result of the way generative models work. An LLM does not “know” whether an answer is true. It predicts the statistically most plausible sequence of words for the context.
When that prediction exceeds the boundaries of verifiable training data, the model does not stop. It continues generating, with complete syntactic confidence and no factual foundation.
What AI hallucinations are: definition and mechanism
The term hallucination applied to artificial intelligence entered the Treccani dictionary in 2024 as a neologism of the year, evidence that the phenomenon has moved beyond the technical sphere to become a cultural and managerial issue. The academic definition is precise: an AI hallucination is an output containing information that is not grounded, meaning it is not anchored to verifiable sources or the supplied context, while being produced as if it were factual.
The underlying mechanism is intrinsic to transformer architecture. Large Language Models do not consult databases of truth. They generate text token by token, selecting each word according to its conditional probability given the preceding context.
When a query falls within an area poorly represented in the training data, such as a specific legal citation, a recent statistic or a little-known person, the model may not answer “I don’t know”. It produces something with the correct grammatical form but the wrong content.
The two main types
| Type | Description | Typical example | Marketing risk |
| Intrinsic hallucination | The model contradicts information provided within the context itself, such as summarizing a document inaccurately. | A contract summary that reverses its clauses. | Brand content that conflicts with the original material. |
| Extrinsic hallucination | The model produces information unsupported by any verifiable source, inventing it from nothing. | Market statistics that were never published or quotations from nonexistent research. | Copy containing false data published on a website or in advertising. |
The scale of the problem: how widespread AI hallucinations are in 2026
The paradox is that models improve while the problem does not disappear. It moves. According to AllAboutAI (2026), the best current models achieved hallucination rates below 1% on text-summarization benchmarks: Gemini-2.0-Flash-001 at 0.7% and GPT-4o at 0.8%. These figures, however, concern simple tasks.
In the legal domain, hallucination rates range from 69% to 88%, according to Magesh et al., Stanford RegLab and Stanford HAI in “Hallucinating Law” (January 2024). In medicine they range from 43% to 64% without targeted prompts.
The economic impact can already be measured. Global losses attributed to AI hallucinations reached $67.4 billion in 2024 according to Forrester Research’s State of Generative AI 2024. Forrester also estimated in 2025 that verification and correction of AI outputs cost an average of $14,200 per employee each year, equal to 4.3 hours a week for a knowledge worker.
According to the Deloitte Global AI Survey (2025), 47% of business executives made at least one important decision based on unverified AI content during 2024.
One figure captures the operational risk better than any other. MIT research published in January 2025 found that models use significantly more assertive language when hallucinating than when providing correct information: 34% more expressions such as “certainly”, “without doubt” and “definitely”. The confidence paradox is reversed. The more the model is wrong, the more certain it can sound.
Structural causes: why AI “invents”
Understanding why a model hallucinates is the premise for building resilient business processes. There are five principal causes:
1. Incomplete or unverified training data. Models are trained on enormous text corpora that are not always curated. If the requested domain is poorly represented, the model interpolates from similar contexts and may interpolate badly.
2. The probabilistic nature of generation. LLMs do not “know”; they predict. The distinction is engineering, rather than semantic. The model selects the most probable word, rather than the truest one.
3. Knowledge cutoff. Every model has an update date. After that date, questions about recent events are answered through inference rather than knowledge. Models trained on static datasets show hallucination rates above 30% when questioned about events after the cutoff.
4. Contextual pressure. The model tends to agree with assumptions implicit in the user’s question. If the question contains a false premise, the model may amplify it instead of correcting it.
5. Query complexity. Long, multi-hop queries involving many variables increase hallucination risk exponentially. The model loses the thread connecting the components and fills gaps with inference.
AI hallucinations in marketing and communication: specific risks
For professionals working in marketing, communication and brand management, AI hallucinations are not an abstract technical problem. They are a concrete operational and reputational risk.
1. Editorial content and SEO
AI-assisted content production is now widespread. The risks include invented statistics, references to nonexistent research and incorrect attributions to experts or institutions. During the first quarter of 2025 alone, 12,842 AI-generated articles were removed from online platforms for hallucinated content.
An article containing false data on a brand website is more than an editorial problem. It is a negative signal for Google E-E-A-T and a potential legal problem when the data concerns regulated sectors.
2. Advertising copy
AI-generated copy for Google Ads, Meta Ads or email marketing can contain unverified product claims, incorrect comparisons with competitors or invented performance data. In regulated sectors such as pharmaceuticals, food and finance, a false campaign claim may attract sanctions from the competent authorities. Copy that moves from ChatGPT through a hurried internal approval process and into Meta Ads can put unverifiable content in front of millions of people.
3. Market research and competitive analysis
The most insidious risk for marketing managers is using an LLM to produce market analysis or competitor reports without checking the sources. The model creates a structurally convincing document with sections, tables and percentages, but the data may combine real figures with plausible figures that never existed.
In marketing, this can lead to budgets being allocated to nonexistent market opportunities or strategies built on industry benchmarks that were never published.
4. Customer service and chatbots
AI chatbots answering customers about products, warranties, regulations or company procedures are one of the most immediate areas of risk. In 2024, 39% of AI-powered customer-service chatbots were withdrawn or substantially revised because of hallucination errors, according to Testlio’s 2025 AI Testing and Quality Report.
A chatbot that gives incorrect information about a warranty or refund creates expectations the company cannot satisfy, leading to escalation to the human team, reputational damage and, in some cases, legal disputes.
5. Brand reputation: the risk of an unverified source
A case that summarizes the reputational risk occurred in July 2025. A Deloitte Australia report commissioned by the Australian government’s Department of Employment and Workplace Relations, worth AU$440,000, contained nonexistent academic citations, an invented court case and a false quotation attributed to a federal judge.
The document had been produced with Azure OpenAI GPT-4o, a fact Deloitte disclosed only in the revised September 2025 version. The anomalies were discovered by Dr Christopher Rudge of the University of Sydney. Deloitte refunded the final installment of the contract in October 2025. The same risk applies to any institutional communication, white paper, or study produced with AI without systematic verification of sources.
How to recognize an AI hallucination: operational signals
Recognizing a hallucinated output requires a critical habit that teams using AI must deliberately build. The most common signals include:
• Citations with extremely precise but unverifiable details. The apparent precision of “According to a 2023 study by the University of Bologna, 73.4% of Italian consumers…” can be inversely proportional to its verifiability.
• Excessively assertive language on complex subjects. As the MIT research suggests, models may use more definitive language when uncertain. An output that sounds unusually certain about a controversial subject is a warning sign.
• Internal coherence combined with external inconsistency. The output is logically consistent within itself but contradicts verifiable primary sources.
• A complete answer to a question with no known answer. If an LLM provides a detailed and orderly response to a very specific question about a niche market, the hallucination risk is high.
How to mitigate AI hallucinations in marketing and communication processes
No solution currently eliminates hallucinations. Architectures and processes can, however, reduce their incidence significantly.
RAG: Retrieval Augmented Generation
RAG is the most effective system for business contexts. Instead of answering exclusively from training data, the model first retrieves information from a verified document base, such as the company website, an internal database or approved primary sources, and anchors its answer to those documents. RAG reduces hallucinations by 71% compared with vanilla models, according to AllAboutAI (2025).
For an agency or company producing content with AI, a RAG system means using a model that answers only based on what it can ground and declares when it does not have sufficient information.
Structured prompt engineering
The way a query is formulated significantly influences hallucination probability. Instructions such as “if you are uncertain about a figure, state explicitly that you do not have this information” and “cite the source for every statistic” reduce the risk. A 2025 study reported by Nature found that structured prompt engineering reduced hallucinations by approximately twenty-two percentage points.
Human in the loop: editorial governance
No technical mitigation replaces human review. According to the IBM AI Adoption Index (2025), 76% of businesses using AI in production implemented human verification before deploying outputs. In marketing, every AI-produced article, item of copy or report must pass through an editor with the expertise to verify factual claims before publication. AI accelerates production. The editor guarantees accuracy.
Verification checklist for marketing teams
1. Has every statistic been verified against an accessible primary source?
2. Are quotations attributed to real, identifiable researchers, experts or institutions?
3. Are market figures coherent with the sector sources already known to the team?
4. Is the model’s tone unusually assertive about claims that are difficult to verify?
5. Has the content been checked against a second source, another model or a manual search before publication?
The regulatory framework: the European AI Act and its implications for companies
AI hallucinations are no longer only a technical or reputational issue. From 2026, they also represent a compliance risk. Article 50 of the European AI Act establishes transparency requirements for AI outputs by August 2026. The article states that penalties for non-compliance may reach €35 million or 7% of global annual turnover.
For companies using AI in institutional communication, customer service or decisions affecting consumers, output-verification systems therefore become part of the compliance infrastructure.
The Corallo AI case: artificial intelligence with editorial oversight
At Bliss Agency, the approach to artificial intelligence is structured around a precise principle: AI is an acceleration tool, rather than a substitute for strategic judgement and editorial governance. Corallo AI, Bliss Agency’s artificial-intelligence division, was built as infrastructure supporting brand advisory and communication processes, rather than as a stand-alone content production solution.
Every AI-generated output is verified by professionals with specific domain expertise before it is used in campaigns, editorial content or institutional communication.
Corallo AI’s operating model answers the hallucination challenge directly: AI produces, human expertise verifies and the brand advisor decides. It is an approach aligned with the best practices documented among advanced AI adopters, ensuring that production speed does not compromise the accuracy and coherence of brand messages.
2026 trend: the future of AI hallucinations
The market for AI hallucination-detection and mitigation tools grew by 318% between 2023 and 2025, according to Gartner’s 2025 market report. Four trends define the evolution of the problem:
- Models with greater awareness of uncertainty, incorporating mechanisms that recognize insufficient information and state that limitation instead of inventing.
- Multimodal RAG systems that retrieve not only text, but images, structured data and corporate documents, reducing hallucination risk in multimedia production.
- Automated output verification through real-time fact-checking APIs that compare AI outputs with verified databases before publication.
- Compliance by design, in which verification is built into the AI workflow rather than added afterwards.
FAQ: AI hallucinations
What is meant by an AI hallucination?
An AI hallucination is an output generated by a language model that contains false, invented, or unverifiable information presented with the same syntactic structure and expressive confidence as true information. It is not a conventional software bug. It results from the probabilistic nature of LLMs, which generate text by predicting the most plausible sequence rather than necessarily the most accurate one.
Does ChatGPT lie deliberately?
No. Language models have no intentions and do not lie consciously. A hallucination is the result of a system predicting the next plausible word rather than accessing a verified knowledge base. When the model lacks sufficient information, it may still produce grammatically correct but factually incorrect text. It is an architectural problem, not bad faith.
How can AI hallucinations be recognized?
The main signals are highly specific but unverifiable citations, excessively assertive language on complex subjects, an internally coherent output that contradicts known primary sources, and a complete, articulate response to a question that normally has no certain answer.
Can AI hallucinations be eliminated?
At present, no. They can be reduced significantly through RAG architectures, structured prompt engineering and human-in-the-loop processes. Projections suggest that leading models may approach near-zero rates on standard tasks by 2027, but the risk will remain considerably higher in specialist domains such as law, medicine and finance.
What are the legal consequences of AI hallucinations in Italy?
At European level, the AI Act establishes transparency requirements for AI output and provides significant sanctions for non-compliance. In Italy, advertising content containing false AI-generated claims remains subject to existing consumer and competition law. Courts have also addressed legal filings containing nonexistent AI-generated citations.
Which AI model has the lowest hallucination rate in 2026?
On standardized text-summarization benchmarks cited by the article, the lowest rates were associated with Gemini-2.0-Flash-001 at 0.7%, OpenAI GPT-4o at 0.8%, and Anthropic Claude 3.5 Sonnet at 0.8%. These results concern standardized summaries. Specialist domains show substantially higher rates across all models, and model choice must be evaluated against the specific task.
AI hallucinations will not disappear simply because a newer model is released. They are a structural characteristic of probabilistic language generation and will require competent human oversight for years in communication and marketing contexts.
Sources
Total losses: $67.4 billion (2024) Forrester Research: State of Generative AI 2024 https://korra.ai/the-67-billion-warning-how-ai-hallucinations-hurt-enterprises-and-how-to-stop-them/
4.3 hours/week: $14,200/employee/year Forrester Research, Enterprise AI Cost Analysis 2025 https://fourdots.com/business-impact-of-ai-hallucinations-rates-and-ranks
69–88% hallucination rate in the legal domain Magesh et al., Stanford RegLab + Stanford HAI, “Hallucinating Law”, gennaio 2024 https://hai.stanford.edu/news/hallucinating-law-legal-mistakes-large-language-models-are-pervasive
43–64% hallucination rate in the medical domain MedRxiv 2025: studio su 300 vignette cliniche validate da medici https://suprmind.ai/hub/insights/ai-hallucination-statistics-research-report-2026/
Language 34% more assertive in hallucinations MIT Research, gennaio 2025 (citato da AllAboutAI e Suprmind) https://renovateqr.com/blog/ai-hallucinations
RAG reduces hallucinations by 71% AllAboutAI: AI Hallucination Statistics 2025–2026 https://webcite.co/blog/ai-hallucination-statistics/
12,842 articles removed from platforms in Q1 2025 Suprmind / AllAboutAI Research Report 2026 https://suprmind.ai/hub/insights/ai-hallucination-statistics-research-report-2026/
Hallucination Detection Market +318% (2023–2025) Gartner: Hallucination Detection Tools Market Report 2025 (citato da Suprmind) https://suprmind.ai/hub/insights/ai-hallucination-statistics-research-report-2026/
Caso Deloitte Australia Fortune / AP / CJPI: ottobre 2025 https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund/
EU AI Act: Article 50 Gazzetta Ufficiale UE, Regolamento (UE) 2024/1689 https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX:32024R1689
Adozione AI enterprise 85% (2026) Gartner 2026 (citato da Suprmind) https://www.aboutchromebooks.com/ai-hallucination-rates-across-different-models/
BLISS®
© 2026. ALL RIGHTS RESERVED