Artificial Intelligence: discover which model minimizes hallucinations the most – a surprising ranking!

In a world where artificial intelligence increasingly infiltrates our daily and professional activities, the issue of error minimization, particularly hallucinations, becomes crucial. Hallucinations, in this context, refer to invented or erroneous information generated by AI models, which continue to pose a major challenge to the reliability of machine learning systems and natural language processing. In 2025, an unprecedented ranking published jointly by Terzo and Visual Capitalist highlights the uneven performance of leading AI models: from the most reliable to the most prone to hallucinations. This guide is based on press extracts submitted to several AIs that had to identify the precise origin of the information with citation of the source and URL.

The result is unequivocal: error rates vary from one to four times depending on the systems used, demonstrating a surprising disparity in the ability to produce accurate and trustworthy content. In particular, some popular tools display still too high hallucination rates, calling into question their use in professional contexts without rigorous human verification. This in-depth analysis offers a new perspective on the reliability of the most popular models, especially when they are used to support strategic decisions or automate complex processes. Through this surprising ranking, it appears that the most accurate AI model does not necessarily correspond to the most publicized or paid one, highlighting the challenges to be met for the future of artificial intelligence.

Hallucinations in Artificial Intelligence: Understanding Origins and Issues

Hallucinations in the context of artificial intelligence refer to cases where an AI model generates incorrect, invented, or unverifiable content. This can range from slightly inaccurate information to completely fictional facts, giving a false impression of credibility. This phenomenon largely results from the operation of models based on neural networks and machine learning, where the algorithm tries to anticipate the continuation of a text based on vast amounts of data, without true understanding.»

One of the key principles of current systems is their training on massively collected data from the internet, books, articles, and other textual corpora. However, these sources are not always free of errors or perfectly structured. When the algorithm attempts to generate a precise response, it combines this information based on statistical probability, which can lead to confusion or erroneous extrapolations. In 2025, despite significant progress in natural language processing, this phenomenon persists, particularly in citation and precise source attribution tasks.

Here are the main reasons for these hallucinations:

• Training data limitations: Models depend on the quality of the input data. Noisy or biased data generate erroneous results.
• Lack of real context or understanding: AI predicts words or phrases based on statistical models without truly “understanding” the content.
• Generalization issues: Some rare or unprecedented concepts can be misinterpreted by the model.
• Absence of effective self-correction capability: Many systems do not signal when they are uncertain, limiting automatic correction.

This context highlights a fundamental challenge for artificial intelligence stakeholders: improving AI performance by strengthening trust in the proposed results, especially in fields like documentary research, automatic report writing, or strategic decision-making. Hallucinations have consequences that go far beyond the technical scope and also touch on the ethics and responsibility of companies and developers.

Origin of Hallucinations	Impact on Reliability	Concrete Example
Noisy data in training sources	Repeated errors in content generation	False attribution of a scientific article to the wrong journal
Statistical modeling without real context	Approximate answers without proof	Reference to erroneous historical facts in an analysis report
Lack of self-assessment of uncertainty	Propagation of unreported errors	AI generating a nonexistent or incorrect URL

Comparative Analysis of AI Models: Which Are the Most Effective at Minimizing Errors?

The recent study by Terzo and Visual Capitalist orchestrated a test on several AI models by providing them with press extracts requiring a precise citation, including the publication name, specific article, and URL. These simple but rigorous criteria allow evaluating each system’s ability to avoid hallucinations.

The results show a significant wide disparity:

• Perplexity: with an error rate of 37%, it is the best performer in the test.
• Copilot: closely follows at 40%, confirming its solid performance.
• Perplexity Pro: slightly rises to 45% errors.
• ChatGPT Recherche: reaches a rate of 67%, revealing significant shortcomings.
• Deep Research: at 68%, it also shows its limits.
• Gemini: at 76% indicates strong difficulties generating reliable factual citations.
• Grok-2: at 77% confirms this trend.
• Grok-3: peaks at 94%, an alarming hallucination rate.

This surprising ranking reveals models sometimes considered performant but that fail to limit errors when it comes to attributing exact sources. An interesting point is that paid versions do not always outperform free options.

AI Model	Hallucination Rate	Paid Version
Perplexity	37%	No
Copilot	40%	No
Perplexity Pro	45%	Yes
ChatGPT Recherche	67%	Yes
Deep Research	68%	Yes
Gemini	76%	No
Grok-2	77%	No
Grok-3	94%	No

Experts emphasize that this ranking should encourage a cautious approach when using AI models for tasks where information security is critical. Perfection is still far away, and human interaction remains indispensable for verifying and validating results.

Why Minimizing Hallucinations Is a Crucial Challenge for the Future of Artificial Intelligence

Reducing errors and hallucinations in AI models has become a central issue to ensure the reliability of models in critical contexts. In 2025, their deployment is widespread across all sectors, whether health, finance, justice, or communication. Yet, each hallucination can cause serious consequences, economically, legally, and socially.

Here are the main challenges related to hallucination minimization:

• Complexity of training data: Integrating higher-quality sources while maintaining a sufficiently large corpus for learning is complex.
• Neural network architecture: Models must constantly adapt to better capture contexts and avoid erroneous generalizations.
• Need for human validation: Integrating collaboration with human experts to improve accuracy and detect hallucinations.
• Transparency and explainability: Users must be able to understand how and why an AI produced certain results.
• Development of automatic verification tools: To identify and correct errors before they are disseminated.

For example, in the medical field, an AI model hallucinating on diagnoses or treatments can compromise patient lives. Similarly, in finance, a misattribution of a source or figure can lead to costly decisions on a global scale. Thus, minimizing hallucinations is vital to ensure the credibility and sustainable adoption of AI technologies.

Challenges	Possible Consequences	Proposed Solutions
Health	Incorrect diagnosis, inadequate treatment	Enhanced validation by medical professionals
Finance	Incorrect economic decisions	Human supervision and automated audits
Justice	False legal interpretation, legal risks	Close collaboration with legal experts
Communication	Spread of false information, loss of trust	Automated fact-checking tools

To progress, researchers are working on hybrid models combining artificial intelligence and human intervention, as well as on automatic quality control techniques using specialized neural networks to detect errors.

The Diversity of Approaches in Combating AI Model Hallucinations

Improving AI performance against hallucinations does not rely solely on better data collection or longer training. Several innovative strategies are implemented to strengthen the precision and rigor of models in their responses.

The main approaches include:

• Integration of verified corpora: Using carefully selected journalistic, scientific, or institutional databases.
• Targeted supervised learning: Training neural networks with labeled samples to better recognize true sources.
• Self-assessment mechanisms: Some systems develop an uncertainty probability which they can signal.
• Model fusion: Combining several different models to cross-check information and reduce the risk of errors.
• Use of chain-of-thought reasoning: To explain their approach and better understand the context before producing an answer.

For instance, Perplexity benefits from rigorously validated documentary bases, explaining partly its advantage in minimizing hallucinations compared to other models. In contrast, Grok-3, despite its power, shows lower effectiveness especially when it comes to precisely citing its sources.

Strategies	Description	Application Example
Verified corpora	Data sorted and validated for quality and reliability	Scientific base used by Perplexity
Supervised learning	Models trained with labeled data	ChatGPT Recherche uses this method
Self-assessment	Detection of uncertainty in the generated answer	Prototype in testing phase in some AIs
Model fusion	Combination for cross-referencing information	Deep Research
Chain-of-thought reasoning	Intermediate explanation of results	Advanced techniques in Gemini

Innovations in these areas should gradually reduce error rates while increasing end-user trust.

Impact of Hallucinations on Business Decision-Making: Risks and Best Practices

In professional contexts, the integration of artificial intelligence cannot happen without measuring the potentially destructive impact of hallucinations. When these AI models are used to generate reports, support analyses, or automate decisions, each error can result in significant costs, time losses, and even competitive weakening.

The major risks include:

• Strategic decisions based on erroneous information: A report produced with incorrect citations can misdirect strategy.
• Domino effect on automated processes: An initial falsified piece of data can ripple across various departments, degrading overall operational quality.
• Damage to reputation: Repeated errors, spread through various channels, can severely harm company credibility.
• Exposure to legal risks: Incorrect source attribution can lead to litigation or convictions.

Consequences	Business Example	Recommended Preventive Measures
Poor strategic decision	Analysis based on unchecked data	Systematic manual validation
Operational problems	Erroneous process automation	Human supervision of AI actions
Loss of credibility	Dissemination of false information	Error detection training
Legal disputes	Incorrect citation leading to legal action	Use of strict verification protocols

To limit these risks, AI performance must imperatively be accompanied by human intervention. A model built on controlled corporate data often proves more reliable than generic solutions. Moreover, establishing internal control and training protocols helps detect AI-produced errors earlier.

Paid vs Free AI Models: A Surprising Battle on Reliability and Hallucinations

A surprising point revealed by the study is the sometimes minimal difference, or even cases where the paid version does not outperform the free version in minimizing hallucinations. In 2025, the trend often encouraging opting for a paid subscription implying better quality does not necessarily guarantee better reliability.

The main reasons are:

• Identity of training data: Several digital models share similar learning bases, regardless of access cost.
• Different publishers’ objectives: Some prioritize volume and speed over thorough result verification.
• Common technical limitations: No model can yet completely eliminate hallucinations.
• Lack of advanced error detection mechanisms integration: Often absent even in premium offers.

Model Type	Tendency in Hallucinations	Expected Advantages	Actual Impact on Reliability
Free	Sometimes as good or better	Accessibility, speed	Variable depending on cases
Paid	Not always better at minimizing errors	Additional features, support	Often disappointing on reliability

For example, Perplexity Pro, the paid version, shows a higher hallucination rate than Perplexity’s free version in the analysis of precise citations. This invites users to analyze offers more deeply beyond the simple price, favoring the intrinsic quality of models.

Best Practices for Integrating Artificial Intelligence While Controlling Hallucination Risks

To fully harness the capabilities of AI models while minimizing risks generated by hallucinations, it is essential to adopt rigorous practices in their implementation and supervision.

Here are some key recommendations:

• Establish systematic verification protocols: Enforce human control for each sensitive production.
• Train teams to detect and report errors: Raise awareness among professional users.
• Use models adapted to business contexts: Favor solutions trained on specific internal data.
• Implement hybrid systems: Couple AI and human expertise for better reliability.
• Regular monitoring and updates: Keep models up to date with fresh and validated data.

A fictitious company, “NovaTech”, illustrates these principles well. As soon as NovaTech integrated an AI model for automatic report writing, it established a dual validation flow, with human experts reviewing each output before release. This process significantly reduced error risks and improved internal trust in the tools.

Recommended Practice	Objective	Concrete Example
Systematic human control	Detect errors before publication	NovaTech reviews each AI report
Continuous training	Increase user vigilance	Monthly sessions for staff
Company-specific data models	Ensure response relevance	Specific training on internal documentation
Hybrid system	Combine AI and human expertise	NovaTech dual validation

Future Perspectives for Reliability and Hallucination Reduction in AI

Advances in artificial intelligence promise significant improvements in model reliability in the coming years. Several avenues are being explored by the scientific community and industry.

Among the expected innovations are:

• Multimodal models combining text, image, and structured data: to anchor answers in richer contexts.
• Better integration of user feedback: enabling AIs to continuously learn from their mistakes.
• Approaches to automatic cross-verification: multiplying sources and confronting answers.
• Advanced techniques in explainable AI (XAI): to understand and justify AI reasoning.
• Increased personalization: adapting models to the specific needs of companies or individuals.

These innovations should reduce hallucination rates and increase trust in AI solutions. Nevertheless, human intervention will certainly remain, in the near future, an indispensable safeguard for data security and process control.

Future Innovations	Expected Benefits	Impact on Hallucination Minimization
Multimodal models	Richer and more reliable context	Reduction of contextual errors
User feedback	Continuous improvement through learning	Decrease in repeated hallucinations
Automatic cross-verification	Stronger validation	Fewer erroneous publications
Explainability (XAI)	Understanding of AI decision	Better user trust
Adapted personalization	More targeted responses	Error reduction

The Complementarity Between Artificial Intelligence and Human Expertise to Gain Reliability

As artificial intelligence advances, collaboration between automated systems and human experts appears as a pragmatic solution to mastering hallucination risks. Using an AI model without control can prove counterproductive, even dangerous.

The benefits of this approach include:

• Double verification: a human expert can identify inconsistencies or errors the AI does not detect.
• Guided learning: human feedback helps refine model training.
• Consideration of business context: often complex and subtle, specific context sometimes escapes algorithms.
• Ethics and responsibility: a human ensures decisions fit within legal and moral frameworks.

In industry, several cases have shown that this complementarity significantly reduces AI system error rates. For example, a legal consulting firm established a workflow where AI prepares a first draft, then an expert lawyer validates and adjusts the content before publication.

Advantages of Complementarity	Description	Application Example
Double control	Limits errors before release	Validation by a legal expert
Model refinement	Feedback on errors for learning	Retraining based on human feedback
Contextualization	Taking business specificities into account	Capturing sector nuances
Ethical responsibility	Ensures compliance and ethics	Human supervision in critical decisions