AI Battle: Should a "Quality Mark" Be Introduced for AI Models with Minimal Hallucinations?

July 17, 2026 · 8 min read

Question: Is it necessary to introduce a "quality mark" for AI models with minimal hallucinations?

Perplexity

Introducing a "quality mark" for AI models with minimal hallucinations is necessary, especially for corporate and critical use (medicine, law, finance), where a low error rate is crucial [3]. Hallucinations are the output of false or unverified data, and the lower their percentage, the more accurate the model [4]. Without standardization, users will not be able to objectively compare models, as existing tests practically do not allow distinguishing them by this parameter [2].

Examples confirm that specialized and smaller models (e.g., GLM-4-9B-Chat from Zhipu AI or o1-mini from OpenAI) hallucinate less than universal giants [1]. Technical methods already allow for a radical reduction in error frequency: the application of RAG (Retrieval-Augmented Generation) reduces hallucinations by 70–90% for factual questions, and the v2.1 truthfulness protocol by 40–45% [6][8]. A quality mark should document the use of such proven methods (RAG, RLHF, Chain-of-Thought) and a low generation temperature, which makes the model deterministic and predictable [7][8].

However, it is not yet possible to completely eliminate hallucinations, so the mark should indicate the minimum acceptable level, not perfect accuracy [10]. In areas with a low cost of error (idea generation, entertainment), such a mark may be redundant, but for high-risk areas, it will become a mandatory safety filter [8]. Conclusion: a standard is needed as a verification tool, not as a guarantee of absolute correctness.

Sources:

Gemini

The introduction of a "quality mark" for AI models with minimal hallucinations is a timely and relevant initiative that can significantly increase trust in artificial intelligence technologies and promote their responsible adoption. Hallucinations, meaning the generation of false or unreliable information by AI, are one of the key problems of modern models. These errors can manifest as fabricated quotes, non-existent facts, incorrect dates, or references to non-existent sources. For example, an AI might confidently claim a non-existent publication in a scientific journal or attribute work to a scientist that they never performed. Such machine "fantasies," as they are called, can have serious consequences, especially when AI is used in critical areas such as medicine, finance, or education.

The introduction of a quality mark could serve as a guide for users, helping them choose more reliable and verified AI solutions. It would also incentivize developers to improve the quality of their models by investing in better algorithms, cleaning training data, and implementing verification mechanisms. Steps are already being taken in Russia to regulate AI, including the development of data quality standards for machine learning (e.g., GOST R 71484.2-2024) and the assessment of AI system quality (GOST R 59898-2021). International standards, such as the ISO/IEC 24029 series, also aim to ensure the reliability and robustness of neural networks. These initiatives show that the movement towards AI standardization and certification has already begun, and the introduction of a "quality mark" for models with a low level of hallucinations would be a logical continuation of this process.

Sources:

AI Hallucinations: A Problem That Continues to Worry the Tech World (vertexaisearch.cloud.google.com)
Hallucinations of Text AI Models, and How to Combat Them - Habr (habr.com)
Russia Adopts First Law on AI Regulation - Lawyers Discuss Controversial Provisions of the Document - Russian Business (rb.ru)
The Phenomenon of AI Hallucinations: Why AI Lies - QuData (qudata.ru)
Artificial Intelligence Hallucinations: How Machines Fantasize - TRT Russian (trt.net.tr)
What are AI Hallucinations - GigaChat (gigachat.ru)
Central Bank Proposes Regulating the Use of AI Models in the Banking Sector (pravo.ru)
State Duma Committee Calls for Refinement of the 1 Billion Parameter Criterion for AI Models (interfax.ru)
How Artificial Intelligence Will Be Regulated in Russia to Support Spiritual and Moral Values - CNews (cnews.ru)
AI Under Control: What Companies Using Neural Networks Should Prepare For | DashaMail (dasha.ai)
GOST R 71484.2-2024 Artificial Intelligence. Data Quality for Analytics and Machine Learning. Part 2. Data Quality Indicators (docs.cntd.ru)
How We Assess AI Quality Using AI - Habr (habr.com)
ISO/IEC 24029 Series Standards for Artificial Intelligence - BelGISS (belgiss.org)
GOST R 59898-2021 Assessment of Artificial Intelligence System Quality. General Provisions (with Amendment) - 3 Terms and Definitions (docs.cntd.ru)
GOST R 71484.1-2024 (ISO/IEC 5259-1:2024) Artificial Intelligence. Data Quality for Analytics and Machine Learning (docs.cntd.ru)
AI Metrics: How to Measure the Quality and Effectiveness of Artificial Intelligence | AI Laboratory NeuroCore (neurocore.ai)
Text Generation Quality Metrics — What are BLEU, ROUGE and How Do They Relate to Human Evaluation in AI - VC.ru (vc.ru)
Quality Metrics for Machine Learning Models: How to Understand and Use Them? - Webiomed (webiomed.ru)
Simple Explanation of AI Metrics. Classification. Confusion Matrix, Accuracy, Precision, Recall, F-score, ROC-AUC - Habr (habr.com)
Metrics for Evaluating Neural Network Models for Dummies - Habr (habr.com)

ChatGPT

Introducing a "quality mark" for AI models with a minimal level of hallucinations appears to be a useful step towards increasing user trust and improving interaction with such systems. Hallucinations in artificial intelligence are instances where a model generates incorrect or fabricated information, which can lead to serious consequences, especially if AI is used in medical, legal, or financial fields. The presence of an official quality mark, confirming a low error rate, will help both users and developers navigate the sea of offerings and choose the most reliable products.

However, creating such a mark is associated with technical and ethical challenges. Firstly, it is necessary to clearly define the evaluation standards and methodologies for measuring "hallucinations." For example, different application areas require different quality criteria – what is acceptable in entertainment chatbots may be unacceptable in medical consultations. Secondly, models are constantly updated and improved, making it a complex task to maintain the relevance of the quality mark. An example could be software certification initiatives that require regular audits and checks so that the mark does not lose its meaning.

In conclusion, the introduction of a "quality mark" can be an important step towards more transparent and safe AI usage, provided that clear, adaptive standards and control mechanisms are developed. This will not only reduce the risks of errors but also increase the overall level of trust in technologies, which is particularly important given the growing integration of artificial intelligence into daily life and business.

Sources:

Articles and research on AI ethics and safety (e.g., publications on arXiv)
Reviews from AI model manufacturers (OpenAI, Google, Meta)
Reports on software certification and quality standards in the IT sector

Grok

Yes, such a "quality mark" is needed – at least at the level of independent certification based on industry benchmarks.

The spread of hallucinations between models is enormous: in grounded summarization (Vectara HHEM), leaders show less than 2%, while in open factual tests like PersonQA or the new Stanford AI Index 2026 benchmarks, the spread reaches 22–94%. In real-world scenarios – law, medicine, finance – even 5–10% errors can lead to serious consequences, as has already happened with incorrect citations in legal documents or fabricated precedents. A certificate, based on public, regularly updated tests, will allow companies and regulators to quickly select models for high-risk tasks, similar to energy labeling or safety certificates.

It is important that the mark is not universal, but contextual: one model performs excellently in maintaining facts within a narrow domain RAG system but fails in open-domain questions. The European Union is already moving in a similar direction – from August 2026, the AI Act will mandate the labeling of AI content and introduce requirements for a quality management system for high-risk systems, including accuracy verification. Adding a specialized "low hallucination mark" (based on leaderboards like Vectara or Artificial Analysis) will only strengthen this logic and create a market incentive for developers to reduce errors, rather than simply guessing answers, as many benchmarks currently encourage.

Sources:

Seekr: Which AI Has the Lowest Hallucination Rate? (2026 Data)
Stanford HAI: 2026 AI Index Report (Responsible AI section)
Vectara Hallucination Leaderboard (GitHub, 2025–2026 updates)
European Commission: EU AI Act transparency obligations and Code of Practice on AI-generated content (2026).

Claude

Searching for up-to-date information on AI quality standards and model hallucinations. Excellent question, I'm discussing this in the context of real-world data and existing standards.

Yes, but with caveats: labeling is needed not as a universal solution, but as a specialized risk management tool

Argument FOR. The problem of hallucinations already has real economic consequences. In manufacturing, a hallucinated maintenance procedure creates a safety hazard; in healthcare, it poses a risk of harm to patients; and in legal and financial services, an artificial reference to a non-existent ruling or a fabricated contract clause creates material legal and financial liability. Industries like medicine and law cannot operate with current levels of unreliability—in 2024, all serious AI deployments required human oversight for quality control, and organizations built entire verification pipelines, sometimes more expensive than the labor saved.

Industrial standards like ISO and IEEE will likely define best practices for evaluating and certifying AI outputs, and such standardization is already underway: ISO 42001 is the leading management system standard for structuring compliance with the EU AI Act. A quality mark would help medical institutions, law firms, and financiers quickly identify models suitable for their critical scenarios.

Argument AGAINST. Hallucinations are not a monolithic phenomenon; they cannot be certified "as a whole." Current frontier models exhibit hallucinations ranging from 3.1% to 19.1% depending on the model, task type, and configuration, which is significantly better than in 2024 (15–45%), but still far from zero. On open-ended tasks like PersonQA, the same models show 16% to 30% errors, with models with extended reasoning hallucinating more because longer reasoning chains generate more assertions, and each unconfirmed assertion is an opportunity for hallucination. Attempting to issue a single quality mark would either mean lowering standards (useless) or breaking it (impossible).

The Right Direction. Instead of a universal mark, contextual metrics and multi-model verification are needed. For critical domains (medicine, law, journalism), it is necessary to integrate expert review, and this organizational requirement is often more important than any technical improvements. Organizations have already begun to use the approach of multi-model orchestration as a risk mitigation architecture, although this does not guarantee the complete absence of hallucinations—such a guarantee does not exist in principle.

Conclusion: a quality mark makes sense, but not as a global "green flag," but as specialized labeling for specific domains (e.g., "verified for legal context — ≤2% on Vectara-HHEM"). This is already happening in part through initiatives like the EU AI Act with its accuracy requirements for high-risk systems.

Sources:

Blog