In the field of artificial intelligence (AI), the term hallucinations is used to describe situations in which language models generate incorrect or fictitious information, but present it as if it were true.
This phenomenon represents a crucial challenge, especially when AI is applied in sensitive fields such as medicine, law or education.
In this context, Dario Amodei, CEO of Anthropic, offered statements that have generated debate in the industry. During the Code with Claude event, Amodei stated that AI models “probably hallucinate less than humans, but they do so in more surprising ways.”
Let’s explore in detail his statements and the interesting implications this may have.
What exactly did Dario Amodei say?
During the Code with Claude event, hosted by Anthropic in San Francisco, Dario Amodei answered a question about hallucinations in language models.
During the response, the manager stated: “AI models probably hallucinate less than humans, but they do so in more surprising ways.” This statement, which directly compares the performance of machines with that of people, has not gone unnoticed and has sparked both interest and skepticism.
Amodei also linked this idea to his vision for the future of AI. He reiterated his prediction that AGI could be reached in 2026, based on a paper he published the previous year.
He used a telling metaphor: “Water is rising everywhere,” suggesting that technological advances are occurring widely and rapidly. Amodei’s statement reflects great confidence in AI’s ability to overcome its current limitations, including the propensity to hallucinate.
What are hallucinations in AI?
Hallucinations in language models arise when an AI generates answers that sound plausible but are erroneous, fictitious or have no basis in reality. For example, you may invent bibliographical references, non-existent historical events, or erroneous technical explanations.
A recent case illustrates the severity of the problem: An Anthropic lawyer had to apologize in court after submitting a fake legal citation generated by the company’s model, Claude, that included made-up names and legal references.
The causes of hallucinations are multiple. These include biases or errors in the training data, the lack of specific context, and the limitations of the model architecture.
Although important progress has been made, hallucinations continue to represent a challenge, especially in contexts where precision is essential.
Do humans or AI hallucinate more?
To evaluate Amodei’s claim, it is worth analyzing studies that compare the performance of AI with that of humans. In the healthcare field, for example, AI has demonstrated great precision.
A 2014 meta-analysis determined that approximately 5.08% of primary care patients in the US are misdiagnosed (PMC). In contrast, some AI models have achieved 95% accuracy in diabetes-related diagnoses.
However, in tasks such as generating bibliographic references, errors are frequent. A study published in JMIR reported that models like ChatGPT-3.5 and 4 make errors in 29% to 40% of citations.
In journalism, a 2005 study found that 61% of news stories in American newspapers contained factual errors, with an average of three errors per erroneous article.
These human errors are often due to carelessness or misunderstanding, while AI models can generate complete falsehoods with apparent safety.
In short, although AI outperforms humans in some contexts, Amodei’s claim cannot be considered universally true. The lack of direct comparative studies between human and artificial hallucinations leaves the question open.
Divergent opinions among experts
Amodei’s words have provoked diverse reactions among professionals in the sector. Some share his optimism. According to UX Tigers, more advanced AI models show lower hallucination rates thanks to more refined training techniques.
However, others disagree. Demis Hassabis, CEO of Google DeepMind, has warned that current models have important gaps and fail even on basic questions.
Additionally, New Scientist published a report in May 2025 warning that hallucinations are getting worse in some recent models designed for complex reasoning tasks.
Advances in reducing hallucinations
To address this problem, the technology community has developed various strategies. One of the most notable is Retrieval Augmented Generation (RAG), which allows models to consult external sources during response generation.
A Stanford study published in 2024 showed that this technique, combined with others, reduced hallucinations by 96%.
Methods such as reinforcement learning with human feedback (RLHF) are also being used, which trains models to prioritize more truthful responses.
Knowledge graphs (KGs) provide structured context that reduces errors, and techniques such as chain-of-thought (CoT) promote step-by-step reasoning that improves model transparency.
Another emerging strategy is to train models to recognize their own limits and be able to answer “I don’t know” rather than inventing an answer. This approach helps reduce the spread of misinformation (Forbes).
However, the definitive solution still depends on the context of use and the degree of external verification available.
No scientific consensus
Opinions within the sector are varied, and technical advances, although promising, do not completely solve the problem. Consequently, a rigorous, interdisciplinary and critical approach is required to ensure that AI is not only powerful, but also trustworthy.
As we approach the possible development of AGI, controlling hallucinations will be key to building trust in the technology and ensuring its responsible integration into society.
This post is also available in: