Humans Do Not Record Reality Perfectly

Artificial intelligence is often criticised for giving false information, showing bias, inventing facts, and presenting wrong answers with great confidence. These problems are real and must be taken seriously. However, this argument has an important weakness. Artificial intelligence is often compared with an ideal human being assumed to be rational, objective, and reliable. In reality, human beings also make many errors in the way they understand the world.

Humans can remember events incorrectly. We can believe misinformation. We can see patterns that are not really there. We can be influenced by our culture, religion, education, family, political views and social environment. We often prefer information that supports what we already believe and reject information that challenges us. We can also be extremely confident about something and still be completely wrong.

This does not mean that human and artificial intelligence hallucinations are the same. They are not. A human hallucination normally refers to seeing, hearing or experiencing something without the expected external stimulus. An artificial intelligence “hallucination” normally means that the system produces information that sounds convincing but is false, invented or unsupported by evidence.

The article therefore introduces a broader concept called Reality Divergence. Reality Divergence means the picture of reality inside a human mind or an artificial intelligence system differs from the best available evidence about reality. Humans and artificial intelligence can both experience Reality Divergence, though for different reasons.

This changes the central question. Instead of asking, “Can artificial intelligence be trusted?”, we should ask, “For this particular task, under these particular conditions, which system is more reliable: a human, artificial intelligence, or humans and artificial intelligence working together?”

This is important because artificial intelligence was created and trained using enormous amounts of information produced by humans. That information contains the best of human knowledge, including science, engineering, medicine, mathematics, history and literature. However, it also contains human mistakes, prejudices, propaganda, misinformation, cultural assumptions and conflicting beliefs. Artificial intelligence therefore inherits part of the human information environment.

The article develops this idea into the Cognitive Mirror Hypothesis. This is the paper’s first major TAPiO concept. The hypothesis suggests that artificial intelligence can partly mirror human civilisation. When we find bias, stereotypes or misinformation inside artificial intelligence, we should investigate not only what happened inside the machine but also where the original information came from. In some cases, artificial intelligence may reflect patterns that already exist in human society.

However, artificial intelligence also has an important advantage. A human being lives only one life and experiences a limited number of cultures, countries and languages. Artificial intelligence can potentially examine information from many different cultures and perspectives. It can compare Malaysian, German, Chinese, African, American and other viewpoints. It can deliberately search for evidence against its own first conclusion. This creates what the article calls cognitive pluralism, where important questions are examined from several serious perspectives before a final decision is made.

Artificial intelligence also has weaknesses. One important problem is sycophancy, where an AI system agrees too strongly with the user. If a chairman already believes an investment is excellent and asks AI to explain why it will succeed, the AI may produce an impressive argument supporting him. This can strengthen the chairman’s existing confirmation bias instead of challenging it.

A better approach is to have AI challenge the proposition. It should first build the strongest argument supporting the investment. It should then try to prove that the investment will fail. It should search for contradictory evidence, identify hidden assumptions and examine what information is missing. Only after this process should it make a recommendation.

This turns AI from an answer machine into an adversarial reasoning partner.

The second major TAPiO concept is Artificial Epistemic Intelligence, or AEI. We usually judge normal artificial intelligence by whether it can produce the correct answer. AEI goes further. It asks whether the AI understands how strongly its answer deserves to be believed.

A strong epistemic AI should be able to explain where its evidence came from, distinguish primary evidence from secondary commentary, identify evidence against its conclusion, recognise uncertainty, identify assumptions and explain what new evidence would change its conclusion. Importantly, it should also be capable of saying, “I do not know.”

This may become more important than simply making AI models larger or more intelligent.

The third major TAPiO concept is the Comparative Reliability Index, or CRI. The CRI proposes a 100-point system for measuring the reliability of human intelligence, artificial intelligence, and human-AI collaborative intelligence.

Accuracy is important, but accuracy alone is not enough. The proposed CRI therefore measures twelve areas: factual accuracy, calibration of confidence, quality of sources, resistance to bias, ability to falsify its own conclusion, willingness to correct mistakes, understanding of context, resistance to manipulation, independence of evidence, awareness of the consequences of being wrong, reproducibility and auditability.

This allows a much more meaningful comparison.

A system that is correct 92 per cent of the time but is extremely confident when wrong, cannot provide evidence and refuses to correct itself may be more dangerous in some situations than a system that is correct 88 per cent of the time but clearly recognises uncertainty and rapidly corrects mistakes.

The article proposes scientifically testing the CRI using approximately 300 professional human participants from areas such as science, medicine, engineering, economics, finance, law, government, business strategy and international affairs. Several leading artificial intelligence systems should be tested against exactly the same problems.

Three main groups should be compared. The first would consist of human experts working alone. The second would consist of artificial intelligence systems working alone. The third would consist of humans working with artificial intelligence through a carefully designed TAPiO Collaborative Intelligence protocol.

The experiment should test factual accuracy, false information, confabulation, confirmation bias, cultural bias, misinformation, uncertainty, source quality, manipulation and the ability to change a conclusion when better evidence becomes available.

Importantly, the research must also try to prove the TAPiO theories wrong. This is essential for scientific credibility. The Comparative Reliability Index, Artificial Epistemic Intelligence and Cognitive Mirror Hypothesis are currently proposed research concepts. They should not be presented as scientifically proven until independent testing and validation.

Existing scientific research also provides an important warning. Humans and artificial intelligence working together do not automatically perform better than the best human or the best AI working alone. The way collaboration is designed matters enormously.

This leads to the article’s most important conclusion.

The future should not simply be Human versus AI.

It should be Human plus AI, deliberately designed to correct each other’s weaknesses.

Humans provide purpose, responsibility, cultural understanding, real-world experience, social judgement, and accountability. Artificial intelligence provides enormous information-processing capacity, speed, comparison, contradiction detection, scenario analysis, and the ability to examine many perspectives.

The machine should challenge human bias. The human should challenge machine confabulation. Both should search for evidence. Both should question their first conclusions. Both should recognise uncertainty. Both should be prepared to say, “I may be wrong.”

The final question is therefore no longer simply, “Who really hallucinates?”

Humans do. Artificial intelligence does. They simply do it differently.

The more important objective is to understand both sets of weaknesses and use that knowledge to build something better.

That new system is Collaborative Intelligence.

Perhaps the greatest contribution of artificial intelligence will therefore not be that machines learn to think like humans. It may be that, by studying the weaknesses of artificial intelligence, humanity finally becomes better at recognising the weaknesses in its own thinking.

Artificial intelligence is often criticised for producing false information, showing bias, and giving incorrect answers with strong confidence. These problems are real. However, this comparison is incomplete because artificial intelligence is often compared with an ideal human assumed to be rational, objective, and reliable. Real human beings are not like this. Humans can have false memories, follow misinformation, defend prior beliefs, respond to social pressure, inherit cultural assumptions and remain highly confident while being wrong. Artificial intelligence has different failure mechanisms. Large language models can generate information that sounds correct but is unsupported, reproduce patterns found in training data, agree too strongly with users, and fail to recognise uncertainty. This manuscript argues that the useful scientific question is therefore not whether artificial intelligence makes mistakes. The useful question is which cognitive system is most reliable for a defined task under defined conditions. It develops three original TAPiO concepts. The Comparative Reliability Index, or CRI, is proposed as a 100-point framework for measuring reliability across twelve dimensions. Artificial Epistemic Intelligence, or AEI, describes artificial intelligence that evaluates not only an answer but also the strength, origin, independence, uncertainty and falsifiability of the evidence behind that answer. The Cognitive Mirror Hypothesis proposes that artificial intelligence partly reflects the knowledge, bias, cultural patterns, misinformation and contradictions of the human information environment from which it is trained. The manuscript then proposes a controlled research design comparing expert humans working alone, leading artificial intelligence systems working alone and humans working with artificial intelligence under a structured collaborative protocol. The goal is not to prove that either humans or artificial intelligence are generally superior. The goal is to measure each system’s characteristic failure patterns and test whether deliberately engineered collaborative intelligence can achieve higher comparative reliability than either acting alone.

Keywords: artificial intelligence, human cognition, hallucination, confabulation, cognitive bias, misinformation, false memory, calibration, epistemology, human AI collaboration, comparative reliability, CRI, Artificial Epistemic Intelligence, Cognitive Mirror Hypothesis

Public discussion about artificial intelligence often begins with a warning: artificial intelligence cannot be trusted because it hallucinates. The warning contains an important truth. Large language models can produce convincing information that is false or unsupported. Even advanced models continue to show this problem. Kalai and colleagues reported in Nature in 2026 that hallucination remains present in advanced language models and argued that common accuracy based evaluation can reward guessing instead of admitting uncertainty. This matters because a system can appear more capable on a benchmark while becoming less trustworthy in a real decision environment.

The wider argument, however, rests on a hidden assumption. Artificial intelligence is often compared with an imaginary human who is perfectly rational. Such a human does not exist. Human cognition also contains systematic weaknesses. Research shows that people can believe misinformation, remain influenced by it after correction, favour evidence that supports existing beliefs, misremember events, and become highly confident in incorrect conclusions. Human error is therefore not an unusual exception. It is part of normal cognition.

This paper does not argue that human hallucination and artificial intelligence hallucination are the same process. They are not. It instead asks a more useful question. When a human, an artificial intelligence system, and a human-AI team face the same problem, which configuration produces the most reliable judgement? This changes the debate from a philosophical contest between humans and machines into a measurable problem of comparative reliability.

The paper proposes that reliability should not be treated as a single measure of factual accuracy. A system can be accurate most of the time but still be dangerous if it is badly calibrated, cannot show where its evidence came from, cannot recognise uncertainty, is easily manipulated, cannot correct itself or repeatedly follows the same hidden bias. Reliability therefore needs a wider measurement framework.

The central proposition of this manuscript is that the future of reliable intelligence may depend less on whether humans or machines are individually more intelligent and more on whether we can build cognitive systems in which the characteristic weaknesses of each are actively detected and corrected by the strengths of the other.

The correct benchmark is not perfection. The correct benchmark is comparative, measurable, and continuously improving reliability.

The word hallucination is used differently in medicine, psychology and artificial intelligence research. In clinical psychology, a hallucination normally means a perceptual experience that occurs without the expected external stimulus. A person may hear a voice when no external voice is present. An artificial intelligence language model does not normally experience this kind of perceptual event.

In artificial intelligence research, hallucination usually means that a system generates information that appears reasonable but is false, unsupported or inconsistent with the available evidence. Farquhar and colleagues used the term confabulation for certain forms of unsupported language model output and showed that semantic entropy can help identify cases in which a model is uncertain about the meaning of its possible answers.

It is therefore scientifically wrong to claim that human hallucination and artificial intelligence hallucination are the same cognitive event. However, both can produce an important practical similarity. Both can produce a representation that does not correspond well with external reality.

To compare these systems without conflating their mechanisms, this manuscript uses the broader term reality divergence. Reality divergence means a material difference between an agent’s representation of reality and the best available evidence about that reality. The term can include false memory, motivated reasoning, perceptual illusion, unsupported machine generation, misinformation effects, source failure and other mechanisms that lead to an incorrect model of the world.

Human memory often feels like a record of the past. Scientific research shows that it is not a perfect recording system. Memory is reconstructive. Later information, expectations, emotion, and how a question is asked can affect what people remember.

Research on misinformation shows that incorrect information can influence later belief and memory, and that correction does not always remove that influence completely. Ecker and colleagues reviewed the psychological drivers of misinformation belief and its resistance to correction, showing that misinformation can remain influential even after people are told it is wrong.

This parallels a major public concern about artificial intelligence. Artificial intelligence is criticised when it gives a wrong answer with high confidence. Humans can also be wrong with high confidence. A person’s statement that he remembers something clearly may be completely sincere and still be inaccurate.

Confidence is therefore not the same as truth. A reliable cognitive system needs calibration. Calibration means that expressed confidence should match actual correctness over many comparable cases. A system that repeatedly expresses 90 per cent confidence while being correct only 60 per cent of the time is overconfident. A system that is correct 90 per cent of the time but expresses only 60 per cent confidence is underconfident. Neither system is well calibrated.

Bias is sometimes described as a problem affecting only irrational, ideological or badly educated people. That view is incorrect. Human cognition uses shortcuts because no person can fully analyse every piece of information that enters the mind. These shortcuts can be useful, but they can also produce systematic error.

Confirmation bias is one example. People often give more attention to information that supports something they already believe. Motivated reasoning is another. People can evaluate evidence differently because of what they want to believe, the group they belong to, or the social cost of changing their position.

Political psychology provides clear evidence. Ditto and colleagues reported in a 2025 Annual Review of Psychology article that partisans show systematic tendencies to seek, believe and remember information that supports their political beliefs and affinities. This does not mean that all political judgement is irrational. It means identity and prior belief can influence how people process evidence.

The same problem can appear in business, investment, government, science and private life. An executive may become attached to a project he created. An investor may ignore negative evidence after making a large commitment. A government may defend a policy because changing it may appear weak. A scientist may become attached to a theory. A person may keep believing something because everybody around him believes it.

Intelligence does not automatically remove these weaknesses. A highly intelligent person may sometimes become better at constructing strong arguments for an incorrect position. This is why intelligence and epistemic discipline must be separated. Intelligence is the ability to process information and solve problems. Epistemic discipline is the ability to change a belief when the evidence no longer supports it.

Artificial intelligence is frequently criticised because it is shaped by its training data. That criticism is valid, but it is incomplete. Humans are also trained by their environment.

A child learns language from other humans. The child learns ideas about right and wrong, danger and safety, authority, family, religion, nationality, social behaviour and cultural expectations. Much of this learning happens before the child develops strong critical reasoning. The child does not conduct a scientific review before accepting its first model of the world. It absorbs information from its environment.

Artificial intelligence learns through very different technical mechanisms, but the analogy still holds. Large language models are trained on enormous quantities of human generated information. This material contains science, mathematics, engineering, history, literature and practical knowledge. It also contains mistakes, propaganda, prejudice, stereotypes, misinformation, political narratives, cultural assumptions, fear and imagination.

Artificial intelligence therefore entered an already imperfect information environment. It did not create human bias. In many cases it inherited statistical traces of human bias. At the same time, artificial intelligence can also create new problems through model architecture, data selection, optimisation, reinforcement processes, safety systems and interaction design.

The relationship is therefore not one way. Human beliefs influence human information. Human information influences artificial intelligence. Artificial intelligence then influences humans. The result can become a feedback system.

Scientific language must be especially careful when discussing religion, spirituality and culturally transmitted beliefs. It would be incorrect to describe all religious belief as hallucination. Christianity, Islam, Buddhism, Hinduism and other traditions contain historical, moral, philosophical and metaphysical claims. Some claims can be examined empirically. Other claims may not be directly testable with current scientific methods.

The same care is required when discussing belief in ghosts, spirits, black magic, luck, supernatural intervention or unusual personal experiences. The research question should not be whether one culture is intelligent and another is irrational. The more useful question is how beliefs are formed, transmitted, protected, changed and connected to identity.

For this reason, the framework separates empirical knowledge, cultural belief, metaphysical belief, superstition, perceptual hallucination and confabulation. These categories can overlap, but they should not automatically be treated as the same phenomenon. This distinction also protects the research from turning a scientific study of cognition into an unnecessary judgement about religion or culture.