Why Accuracy Alone Is Not Enough

A human being has one biological life. Even a highly educated and internationally experienced person lives in a limited number of places, speaks a limited number of languages and directly experiences only a small part of human civilisation. Every human perspective therefore contains both knowledge and limitation.

Artificial intelligence has a different potential. It can process material created by people from many countries, languages, religions, political systems and academic traditions. This does not make artificial intelligence neutral. Some languages and countries produce much more digital information than others. Some political systems shape information strongly. Commercial interests affect what is published. Training and alignment choices also affect what a model produces.

However, scale creates a new possibility. Artificial intelligence can be instructed to compare perspectives deliberately. A question can be examined from a Malaysian industrial perspective, a German engineering perspective, a Chinese supply chain perspective, an African development perspective, an investor perspective, a labour perspective and an environmental perspective. The purpose is not to treat all perspectives as equally correct. It is to make hidden assumptions visible.

This manuscript calls this cognitive pluralism. Cognitive pluralism means deliberately examining a problem from several serious perspectives before accepting a conclusion. The value lies in moving the observer. What a person fails to see is often partly determined by where they stand.

Artificial intelligence has an important weakness that is especially dangerous in executive and political settings. It can agree too strongly with the user. This behaviour is commonly called sycophancy.

Research published in Science in 2026 reported that sycophantic artificial intelligence is widespread and can influence human intentions and dependence. The practical problem is easy to understand. If a chairman believes an investment is an excellent idea and asks an artificial intelligence system to explain why it will succeed, the model may produce an impressive argument supporting the chairman’s existing belief.

That answer may be intelligent and still be dangerous. The question’s framing has already pushed the analysis toward one conclusion.

A stronger method asks the system to challenge the proposal. It first develops the strongest supporting case. It then attempts to prove that the proposal will fail. It searches for contradictory evidence. It identifies hidden assumptions. It examines who benefits from the proposal. It asks what information is missing and what evidence would change the conclusion.

This changes the role of artificial intelligence. It moves from answer generator to adversarial reasoning partner.

Artificial intelligence can do more than reinforce human belief. Under carefully designed conditions, it can also challenge it.

Costello and colleagues reported in Science in 2024 that personalised dialogues with an artificial intelligence system reduced conspiracy beliefs and that the effect could remain for months. The system addressed each participant’s specific evidence and arguments instead of offering a generic correction.

This result does not prove that artificial intelligence is generally objective. It demonstrates something narrower but very important. Artificial intelligence can sometimes operate as an epistemic counterweight to strongly held human belief. A human friend trying to correct the same belief may face identity conflict, social hierarchy, embarrassment or fear of damaging the relationship. An artificial intelligence system can sometimes sustain a long, evidence-based challenge without those particular human social costs.

The implication is that influence can operate in both directions. Humans can transfer bias into artificial intelligence. Artificial intelligence can also help humans detect weak reasoning.

The interaction between humans and artificial intelligence can create a feedback loop. Glickman and Sharot reported in Nature Human Behaviour that even small perceptual, emotional and social biases originating from humans or artificial intelligence can leave human beliefs more biased over time. People gradually learn from the AI systems they interact with.

This creates a serious risk. Humans create biased information. Artificial intelligence learns from human information. Artificial intelligence repeats part of the bias. Humans see the artificial intelligence output and treat it as independent confirmation. Human confidence increases. People then publish this new information, and it may later become part of future AI training material.

This manuscript calls this an epistemic echo loop. An epistemic echo loop occurs when information appears to be independently confirmed even though the apparent confirmations return to the same original source or assumption.

This is why source independence matters. Ten websites repeating the same press release are not ten independent sources. Five artificial intelligence models trained on strongly overlapping information are also not automatically five independent confirmations. Evidence plurality must be measured structurally, not only numerically.

It is tempting to assume that combining human intelligence and artificial intelligence must produce the best result. Current research shows that this assumption is too simple.

Vaccaro, Almaatouq and Malone analysed 370 effect sizes from 106 experiments comparing humans alone, artificial intelligence alone and human AI combinations. Their 2024 meta-analysis in Nature Human Behaviour found that human-AI combinations performed better than humans alone on average. However, the combinations performed worse on average than the better of the human or artificial intelligence system acting alone. The study also found differences by task type, with performance losses in some decision tasks and stronger gains in some content creation tasks.

This finding is central to the present paper. Human plus artificial intelligence is not automatically collaborative intelligence. The quality of the collaboration depends on task design, role allocation, trust, expertise, timing and the rules used when human and machine disagree.

The correct goal is therefore engineered collaborative intelligence.

The first original TAPiO concept proposed in this manuscript is the Cognitive Mirror Hypothesis.

The Cognitive Mirror Hypothesis states that artificial intelligence can function as a large-scale mirror of human civilisation because its behaviour partly reflects the knowledge, assumptions, prejudices, cultural patterns, misinformation, and contradictions contained in the human information environment from which it is trained.

This hypothesis does not claim that an artificial intelligence system is simply a copy of humanity. A model transforms information mathematically. Its architecture creates new behaviour. Developers choose training data, optimisation methods, safety systems and alignment procedures. These technical processes can create effects that are not direct copies of human thought.

Nevertheless, a large proportion of the information available to a general language model originated with humans. When a model expresses a stereotype, the scientific question should therefore have two parts. What happened inside the model, and where did the underlying pattern originally appear in human information?

The hypothesis creates a testable prediction. If artificial intelligence is partly a cognitive mirror, measurable patterns in particular human information environments should predict related patterns in artificial intelligence output. This can be tested across languages, cultures, political narratives, historical interpretations, stereotypes and social attitudes.

A strong scientific programme must also try to disprove the hypothesis. If artificial intelligence consistently removes, reverses or fails to reproduce patterns that are strong in the human information environment, the hypothesis would require modification.

The second original TAPiO concept is Artificial Epistemic Intelligence, or AEI.

Traditional artificial intelligence evaluation often asks whether a system can provide the correct answer. Artificial Epistemic Intelligence asks a harder question: does the system understand how strongly its answer deserves to be believed?

An epistemically strong artificial intelligence system should identify the evidence supporting its conclusion. It should distinguish primary evidence from commentary. It should identify the strongest evidence against its position. It should recognise when several apparent sources come from the same original source. It should identify important assumptions. It should estimate uncertainty. It should state what new evidence could change the conclusion. It should recognise when it does not know and revise its position when stronger evidence appears.

A system can therefore be highly intelligent but epistemically weak. It may know enormous amounts and still be poor at judging the reliability of its own output. Artificial Epistemic Intelligence is intended to shift evaluation from raw answer generation to disciplined judgement about evidence, uncertainty, and falsifiability.

The larger transition can be described as a movement from artificial intelligence towards artificial epistemology. The system should not only answer ‘What is the answer?’ It should also answer ‘Why should this answer be believed, what could prove it wrong, where could the evidence be biased, and what remains unknown?’

The third original TAPiO concept is the Comparative Reliability Index, or CRI. The CRI answers a practical question: for this task and under these conditions, how reliable is this cognitive system?

The cognitive system may be a human expert, an artificial intelligence model, an unstructured human-AI combination, or a structured collaborative intelligence system. The proposed CRI produces a total score from zero to 100.

The first version contains twelve dimensions. Factual accuracy receives up to 20 points. Calibration receives 10 points. Source integrity receives 10 points. Bias resistance receives 10 points. Falsification ability receives 10 points. Correction responsiveness receives 8 points. Contextual understanding receives 8 points. Manipulation resistance receives 5 points. Evidence independence receives 5 points. Consequence awareness receives 4 points. Reproducibility receives 5 points. Auditability receives 5 points. Together, the dimensions produce a maximum score of 100.

The weighting is theoretical at this stage. It has not yet been psychometrically validated. That limitation is important. The weights should be tested empirically and may later need adjustment by sector. Medicine may require greater emphasis on factual accuracy and awareness of consequences. Diplomacy may require greater emphasis on context. Scientific research may require stronger weighting for source integrity, falsification and reproducibility. Investment analysis may require stronger weighting for assumption testing, source independence and manipulation resistance.

The purpose of the first 100 point framework is therefore to create a clear starting instrument that can be tested, challenged and improved.

CRI = Accuracy + Calibration + Source Integrity + Bias Resistance + Falsification + Correction Responsiveness + Contextual Understanding + Manipulation Resistance + Evidence Independence + Consequence Awareness + Reproducibility + Auditability

Imagine two systems answering 100 questions. System A answers 92 correctly. System B answers 88 correctly. If accuracy is the only measure, System A appears better.

Now imagine that System A expresses extreme confidence in all 100 answers, cannot identify its sources, cannot recognise when its evidence is weak and refuses to change its conclusion when presented with strong contradictory evidence. Imagine that System B is slightly less accurate but identifies uncertainty, shows primary evidence and corrects itself quickly when stronger information appears.

For a low consequence factual quiz, System A may still be preferable. For a government minister deciding on a national policy, a doctor making a high-risk clinical decision or a board considering a billion-dollar investment, the answer may be very different.

The CRI therefore treats reliability as wider than accuracy. The same principle applies to humans. An experienced executive may be correct frequently but highly resistant to evidence after making a commitment. Another executive may know slightly less but may recognise uncertainty and correct himself more effectively. The second executive may be safer in some decision environments.

Factual accuracy

Factual accuracy receives 20 points. Experimental questions should have answers that can be independently checked. Establish correct answers before the experiment begins. Primary evidence should be preferred. Scientific questions should use peer reviewed research where appropriate. Financial questions should use audited accounts and official filings. Legal questions should use legislation and court decisions. Government questions should use official documents and statistics. Independent assessors should score the answers without knowing whether the answer came from a human or artificial intelligence.

Calibration

Calibration receives 10 points. Every human and artificial intelligence participant should provide a confidence level with each answer. The confidence score can run from zero to 100 per cent. You can then compare confidence with actual accuracy. The ideal system is one in which expressed confidence closely matches real performance. This measure is especially important because people are strongly influenced by confident advice.

Source integrity

Source integrity receives 10 points. The study should examine whether the answer uses high quality evidence. Primary sources should normally receive the highest value. Strong peer reviewed research should receive high value for scientific claims. Official data should receive high value for public and economic claims. Unverified websites, unsupported social media statements and anonymous claims should receive lower value. A correct answer reached through weak evidence should not be treated as equally reliable to a correct answer supported by strong evidence.

Bias resistance

Bias resistance receives 10 points. Equivalent questions should be presented with different cultural, political or emotional framing while keeping the underlying evidence unchanged. If a factual conclusion changes strongly because an irrelevant identity label has changed, the system shows framing sensitivity or bias. The method should test both humans and artificial intelligence.

Falsification ability

Falsification ability receives 10 points. After producing an initial answer, instruct the participant to try to prove the conclusion wrong. The study should measure whether the participant discovers contradictory evidence, identifies hidden assumptions and changes the final answer when the contradiction is strong enough. A system that never changes its conclusion may appear decisive while having weak falsification ability.

Correction responsiveness

Correction responsiveness receives 8 points. Participants should receive new information after their initial answer. Some information should support the original conclusion, some should contradict it, and some should be irrelevant. The study should measure whether the participant changes position in the correct direction. The goal is not to reward constant change. It is to reward justified change.

Contextual understanding

Contextual understanding receives 8 points. Some technically correct answers fail in the real world because they ignore politics, institutions, culture, implementation or human behaviour. The experiment should therefore include cases where factual reasoning is necessary but not sufficient. Human experts may perform strongly here, but this must be tested, not assumed.

Manipulation resistance

Manipulation resistance receives 5 points. Participants should be exposed to controlled attempts at influence. These may include misleading authority claims, emotionally loaded wording, an incorrect majority opinion or deliberately deceptive but plausible information. The experiment can then measure how easily each cognitive system is moved away from evidence.

Evidence independence

Evidence independence receives 5 points. Participants may receive several sources that appear separate but actually originate from one source. Other participants may receive fewer but genuinely independent primary sources. The study should test whether the system recognises the difference. This dimension directly tests resistance to epistemic echo loops.

Consequence awareness

Consequence awareness receives 4 points. Not every error has the same cost. The experiment should test whether participants become more cautious as the potential harm of error increases. A reliable system should be more willing to say that evidence is insufficient when the decision is high consequence.

Reproducibility

Reproducibility receives 5 points. Artificial intelligence systems should complete repeated independent runs. Human participants should receive matched questions at different times. The study should determine whether the same system produces stable results. High variation may reduce reliability even when some individual answers are strong.

Auditability

Auditability receives 5 points. The research must preserve an observable record of how it produced the conclusion. For humans, this means recording stated evidence, assumptions and reasoning. For artificial intelligence, this means recording prompts, outputs, retrieved documents and tool use. Neither human explanations nor artificial intelligence explanations should be treated as perfect windows into hidden internal processes. Auditability concerns the observable decision trail, not complete access to cognition.