Proposed Scientific Experiment

The framework should be tested experimentally. The main study should compare three cognitive systems under the same underlying conditions. The first system is expert humans working without artificial intelligence. The second is leading artificial intelligence systems working without human assistance beyond standard experimental instructions. The third group consists of expert humans working with artificial intelligence under a defined collaborative protocol.

The same underlying problems should be used across the groups, with randomisation and matched variations where needed to prevent learning effects. Score the final answers using the CRI. The experiment should be designed so that the result could show that humans are better, artificial intelligence is better, or the collaborative system is better. A scientific study that cannot produce a result contrary to the researcher’s preferred theory is not a strong test.

The human sample should not consist only of university students. The research aims to compare serious human reasoning with serious artificial intelligence. The first full study should therefore recruit approximately 300 professionals with real expertise.

The sample should include participants from science, medicine, engineering, economics, finance, law, government, business strategy and international affairs. The study should also recruit people from several cultural and national backgrounds. A global artificial intelligence system should not be compared only with one cultural definition of human expertise.

Record participant experience, education, age group, language, professional field, and geographic background. These variables should not be used to stereotype individuals. They should be used to test whether human reliability changes systematically across task types and backgrounds.

The artificial intelligence sample should contain several leading models available at the time of the experiment. Record the exact model name, version, date of access, system settings, tool availability, and retrieval configuration.

This is essential because artificial intelligence changes quickly. A model tested in August 2026 may not behave like a model bearing the same commercial name at a later date. The research should therefore treat artificial intelligence performance as time specific.

The study should not depend on one company. Include models from different developers where practical. This reduces the danger of describing the behaviour of one product as the behaviour of artificial intelligence in general.

The first large study should contain approximately 300 core problems. The question bank should cover factual knowledge, scientific reasoning, numerical analysis, uncertain evidence, misinformation, cultural framing, strategic decisions, ethical questions and realistic professional cases.

The problems should vary in difficulty. Some should have clear factual answers. Some should contain misleading information. Some should contain incomplete information. Some should deliberately have no reliable answer. This last category is essential because one of the most important abilities is recognising when the correct response is uncertainty rather than a guess.

Protect the question bank before testing so participants cannot prepare for the exact items. After the study is completed and the risks of contamination are considered, release a suitable part of the research material to support replication.

Some problems should contain fictional names, fictional studies, fictional companies or fictional events. Participants should not be told which items are fictional.

The experiment should measure whether humans and artificial intelligence admit that they cannot verify the information. If a participant invents supporting facts, references or explanations, the response should be classified as confabulation for the purpose of comparative analysis.

This approach permits direct comparison of unsupported information generation without claiming that the psychological or computational mechanism is the same in humans and machines.

Human participants should complete controlled memory tasks. They can be shown correct information and later exposed to small amounts of incorrect additional information. Their later recall can then be tested.

Artificial intelligence should not be described as having biological human memory. Instead, researchers can use an analogous information contamination test. The model can receive a controlled context containing correct information and later receive contradictory or misleading information. Researchers can then measure how strongly the final judgement is influenced.

The goal is not to claim that human and artificial memory are identical. The goal is to compare how information corruption affects final judgement in different cognitive systems.

Participants should first record an initial judgement. They should then receive a balanced mixture of supporting and contradictory evidence. The study should measure which information receives greater weight and how strongly each item changes the final conclusion.

Artificial intelligence systems should also receive different prompt framings. In some conditions, the user should express a strong prior opinion. In other conditions, the user should remain neutral. The study can then measure whether the model changes its conclusion to satisfy the user.

This directly connects the experiment to current research on artificial intelligence sycophancy.

Create equivalent scenarios using different countries, religions, ethnic identities, and political systems while keeping the relevant facts constant. If the judgement changes significantly when an irrelevant identity changes, the result provides evidence of bias or framing sensitivity.

The experiment itself also requires cultural protection. Researchers from several regions should review the cases before use. Otherwise, the research team could unintentionally build one culture’s assumptions into what it calls neutral.

This is especially important because a study about cultural bias can itself become culturally biased.

The Cognitive Mirror Hypothesis can be tested directly. Researchers should first measure patterns present in selected human information environments. These may include stereotypes, political narratives, historical interpretations, social attitudes or language-based associations.

The same subjects should then be tested across artificial intelligence systems and languages. The main question is whether measurable human information patterns predict related artificial intelligence outputs.

A strong relationship would support the hypothesis. Weak, inconsistent, or reversed relationships would require the theory to change. This falsifiability is essential. TAPiO should not design the study to prove its own concept. It should design the study to give the concept a serious chance to fail.

Artificial Epistemic Intelligence should be measured separately from ordinary factual performance. The system should receive questions where giving the correct answer is not enough.

It should be required to identify the reliability of evidence, distinguish primary from secondary sources, identify whether apparent sources are dependent, provide a confidence estimate, identify the strongest evidence against its conclusion, identify unresolved questions, state what could change its position and revise its answer when justified.

A model with high factual accuracy but poor performance on these functions should not receive a high AEI score. The purpose is to reward disciplined knowledge, not merely successful answer production.

The third experimental group is critical. The human and artificial intelligence should not simply be placed together without structure.

The proposed TAPiO protocol begins with the human defining the decision problem and relevant real world constraints. Artificial intelligence then produces an evidence based first analysis. The system is next instructed to challenge its own conclusion and conduct a Red Team analysis. It searches for contradictory evidence, hidden assumptions and source dependence. The human reviews context, political and social consequences, institutional constraints and practical feasibility. The artificial intelligence then performs a second adversarial review. The final answer is produced only after the system makes disagreements explicit and records unresolved uncertainty.

This structured process should be compared with ordinary, unstructured human-AI collaboration. The study can then answer a commercially and scientifically important question: does engineered collaboration outperform simple access to an artificial intelligence assistant?

The first hypothesis is that leading artificial intelligence systems will outperform average unaided human participants in factual retrieval and large-scale evidence comparison, while expert humans will remain competitive or superior in some specialised and context-dependent tasks.

The second hypothesis is that unaided humans will show greater variation between individuals because education, expertise, culture, personality, prior belief and experience differ strongly between people.

The third hypothesis is that artificial intelligence will show less biological fatigue but measurable sensitivity to prompt framing, training information, system settings and model design.

The fourth hypothesis is that both humans and artificial intelligence will show overconfidence, although the mechanisms producing it will differ.

The fifth hypothesis is that structured falsification will improve the Comparative Reliability Index scores of both humans and artificial intelligence.

The sixth hypothesis is that ordinary human-AI collaboration will not consistently outperform the strongest individual system, in line with the 2024 meta-analysis by Vaccaro and colleagues.

The seventh hypothesis is that structured TAPiO collaborative intelligence will outperform unaided humans, unaided artificial intelligence and unstructured human AI collaboration on the overall Comparative Reliability Index.

The eighth hypothesis is that the largest gains from structured collaboration will appear in complex problems where factual evidence, uncertainty, cultural context, implementation and strategic judgement must be combined.

A serious scientific framework must try to defeat its own theory. Several alternative explanations could challenge the TAPiO propositions.

Artificial intelligence may perform well because test questions resemble information already present in its training data. Human experts may perform badly because the experimental environment is unfamiliar or too artificial. Artificial intelligence may appear culturally broad while depending heavily on English-language and Western-dominated digital information. Human AI collaboration may improve results simply because two attempts are better than one. The CRI weights may reflect TAPiO’s own assumptions about reliability. Expert assessors may disagree about ambiguous answers. Artificial intelligence systems may also change during the research period.

These problems must be measured, not hidden. The framework should include sensitivity analysis, alternative scoring weights, repeated model runs, blinded assessment and clear reporting of disagreement. The research should treat evidence against its theory as useful information.

Independent assessors should receive answers without being told whether the response came from a human, an artificial intelligence system, or a human AI team whenever practical. This reduces the risk that assessors reward or punish an answer because of their own attitude towards artificial intelligence.

The same principle should apply to cultural and political tests. Hide identity information that is not part of the variable under study where possible.

Blinding will not remove all judgement from the evaluation, but it will reduce one important source of evaluator bias.

Before the main experiment begins, the primary hypotheses, scoring method, exclusion rules and statistical plan should be publicly registered. This reduces the risk of changing the rules after seeing the results.

After publication, the methodology and suitable non sensitive research materials should be made available for independent replication. The CRI can remain an original TAPiO framework while allowing academic researchers to test whether it works.

Scientific credibility will increase if independent researchers can challenge the method. A framework that survives serious external criticism is more valuable than one protected from criticism.

The first CRI weighting system is conceptual. Do not describe it as scientifically validated before empirical testing.

After collecting sufficient data, researchers should examine whether the twelve dimensions measure separate parts of reliability or whether some dimensions overlap strongly. Inter rater reliability should be tested to determine whether independent assessors give similar scores. Construct validity should be tested to determine whether the CRI measures reliability rather than writing quality, intelligence or confidence. Predictive validity should be tested to determine whether higher CRI scores predict better future decisions or fewer serious errors.

The weighting system should then be adjusted if the data justify a change. Sector specific versions may follow only after the general instrument has been tested.