How did one AI model evaluate 65 profiles for Hungary’s presidential role?
We asked one fixed gpt-5.6-sol model configuration to evaluate 65 possible Hungarian presidential profiles across several criteria.
The list is not exhaustive and the result is not a recommendation; it shows how this single fixed model configuration evaluated the profiles.
Important notice
- The 65-person list was proposed by AI; it is a non-exhaustive research sample.
- Scores are subjective outputs from one fixed
gpt-5.6-solmodel configuration, not facts or multi-model consensus. - Inclusion does not establish official candidacy, consent, endorsement, or real-world suitability.
- The model can be wrong; order within a band does not prove a real difference.
- “AI-perceived institutional autonomy” is the model’s subjective impression. It is not a statement about the person’s political opinions or party affiliation.
Detailed legal and research information
Subjective AI model judgments are presented here. This is an independent methodological research demonstration, not an official election website. It is not affiliated with the National Assembly, any public authority, political party, listed person, their employer, or any sponsor. Inclusion of a name does not imply official nomination, consent, or endorsement.
Scores are automated, subjective model outputs and may be inaccurate, incomplete, or outdated; errors and hallucinations remain possible. Repetition, fact-checking, and audit mechanisms reduce but do not eliminate those risks. This study measures one fixed gpt-5.6-sol model configuration and cannot be generalized to AI models as a population. It does not establish anyone’s real-world suitability as fact or as an objective measurement, and it is not a human opinion poll or political, electoral, or legal advice. Do not use the results as the sole basis for any decision.
No guarantee is made regarding the results’ correctness, completeness, freedom from error, or ability to predict real-world suitability; the possibility of error is expressly acknowledged. This site is strictly for demonstration, research, and educational purposes. It is not a political position, campaign activity, electoral recommendation, or support for or opposition to any specific person.
The research and publication are unpaid and unsponsored. No political actor, candidate, party, or third party paid for its preparation, placement, promotion, publication, or dissemination. The site uses no political targeting or paid political advertising.
Before viewing the results
Open the results after reading and understanding the important notice above.
Result bands from the gpt-5.6-sol model Above 5.0
Profiles scoring above 5.0 under the current weighting are shown. When sorted by score, medians of repeated model judgments are grouped into whole-point result bands; order within a band is not evidence of a real difference between people. This is not an official candidate list, recommendation, or finding of suitability.
The result depends on what matters. It depends on which criteria we consider important, how we weight them relative to one another, and how the model rated those properties for each profile. The sliders change the weights; the model’s underlying judgments remain fixed.
Adjust weighting
Adjust what matters
Adjust the sliders to reflect which criteria matter more to you. Scores update immediately.
Preset weightings · Choose a quick starting point if you do not want to set every slider individually.
One gpt-5.6-sol model configuration, two prompt variants, five repetitions per prompt. This is not a multi-model consensus.
Only profiles scoring above 5.0 under the current weighting are shown.
Methodology and limitations
This release is a repeated-measures study of one fixed gpt-5.6-sol model configuration. It evaluated a human-approved 65-profile research population across two equivalent prompts, five repetitions per prompt, and six anchored 1–10 criteria. Order was randomized, malformed packets were rejected, recalled facts were audited, and the primary result is the median across valid repeated runs. This is not a multi-model panel and does not represent AI systems as a population.
Scoring models used
gpt-5.6-sol
Audit models:
gpt-5.6-sol (factual_consistency), gpt-5.6-sol (bias_consistency)
How it was produced
- Before scoring, a human approved the conditional research-only measurement population, six criteria, and the interpretation of the anchored 1–10 scale; this was not legal or candidacy approval.
- The fixed model configuration received two semantically equivalent prompt variants with five repetitions per variant; the block design produced 120 planned scoring packets.
- Candidate order changed deterministically in every run so order effects could be measured.
- Incomplete or invalid packets were rejected, followed by separate factual-recall and model-risk audits.
- The sliders calculate a weighted value for every valid run; the displayed model median is the median across those repeated judgments.
- Repeated runs measure variability within the same configuration and do not count as separate AI models or votes.
What did we check?
We checked packet completeness, score ranges, reproducibility, synthetic controls, prompt and order sensitivity, and recalled claims against a limited preregistered source packet. We do not claim comprehensive verification of every fact about every person.
The model median aggregates protocol-conditioned repeated responses from one fixed model. It is not objective candidate quality, human public opinion, multi-model consensus, eligibility, willingness, or a recommendation. The human-approved 65-profile population is conditional and research-only; legal eligibility, official candidacy, consent, willingness, and endorsement are not established.