Research · Evidence · AI Evaluation

What does the source actually support?

I tend to begin with a question that sounds simple: what does the source actually support? Much of my work has grown from noticing the distance between a confident claim and the evidence underneath it.

I research across AI, education policy, developmental psychology, neuroscience, human rights and misinformation. I use AI to widen the search space, surface counterarguments and accelerate comparison; consequential claims go back to primary or authoritative sources, and uncertainty stays visible when the evidence does not justify certainty.

Independent researcher & writer Native Czech · English CET / CEST · Remote

Selected research evidence

Four questions I have actually worked on — and the public work behind them.

When an authority cites an authority, where is the evidence?

Pillars of the Czech Curriculum: Evidence, Ideology, or Trust in Authority?

OECD · UNESCO · CASEL · DigComp · policy evidence

I traced frameworks cited as the evidence base for a national curriculum and separated institutional authority, normative goals and empirical support. The analysis asks which propositions are empirical, which are normative, what methodologies sit underneath them, and how far the evidence can reasonably be generalized.

seberizenevzdelavani.cz/pilire-rvp-dukazy-nebo-ideologie-a-vira-v-autority/

When the law permits something, does the system actually allow it?

Self-Directed Education in the Czech Republic: Is It Possible?

law · education policy · definitions · practical constraints

I compared public claims about educational freedom with legislation, curriculum rules and the practical conditions created by the wider regulatory system. The central distinction is between formal legal possibility and what the surrounding system makes possible in practice.

seberizenevzdelavani.cz/seberizene-vzdelavani-v-cr/

What does “control” mean when model behavior is learned, emergent and partly opaque?

Autonomous Learning and Emergent Behavior in AI: The Illusion of Control Is the Present, Not the Future

AI · emergence · capability evaluation · uncertainty

I examined the gap between controlling architecture and training conditions and being able to predict, explain or reliably evaluate model behavior after training. The question is not whether humans influence these systems, but what kind of control that influence actually justifies us claiming.

michaelarerichova.cz/en/lack-of-control-over-ai/

What should make us distrust a persuasive piece of content?

How to Survive the Deepfake Jungle: 10 Questions to Free Yourself from AI Manipulators

misinformation · provenance · corroboration · cognitive bias

I built a practical verification framework around provenance, missing context, independent corroboration, incentives, emotional manipulation, cognitive bias and the possibility that apparently persuasive material may itself be AI-generated.

michaelarerichova.cz/en/deepfakes-from-ai/

How this transfers to model-output evaluation

The underlying task is familiar: isolate the claim, determine what evidence would resolve it, find the strongest available source, compare the wording against that source, identify unsupported scope, missing qualifiers or contradictions, and document a verdict without overstating certainty.

I would apply the same method to AI-generated answers: factuality evaluation, source verification, claim decomposition, edge-case analysis, adversarial checking and rubric-guided review.

This is a transfer of my research method to model evaluation — not a claim of prior production annotation, inter-rater calibration or professional red-team experience.

Micro fact-check

A plausible claim that fails on scope

Claim: “The EU AI Act became fully applicable on 2 August 2026.”

Misleading

Why

2 August 2026 is the Act’s general application date, but “fully” overstates what happened on that date.

Even under the original Act, Article 6(1) and the corresponding obligations were set to apply only from 2 August 2027, while Article 111 contained transitional periods extending as far as 2030.

Regulation (EU) 2026/1744 subsequently amended the timeline again. Chapter III, Sections 1–3 — with the exception of Article 6(5) — apply from 2 December 2027 for systems classified as high-risk pursuant to Article 6(2) and Annex III, and from 2 August 2028 for systems classified as high-risk pursuant to Article 6(1) and Annex I.

The timeline also continues forward

New Article 5 prohibitions apply from 2 December 2026 (Art. 1(40)(a)).

Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content that were placed on the market before 2 August 2026 must take the necessary steps to comply with Article 50(2) by 2 December 2026 (Art. 1(39)(b), inserting Article 111(4)).

What failed?

Nothing in the claim looks absurd. That is precisely why it deserves checking. The date is broadly meaningful; the failure is the word “fully”. It compresses a staggered legal timeline into a single event. The error is one of scope, not fabrication.

Primary legal source

Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026, Arts 1(39)(b) and 1(40), OJ L, 2026/1744, 24.7.2026.

data.europa.eu/eli/reg/2026/1744/oj

Original Regulation (EU) 2024/1689: data.europa.eu/eli/reg/2024/1689/oj

Checked against the operative legal text on 7 September 2026.

I am not interested in replacing uncertainty with confidence. I am interested in finding the boundary between what we know, what we infer, and what the source does not allow us to say.