Was den Menschen zum Menschen macht

AI Self-Validation

5 Min. Lesezeit

🇩🇪 Deutsche Version: KI-Selbstvalidierung

AI self-validation is the attempt of an AI system to establish the validity of its own results by itself — without an examining authority independent of it. It is the literal enactment of the Münchhausen metaphor: the system is supposed to pull itself out of the swamp by its own hair. According to the Münchhausen trilemma, this attempt necessarily falls prey to one of the three horns — as a rule to the logical circle, for generator and examiner are the same system.

Three forms

  • Intrinsic self-correction — the model checks and improves its own answer without external feedback (“think again”).
  • Self-judging — a language model acts as a “judge” over outputs that stem from itself or from the same model family.
  • Training on its own outputs — a model learns iteratively from data that it (or its kind) has generated itself.

The threefold empirical refutation

All three forms have by now been tested independently of one another — with the same result:

  1. Self-correction fails. Huang et al. show (ICLR 2024) that the performance of large language models on reasoning tasks frequently deteriorates under self-correction without external feedback. Procedures such as “Self-Refine” (Madaan et al. 2023) work in the underlying studies only because they use an external signal — test cases, reference solutions, checking programs; the survey by Kamoi et al. (2024) reaches the same conclusion.
  2. Self-judging is biased. Panickssery et al. (2024) demonstrate that language-model evaluators recognize their own generations and systematically favour them (self-preference bias) — correlated with the capacity for self-recognition. If generator and examiner belong to the same model family, the “examination” is a circle; only judgment by a different system of justification provides a remedy.
  3. Training on oneself collapses. Shumailov et al. (Nature 2024) show model collapse: iterative training on self-generated data leads to irreversible defects — the tails of the original distribution disappear, the model loses information about the real world. This is the Münchhausen diagnosis at the population level: a system that learns only from itself drifts away from reality. Regularly feeding in genuine, external data prevents the collapse.

Where self-improvement genuinely works

The counter-cases confirm the diagnosis instead of refuting it: wherever AI systems demonstrably improve themselves, an external, formal verification signal exists. AlphaProof has every proof machine-verified by the proof checker Lean; AlphaZero stands on the external rulebook of the game with its unambiguous winning condition; the self-modifying “Darwin Gödel Machine” (Zhang et al. 2025) validates every change to itself against external test suites. Self-improvement succeeds exactly where the firm ground comes from outside (cf. verification authority).

Assessment in terms of personhood ontology

AI self-validation confirms the tool character of artificial intelligence from the epistemological side: a system with merely derived intentionality cannot establish the validity of its outputs, because validity is a relation to reality — and precisely this relation (adequacy) is enacted only in the personal judgment. This is not a devaluation of the systems but a determination of their place: they generate candidates (cf. AI-assisted knowledge generation); validity comes from outside.

Counter-argument

The title of the Huang study ends in “Yet” — not yet. Future systems with closer interaction with the world could extend the reach of formal checking procedures and shift the human share further. The reply: this is to be expected and does not change the diagnosis. Wherever a system improves “itself”, it does so against an external signal — and then it is no longer self-validation but externally anchored justification. What is revisable is the empirical basis; the conceptual diagnosis — self-justification is circular — is not. Three limits remain structural in addition: the positing of the criteria of success, the grounding of signs in a practice, and the attribution of responsibility, which falls on persons alone.

Ontological classification: Special case of self-justification; distinct from externally anchored justification; necessarily falls prey to one horn of the Münchhausen trilemma.

Sources: Generated by querying the Personhood ontology. Research as of 14 July 2026 (research report The Münchhausen Problem in LLM-Assisted Ontology Engineering).

Further sources:

  • Huang, Jie; Chen, Xinyun; Mishra, Swaroop; Zheng, Huaixiu Steven; Yu, Adams Wei; Song, Xinying; Zhou, Denny (2024): Large Language Models Cannot Self-Correct Reasoning Yet. ICLR 2024.
  • Kamoi, Ryo; Zhang, Yusen; Zhang, Nan; Han, Jiawei; Zhang, Rui (2024): When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs. Transactions of the Association for Computational Linguistics 12, pp. 1417–1440.
  • Panickssery, Arjun; Bowman, Samuel R.; Feng, Shi (2024): LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024.
  • Shumailov, Ilia; Shumaylov, Zakhar; Zhao, Yiren; Papernot, Nicolas; Anderson, Ross; Gal, Yarin (2024): AI models collapse when trained on recursively generated data. Nature 631, pp. 755–759.
  • Madaan, Aman et al. (2023): Self-Refine: Iterative Refinement with Self-Feedback. NeurIPS 2023.
  • DeepMind (2025): Olympiad-level formal mathematical reasoning with reinforcement learning (AlphaProof). Nature, 12 November 2025.
  • Zhang, Jenny; Hu, Shengran; Lu, Cong; Lange, Robert; Clune, Jeff (2025): Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. Preprint.
  • Albert, Hans (1968): Traktat über kritische Vernunft. Tübingen: Mohr Siebeck.

See also

Generated by querying the Personhood ontology.

Graphansicht

Backlinks