For builders integrating LLMs into serious technical workflows, the biggest risk isn't that the AI is wrongβit's that the AI acts confidently from an unresolved wrong state. In a new independent publication, Alexanja Senke argues that we need to treat Human-AI research like critical infrastructure: it must fail closed. This means the system defaults to a safe, known state when errors occur, rather than hallucinating its way through gaps in logic.
The Core Principle: Failing Closed
The central rule for this architecture is simple but demanding: The AI may be wrong. It may not act from an unresolved wrong state. In traditional software engineering, fail-closed systems prevent unsafe operations when errors are detected. Senke applies this logic to intellectual inquiry. Instead of hiding abandoned hypotheses or broken proof paths, a fail-closed research process keeps them visible. Contradictions, numerical evidence that does not generalize, and conclusions that must be downgraded are treated as essential parts of the research record. If a system hides its failures, it cannot claim to govern its knowledge.
Building the Co-Intelligence Architecture
Senke details a compact Human-AI research architecture that moves beyond prompt engineering tricks. It relies on specific methodologies for state control and verification. Key components include reconstruction before agreement, hidden-assumption audits, and parallel-path verification. The framework uses explicit claim boundaries and evidence maturity reasoning to stress-test conclusions. For builders, this translates to a workflow where you don't just ask the AI for an answer; you ask it to reconstruct the logic and identify the weakest link before you accept the output. The separation between public, protected, and unresolved work is controlled and transparent.
Discipline Over Institutional Affiliation
This approach is targeted at independent researchers, engineers, developers, and founders who need rigor without the overhead of traditional institutional gatekeeping. The argument is practical: formal institutional affiliation is not a prerequisite for building a disciplined research process, but discipline itself is non-negotiable. The future of Human-AI research depends less on more powerful models and more on better boundaries, better state control, and better falsification. The methodology explicitly preserves the word UNKNOWN when evidence is insufficient, preventing fluency from being mistaken for truth.
Key Takeaways
- Implement fail-closed logic in AI workflows by preventing the model from acting on unresolved wrong states.
- Preserve failed hypotheses and contradictions in your research record; they are data, not defects.
- Use techniques like hidden-assumption audits and parallel-path verification to stress-test AI outputs before acceptance.
- Rigorous state management and falsification are more critical to reliable AI systems than raw model power.
The Bottom Line
If a system cannot be stopped, it is not governed. Senkeβs framework proves that treating AI errors as visible state variables, rather than hiding them, is the only way to build trustworthy co-intelligence.