The Research Programme: We Read the Work
How people lose command of AI without noticing, and what that does to their work and their capability. Real transcripts, not surveys. The known flaw is the ground truth. Nothing publishes below research-grade reliability.
Most AI research studies the machine. We study the interaction.
The programme asks one question: how do people lose command of AI without noticing — and what does that do to their work, their skills, and their judgment? The failures that matter leave no trace, which is why they need instrumentation rather than introspection.
Three findings define the field
- 79% of AI failures leave no visible trace (Stanford/Bigspin, 100,000 real conversations). Ordinary monitoring catches roughly 12%.
- N=1,372: people adopt AI output without engaging their own reasoning — and confidence rises even when the AI is wrong (Wharton).
- Drift: repeated use recalibrates the "something's off" instinct to match the machine. The profile cannot be self-diagnosed.
The method: we never ask, we read the work
Real transcripts, not surveys. The unit of study is the episode — a moment where a decision rode on AI output. The keystone of the method: you cannot plant errors in real work, so the catalogue of known machine flaws is the ground truth. Signals are read first and labelled second, which strips out rationalisation. Nothing publishes below Cohen's kappa >= 0.70 inter-annotator reliability.
What the programme produced
The Direction framework: the Four Moves, the delegation dial as context, cognitive surrender as the failure mode — the applied model of staying in command that everything at Ilanooo stands on. Plus an original contribution the literature did not have: a failure archetype we named silent tool failure — a tool errors, and the human moves on.
The curation cornerstone behind it: 21 studies across 30 years of the human–AI interaction record, read in one place.
The instruments
The research runs on instruments anyone can use. Paste one working conversation into DirectAI and get an instant reading: are you directing the machine, or surrendering to it? The full Direction profile follows — your band, move by move, with the delegation dial as context. For children, the same capacities grow through Creates — made visible as a portfolio instead of a score.
Findings ship when they clear the reliability bar, never for a moment's attention.