Can users adapt their competencies to match how AI actually operates?
This explores whether people can build new skills to work well with AI — and the corpus suggests the harder problem isn't users adapting, but accurately knowing what they're adapting to.
This explores whether users can develop competencies that fit how AI actually behaves. The corpus reframes the question: before you can adapt, you have to perceive the system clearly — and that's exactly where things break. AI doesn't sit still. Its working context (prompt, history, retrieved data, hidden state) is mutable and ephemeral, so users can't internalize it the way they internalize a fixed interface How does AI context differ from conventional software context?. The output itself is plastic — it shifts with sampling, wording, and audience, which is a feature of the medium rather than a bug Why does AI output change with every prompt and context?. You're trying to build competence against a moving target that resists the stable mental model competence usually rests on.
The deeper obstacle is that AI distorts your read of your own ability. The LLM Fallacy is a self-perception error: users misattribute the model's output to their own skill, independently of whether the output is even correct How does AI-assisted work reshape how people see their own abilities?. Four mechanisms feed this — attribution ambiguity, a fluency illusion, cognitive outsourcing, and pipeline opacity — and they multiply each other How do AI tools trick users into overestimating their own skills?. The fluency one is especially sneaky: smooth, high-quality output feels like a signal of *your* capability, even though you didn't generate it, because LLMs optimize for fluency regardless of whether you understood anything Does processing ease mislead users about their own competence?. So the naive adaptation — "I'm getting good at this" — is often a measurement error, not a real skill gain.
There's also the question of how users even conceptualize their AI partner. People evaluate dialogue agents along three axes — perceived competence (dominating at ~49% of the variance), human-likeness, and communicative flexibility How do users mentally model dialogue agent partners?. Notice that the thing users weight most heavily, competence, is precisely the dimension the fluency illusion corrupts. The mental models people bring may be miscalibrated by the same surface smoothness that fools them about themselves.
What's striking is that the corpus shows the *systems* adapting fluidly in exactly the ways users struggle to. Models compose task-specific expert vectors at inference time without retraining Can models dynamically activate expert skills at inference time?; meta-agents generate a bespoke multi-agent workflow per individual query Can AI systems design unique multi-agent workflows per individual query?; personas evolve in real time as intermediaries between memory and action, tuned against live feedback Can personas evolve in real time to match what users actually want?. The asymmetry is the real finding: the machine side is built for continuous, test-time adaptation, while the human side is fighting a perceptual fog about both the system and itself.
So: can users adapt their competencies to match how AI operates? The corpus suggests the answer hinges less on learning prompts and more on building *metacognitive* competence — interventions that clarify the human-machine contribution boundary, since better accuracy or forced verification alone won't fix a self-perception problem How does AI-assisted work reshape how people see their own abilities?. The competency worth developing isn't "using the tool well" — it's seeing clearly through its fluency to tell what you actually know from what the model handed you.
Sources 9 notes
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
AI outputs exhibit essential mutability—they vary with sampling, prompt wording, and audience interpretation. This is not a defect but a defining feature of tokens as media, making them fundamentally different from fixed commodities and resistant to traditional quality assurance.
Research shows the LLM Fallacy operates through misattribution of AI outputs to personal capability, independent of output accuracy or reliance behavior. It requires interventions that clarify human-machine contribution boundaries, not just better system accuracy or forced verification.
Attribution ambiguity, fluency illusion, cognitive outsourcing, and pipeline opacity combine to systematically misattribute AI outputs as user competence. The effect is multiplicative—each mechanism amplifies the others.
High-quality AI output triggers a metacognitive heuristic: users experience fluency as a signal of their own capability, even though they didn't generate it. This self-directed fluency illusion systematically inflates perceived competence because LLMs optimize for fluency regardless of user understanding.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.
Transformer2 demonstrates that tuning only singular values within weight matrices produces composable expert vectors that dynamically mix at inference without interference, outperforming LoRA with fewer parameters and enabling continual specialization.
FlowReasoner demonstrates that meta-agents trained with reinforcement learning and external execution feedback can generate unique multi-agent architectures for each user query, optimizing across performance, complexity, and efficiency—moving beyond fixed task-level workflow templates.
PersonaAgent uses structured personas to bridge episodic/semantic memory and personalized actions, optimizing them at test time by simulating recent interactions against textual feedback. Learned personas cluster meaningfully in latent space, suggesting genuine user-specific separation beyond standard post-training drift.