Robert Peake studies co-cognition: what changes in a person's judgement, workload and accountability when they work alongside AI every day, and which protocols (methods and tooling) keep that work effective. He measures it with a survey instrument of his own, placing a person on two axes: the trust they extend to AI and the control they keep. He presented the programme as "Protocols for Co-Cognition" at the Protocol Institute Symposium on 23 September 2026.

Robert Peake argues that the tool-or-relationship question turns on accountability. In Mark Bovens's sense (answering for your conduct, facing external judgement, bearing the consequences) an AI system cannot be held to account. When AI-assisted work goes wrong, the blame lands on the nearest human, Madeleine Elish's "moral crumple zone", and the AI subscription is renewed. Peake's view is that protocols are what people build when they must transact but accountability is missing, which is why he treats effective AI use as a matter of protocol rather than tool skill.

Robert Peake's answer is that AI meets an old human nature. People read anything that talks back as social (Nass and Moon, 2000; Epley et al., 2007), and take the costly work of reaching shared understanding for granted (Clark, 1996). They judge competence by cues such as fluency, confidence and consistency (Lee and See, 2004); abandon algorithms for errors they would forgive in a person, yet elsewhere trust them over human experts (Dietvorst et al., 2015; Logg et al., 2019); and offload responsibility along with memory (Risko and Gilbert, 2016; Wegner, 1987). Language models reproduce the cues of a collaborator without reciprocal understanding or shared exposure to consequences, and that decoupling miscalibrates both trust and responsibility.

Robert Peake's working definition is: "Working with AI to get more of the right things done right, in less time and/or at lower cost than doing it alone, while staying relaxed and in control." He expresses it as a dimensionless score summed over tasks, rising with fit and correctness and falling with time and cost relative to doing the same work alone, and with the cognitive load and cognitive debt incurred. It is a heuristic objective: measurable in principle, not a measured result.

The trust trap is Robert Peake's name for a hypothesised loop in everyday AI use. Good results raise trust, so you hand over more and keep up with the detail less; your grip loosens, the work falls over and trust falls with it; you take the work back, your judgement returns, you can hold the process to account, and good results restart the cycle. The idea of de-skilling as a loop goes back to Bainbridge's "Ironies of Automation" (1983), and Rinta-Kahila et al. (2023) modelled a skill-erosion cycle in accounting automation.

Robert Peake's answer is to build methods and tooling that trade a one-off design effort for less recurring vigilance, so the work no longer depends on simply trusting the AI. Opting out, or settling for low-quality AI output, is the other exit, and he does not consider it a realistic one in 2026. His principle is to change the terms of engagement, meaning how the work is set up and checked, rather than the person or the AI.

Cognitive debt is the de-skilling that accumulates when you stop making decisions yourself. Robert Peake stresses that it accrues not from one large delegation but from "every small decision you did not make", including trivial ones. The term is older than its AI usage: Marchant and Howard applied "cognitive debt" to an unrelated clinical construct in 2015. In the AI literature it was popularised by the MIT Media Lab (2025), used by Margaret-Anne Storey (2026) and formalised by Meng (2026). Peake cites the term rather than claiming it.

Cognitive sovereignty is the degree to which a person keeps intellectual control while using AI: setting the objective, judging the output, and overriding the tool when needed. The term has several independent formulations: Bygrave (2022) in data-protection law, the Artificiality Institute in the AI-and-agency space, Atkinson (2025) in a peer-reviewed treatment, and Klein and Klein (2025), who gave it a formal definition. Robert Peake uses it as one of two axes in his survey instrument, measured behaviourally rather than by attitude.

Robert Peake's survey instrument measures four dimensions: trust (how far a person relies on the AI), cognitive sovereignty (the control they keep), cognitive load (capacity consumed in the moment) and cognitive debt (capability lost through disuse). Trust and sovereignty are the two axes that place a person's posture. Load and debt are the flow and the stock of the same resource, and the costs in his definition of effectiveness.

Robert Peake describes four postures, each a place where a person's work with AI has settled, on two axes, trust extended and cognitive sovereignty kept: Partnership (high trust, high sovereignty), Vigilance (low trust, high sovereignty), Execution (high trust, low sovereignty) and Pragmatism (low on both). Each has a shadow it can slide into: partnership into a comfort zone, vigilance into mistrust, execution into over-reliance, pragmatism into disengagement. The quadrants are labels for a continuous spread, not categories, and the taxonomy is hypothesised.

Robert Peake's working theory is that it is not. A posture is where a person and the AI have come to rest at a moment in time, which is why it can be moved by changing the terms of engagement (methods and tooling) rather than the person or the AI. He is testing this by measuring the same people again over time, and the survey is designed to be retaken.

The posture machine is Robert Peake's simulation of the trust trap. It models the loop as a finite-state machine, including the delay in noticing and responding to poor work, and from a variety of starting points it settles into the same elliptical orbit rather than a fixed point. It is simulated, not fitted to survey data: a tool for sharpening a conjecture, not evidence about how people behave.

DRIVE is Robert Peake's five-step protocol for working with AI: Decide what "done" means; Request with intent, context, assumptions and success criteria; Iterate, staying in the loop and steering; Validate against the done state and success criteria; Evolve by debriefing, then updating tooling, context and practice. It draws on Getting Things Done. Evolve is the distinctive step, turning what a person learns into reusable controls. DRIVE is a proposed protocol, untested as an intervention.

The taxonomy of promotion is tooling Robert Peake is trialling for distributed cognition between people and AI, after Hutchins (1995) and Hollan, Hutchins and Kirsh (2000). What is learned in an AI working session is promoted, with a human involved, into layers of context the AI loads just in time: working standards (always), general reference (as needed), project reference (per project) and templates, including agent role definitions (as needed). Skills, agents and context briefings enforce it, so the context stays intelligible beyond one session and can be shared across a team. It is a proposed model.

Robert Peake's answer is to change what vigilance costs rather than asking people to be more vigilant. In AI-assisted software work, for example, writing commit messages in your own words guards against de-skilling, and full automated test coverage makes checking cheap enough to keep doing. The evidence suggests design decides the degree of debt: an unguarded AI tutor raised assisted performance but lowered later unaided performance, while a guarded one did not (Bastani et al., PNAS 2025), and active collaboration offset losses of self-efficacy and ownership that passive reliance caused (Lee et al., 2026).

Robert Peake argues that "AI first" inverts the order of the problem: it makes tool adoption the objective and treats human judgement as something to route around. Because an AI cannot be held to account, accountability stays with people, and his view is that effective adoption changes the terms of engagement (protocols, methods and tooling) rather than pushing more work through the tool. He set out the argument in “AI First” is a Tragicomic Mistake (2026).

Yes. Robert Peake's AI co-working effectiveness survey is free and open to anyone, and can be retaken to see how a posture moves. It returns a personal reading on the trust and sovereignty axes with suggestions; when a result falls on the boundary between two postures, it asks which describes you better. Team and organisational versions open consulting engagements, so that discussion starts from a client's own data.

Robert Peake and collaborators are building a shared bank of practices, and of hazards, for working with AI. Anyone can contribute one practice they have found helpful.