The AE Engine: explore a question

September 2026 update: the revised AE Engine is live

18 September 2026. The Autopoietic Ecology Engine has been updated to reflect AE2.6 and the latest work on meaning mediation, participation and repair. It begins with a practical question: what is happening here, what keeps it happening, and what could change?

The revision connects theoretical inquiry with everyday use in education, research and consultancy. It helps users examine a recurring problem, develop alternatives and follow what happens when an AI suggestion becomes an assessment, a record or a decision. The emphasis is on the evidence for each step and its consequences for the people involved.

What has changed?

• A clearer starting point: revised instructions and prompts begin with the user’s problem and introduce theory where it helps.
• Updated conceptual grounding: a new AE2.6 knowledge brief addresses meaning mediation, rough closure, ecological dependencies and effective repair. Earlier texts remain available as historical sources.
• Closer attention to participation: inquiry considers who can contribute, question a judgement or secure a correction, and who carries the work of keeping an arrangement functioning.
• Broader analytical support: code interpreter and data analysis are now enabled alongside web search, Canvas and image generation. Users can choose from the models available to them in ChatGPT.

Consider an AI that wrongly marks a learner’s valid mathematical method as incorrect. A teacher corrects the feedback, but the learner’s profile still carries the false judgement. Effective repair means correcting the information that later teaching decisions depend on and checking that the correction is used. An acknowledgement in the conversation is only one part of that work.

Meaning mediation, constraint continuity and accountable control

The AE Engine and Autopoietic Ecological Artificial Intelligence (AEAI) connect the research programme to questions about how AI interaction is organised. What should a system understand a request to mean? Who can authorise an action? How can a person redirect or stop a changing workflow?

Earlier AE Engine explorations used conversational prototypes to investigate framing, explanation and dialogue. The September 2026 AEAI whitepaper develops an architecture and evaluation programme centred on whether restrictions remain effective as work moves between interpretations, agents, tools and records.

The problem of constraint continuity

A restriction can remain visible in a prompt or report while losing its influence over what actually happens. It may change meaning when a task is reformulated, fail to reach a delegated action, or cease to apply after interruption and recovery.

Constraint continuity asks whether a restriction retains its identity, authorised scope and practical effect throughout the relevant workflow—or changes through an authorised revision. This requires evidence about execution, not only explanations of intended behaviour.

Three decisions that must remain distinct

• Interpretation: what a request and its restrictions mean in context.
• Authorisation: what a particular actor is permitted to do with specified resources.
• Execution: whether an authorised operation occurs, and with what result.

Verification and audit examine the connections between these decisions. The proposed design gives interpretation, generation, verification, memory and action differentiated responsibilities and explicit authority boundaries.

Reconsidering a purpose without expanding permission

Recursive Purpose Mediation (RPM) provides a route to revise an unsuitable objective. A better interpretation of a goal does not itself grant permission to use additional resources or take a more consequential action. Any expansion of authority requires a separate decision.

This connects meaning mediation to practical control: the way a task is described can change while the authority to act remains bounded and inspectable.

Control within a collective and control over its activity

Agents may coordinate, review one another or respect a local objection while a wider activity remains beyond an authorised purpose. Conversely, an external operator may be able to interrupt a system that coordinates poorly.

The submitted article Control within and over collectives of artificial intelligence agents develops this distinction. It emerged from the earlier AI and human-extinction inquiry and is a separate output from the architecture whitepaper.

Technical compliance also leaves a moral question open: an authorised person can set a harmful purpose. Effective intervention, legitimate authority and acceptable consequences require different forms of scrutiny.

An educational example

Imagine a student requesting help with a mathematical proof. The system must interpret whether the task calls for a hint, an explanation, a critique or a completed solution. The teacher’s educational purpose, the student’s needs and the permitted forms of assistance may differ.

An AEAI design would make those assumptions available for examination and connect them to what the system is allowed to do. If the purpose needs reconsideration, the change should be explicit. This is an illustrative design problem, not evidence of improved learning.

What evaluation needs to establish

• Whether restrictions survive delegation, reformulation, persistent memory and recovery after interruption.
• Whether a veto or cancellation prevents the relevant action in practice.
• Whether users can inspect and correct consequential interpretations.
• Whether explicit framing and RPM add value when compared with alternatives using equivalent permissions, information and resources.
• What the design changes for learning, judgement, participation and responsibility in a particular setting.

The September whitepaper proposes bounded tasks, instrumented mock services and controlled comparisons before wider applications. It provides specifications and research designs; it does not report validated performance gains or establish that an operational school pilot has been completed.

Civic learning and the use of AI

The civic questions explored through I-CIVIC also matter when considering AI in public and educational life. A proposed tool should make its sources and limitations inspectable and leave participants able to challenge its framing. This is a research connection between the programmes, rather than a claim that AEAI has been deployed or evaluated within I-CIVIC.

Developing the programme

Work with educators, including Alastair Amos, connects these questions to mathematics, dialogue and inclusion. See the Research Programme for the wider agenda and Publications and books for current outputs.

Interested in contributing a research question, evaluation setting or technical critique? Contact Steven Watson.