A figure walks towards the Dark Tower beneath an orange sun.
Back to the main site

How I use artificial intelligence.

This is the process today:

I define the objective, the success criteria and how to check them; the agent works out how; and every additional instruction has to earn its place with evidence that it is needed.

Seven lessons follow from this. Three concern judgement (what remains human, why it needs to be verified, and where the model fails). Two concern design (minimal instructions, making use of latent capabilities). Two concern practice (how to write instructions and how to scale up to many agents).

Personal

The tool helps. The judgement remains mine.

  1. Define the task

    Decide what needs to be resolved and with which information, without leaving the choice of objective to the tool.

  2. Review the output

    Check against identifiable sources, verify calculations and distinguish data, inferences and matters still needing verification.

  3. Explain the limits

    Record relevant decisions and AI interventions. If I cannot support a result, say so before presenting it as a conclusion.

Where it is applied

  1. Añelo report, final edition: each claim with its type, a 610-row registry, cross-review between two AI systems with separate roles and human sampling.

  2. New edition of the electricity report, in progress: a data manifest with a hash for each file, a claims registry, updates by edition and cross-verification.

The method is set out in the manual Traceable AI workflows.

I want technology to expand our capabilities and autonomy.

Comments or corrections? Get in touch.

Email · LinkedIn · WhatsApp