AI and the Everyday World

By Neil Richardson (retired, NHS Kirklees Corporate Services 2003-2019)

In a recent one-hour lecture (1), respected computer scientist and author Professor Michael Wooldridge advised his audience that the gist of his talk was not to deride the discipline AI or seek a return to the years when computer algorithms were in vogue. Instead, assisted by helpful screen images, his presentation informed an appreciative audience that AI systems often produce strange and unpredictable results. In 2020 the Large Language Model ChatGPT3 received applause when it answered questions involving mathematics and reasoning. Remarkably, the phrase ‘taller than’ emerged from GPT3 even though not built into the LLM design. Exemplifying the system’s scope, it can find all the triples (a, b, p) of positive integers with prime p when a to the power p = b! + p.

This capacity to solve problems relies on very powerful computers (GPUs) having access to a staggeringly large volume of training data. Fifty years ago, a computer file might hold four hundred personnel records, each showing an employee’s name, date of birth, national insurance number, et al. In 2026, sourced principally from the digital web, the file deployed in GPT3 projects totalled forty-five terabytes, equivalent to five hundred billion words, or 45 million long novels – none of these three measures being easily imagined by mere mortals. From the vast data accessible to LLM, the system works like a phone’s autocorrect, assigning probabilities to parts of sentences to log the most likely combinations of symbols.

Contemporary AI: an incredible technical performance. However, Professor Wooldridge’s closing remarks reveal how mysterious and non-conscious LLMs are painfully inconsistent, frequently deliver errors, and are highly overconfident, despite serious limitations in their capacity to revise artificial beliefs from new information. This list reads like a comedy drama where some dated business computer sends an acceptance letter to the worst possible candidate. Elsewhere, decades before the design of ChatGPT3, another approach to real-world troubles had also achieved results.

At Lancaster University, a novel means to deploy models was developed during research within troubled corporations by Professor Peter Checkland and colleagues, based at the university’s Department of Systems Engineering. Persuading university staff to act as itinerant consultants while also moving away from advanced statistics to A3 flipcharts meant the department’s research was somewhat radical. As a flexible set of principles for organizational enquiry, Soft Systems Methodology (SSM) guides users in the art of freehand sketches: coherent models to hopefully stir discussion about clients’ hard-to-define problems. Each model was an arrangement of verb-phrases (called activities) put forward for consideration. Merely marks on paper, a sketch which failed to gel with local experience was quickly committed to the waste-paper basket.

To assist learning, the modelling language was English not mathematics, illustrated by the activity Travel to university conference. This abstract high-level notion could be assessed and performed by a solitary commuter via first-class train, private car, or by several other ways. In group studies, contributors can debate the ‘how’ of each activity. Seven or eight linked activities were usually sufficient to form a complete SSM model. In a tentative model to construct design D, three activities are Agree the build D based on available materials and labour  2; Obtain materials and labour  3; and Find out about materials and labour  1. Arrows from activity 1 to 2 and activity 2 to 3 indicate logical dependency and rational transfers of data. Assuming the sketch was received as meaningful to ongoing concerns, a contrast of model with how the organization is seen by its members asks pertinent questions:

       Is this activity performed in some way? If not, ought it to be?

       How well is it performed?

       Who is in charge?

       Does the activity have information needs?

       Any ideas for change?    

Peter Checkland’s Methodology does not attempt to create precise accounts of our complex and inconstant human world, replete with diverse concerns for numerous groups. While in the main AI models manipulate mathematical representations of the world, SSM models are only rational conjectures to assist systemic enquiry into ever-changing social scenarios – research which may take directions not imagined during the opening days. Occasionally, strong agreement over a model’s content among learners goes beyond modest change to help design new corporate systems, as that noun is usually understood in business settings.

1  The Royal Society’s Michael Faraday Prize Lecture,
    This is not the AI we were promised, February 2026