AI is everywhere; from algorithms to predictive text to face recognition software & ‘customer support’. The sum total of digitised knowledge colonised largely without consent into a complex, cross-referencing index aiming to cover all possible variants.

Data is vital to search returns.  Having worked with keywords for image search since 1998, my work, along with many others has been ‘appropriated’ to advance AI. It feels like theft.  Images which allowed us to illustrate and enhance, evidence and reflect, in an AI context become ‘input data’. Images which used to ‘speak a thousand words’ have become the tool used to generate those words themselves.

Large Language Models (LLMs) rely on vast data sets. AI uses data in captions, IPTC fields and keywords, combined with pixel search (diffusion) to return results. Data is then matched to facts about that data. Neural networks overlay inputs and adapt outputs in epochs of exhaustive pattern variables to configure results. Accuracy is assessed with iterations of supervised ‘gradient descent’. By keeping metrics simple and attributable the proportion of true positives is increased. This process relies largely on the quality of data and the specialism of those training the model.

In a process called ‘overfitting’, AI recognises only that which it knows. Unless you are able to thoughtfully consider what you need AI to do, it is restricted in its potential. Anything unseen or unattributed – including all cultural and historic experience and perception which has not been digitised is missing. Where data is missing – approximations are returned. The pre-exisiting bias of datasets excludes large amounts of human experience and builds a future which discounts that of which it is unaware.

At present AI grasps neither the subtleties of language or what it is to be human. Concepts and emotions which occur in any context do not follow predictable rules of pattern recognition. Synonyms must be exactly and precisely interchangeable for words they represent. Differences may be subtle but they’re relevant. Skeletal is not skeleton. AI does not know this.

The deployment of AI to creative industries; writing, publishing, the arts and music, which humans enjoy as a process – demonstrate the interest and preoccupations of those whom it serves – definitely at the higher end of Maslow’s hierarchy.  The machines have all the fun and don’t even know they’re enjoying themselves. Life is about experiences and how we experience them. AI can’t live life for us.

Many start-ups building on the groundwork of frontier models controlled by very few and rich people are free for now to encourage take-up but in time, fees will be levied. With externalised costs an AI query uses 10 times more power for its calculations than normal search, lithium is required for batteries, and water for cooling in remote data centres. Google have increased their carbon emissions by 48% since 2019.

The purpose of my work as a keyword specialist has been to return accurate and relevant results. The purpose of AI in automating keywords is to reduce costs and overheads and process routine tasks. Beauty and accuracy are in the details.  AI can’t ask questions no-one has thought of, or challenge common conceptions, neither can it question its own results. A wrong caption or published date can corrupt an actual date. An unknown person may be assigned as a known person. Broad strokes miss detail. Vast results and omissions may degrade search experience and even reshape history. As language changes and search terms evolve, some degree of automation is welcome but the lasting value of AI will be to complement our work not to replace us from it.