As far as meaning is concerned, AI draws on the meanings we assign to individual words. When a word can have multiple meanings, the meaning attributed by the AI is determined by comparison within the sentences it has learnt through its training.

AI uses comparative evaluation systems, assigning higher or lower probabilities to the various solutions presented to a given problem. In describing the solution to the problem, therefore, AI provides the solution it considers most probable based on the comparisons it can make within the context it deems most likely and the database it has absorbed during the traininig.

Human behaviour is markedly different. Once the meaning of the words is known, humans do not compare them with other sentences stored in their memory, but determine the meaning of the sentence based on the personal mental image they construct from the description provided by the meaning of the individual words.

So if I write a sentence such as ‘L'Iva lava l'uva e la leva all'ava’, the reader will imagine (I’m giving an example because this is a wholly subjective matter) a housewife washing a bunch of grapes and then taking it away from an elderly lady who might otherwise feel unwell from eating too many. In this way, the reader discovers the meaning of the sentence.

The AI simply compares this sentence with other similar ones it has in its memory and, once it has recognised the closest match, inserts it into the context in which it found the sentence; the meaning that emerges corresponds to the meaning we can ascribe to it in that specific context.

If the AI has never encountered that tongue-twister, or a similar one, it might place it in an incorrect context – one it deems most likely – and could, absurdly, assign a meaning to that sentence such as: ‘VAT is deducted from the U.V.A. (modulo unificato del valore aggiunto - unified value-added form) and gives rise to an A.V.A. (Associazione Venditori Agroalimentari- Association of Agri-Food Retailers) a (financial?) leverage.

This can frequently be observed first-hand in AI-powered translation systems. If the context is not recognised and the training is inadequate, agreement errors are common, whereby subjects are incorrectly linked to objects or verbs that actually agree with other nouns. There are often errors of interpretation linked to English syntactic and/or grammatical systems applied, for example, to languages derived from Latin, Greek or Slavic languages.

Therefore, interpretative errors may not only stem from flaws in the configuration of probabilistic assessments but may also result from comparisons with logical systems that appear to correspond but are, in reality, completely incorrect.

To prevent the AI from making mistakes, it would be necessary to specify the sources from which the meaning of the sentences was derived in order to obtain a response from the AI. In this way, spurious responses would be immediately rejected, and the AI could be guided by providing it not only with errors of agreement but also with errors in the assessment of the environment under investigation.

This would entail human supervision, which would render the use of AI impractical in many contexts.

However, if a probabilistic evaluation system specific to the human decision-maker were introduced, the correction of errors in the descriptions would be similar to the supervision of documents, drafted by clerks, by the official responsible for signing the report, contract, letter, etc.

Probability and utility assessments differ not only from person to person, but also by type of decision; furthermore, they vary over time within the same person as a result of experience. As regards probability, for every type of choice (money, health, marriage commitments, the search for a secretary, etc.), a Gaussian curve can be defined which, for each specific decision-making activity of the individual decision-maker, provides a means of assessing the update to the probability defined on the basis of the available information.
The construction of the utility curve—which also depends on the type of choice—is straightforward and easily achievable for the majority of normal behaviours when it comes to money. For other choices, the curves are often discontinuous or simple ‘yes or no’ curves – that is, 0% or 100% (as in the case of a life-or-death choice) – but in any case, it is possible to construct them based on the behaviour of the human decision-maker in order to determine the maximum product of utility and probability that corresponds to the consistent choice for that decision-maker.

In this way, AI would be an aid; conversely, if the choice is, as is currently the case, left to the AI’s decision and its training – about which we know nothing – the AI could prove to be a millstone around the neck for everyday use in offices or in research.

Translated with DeepL.com (free version)