From Human Judgement to AI Calculation

Matthias Müller-Prove 4/2023 :: v1

v1.1 at LinkedIn

Once Joseph Weizenbaum asked the following question [Joe Weizenbaum at MEDICHI in Klagenfurt] : Take a cow with three legs. Provide a prothetic leg to the cow. How many legs does the cow have now? Three or four? – The answer is still three. Three legs and one prothesis. The artificial leg might take over the function of a leg to provide support and balance during walking. But it remains an external element to the cow’s body. It is a category error to count the artificial leg as a real one because it does not have all the properties of a leg, for instance no fur, no touch senses, no scars from years before. Calling the prothesis a leg is a metaphorical term to indicate the intended purpose of the tool or machine.

It is the same as calling certain concepts of graphical user interface “desktop”. Desktop computing was introduced as a metaphor to ease the interaction between people and machines in times when the prevailing paradigm was CLIs – command line interfaces with green text on black screens.
Let’s hold on for a moment and celebrate the 40th anniversary of Apple Lisa. A year later, 1984, Apple has introduced the Macintosh. Two years before was the launch of Xerox 8010 Information System, the Xerox Star; all of them offered Desktop-GUI systems.

Back to the future of AI and generative text production services provided by systems like chatGPT and others. Like the fourth leg or like files and folders and the trash can… it is a category error to treat the output of AI text generation as if it would have been written down by somebody. Forever there was an invisible contract between author and reader of texts, i.e. the author’s intend was to capture a certain idea and to convey a certain meaning to a future reader. If any reader reads any text, her assumption was that the author was writing the text for a reason. This is true for any kind of texts, regardless whether it be science, science fiction or fiction, whether it be news of poems, whether it be valid reasoning or fantasy.
This invisible assumption no longer holds true for texts which are produced by artificial neural networks. Such texts are not written, they are not told. In the context of AI, the verb ‘to write’ is a metaphor for something that looks similar to the human activity of writing a text. AI systems do not tell stories. They “tell stories.” They virtually tell stories. Neither do they tell the truth nor do they lie. They do not even virtually lie because they lack the (cap)ability to handle truth in a moral and responsible sense.

We need to be very clear not to mix up the concepts of human writing and the output of AI systems. The latter generates intriguing letters, but they are text without a single author. There is no authority to stand up and defend the thoughts that are captured in the text. AI texts are generated based on millions of millions of training texts and clever matrix algorithms to effectively and efficiently generate sound and likely strings, that resemble human texts.

If we mix such strings within scientific text bodies, we cannot be confident anymore that the reader, who is asynchronously reading the lines, engages with an author and her ideas. Any attempt to understand the author might fail because there is no an author anymore. The remote dialog cannot be established because the authorship is unclear and the meaning might just be projected onto the text by the reader. If a channel between author and reader cannot be established with certainty then a scientific discourse is severely hampered.

AI texts have not author. Furthermore they represent something else than crowd intelligence or the wisdom of the crowd in which one in many might have the idea that was needed to solve a problem and other support and promote the suggested solution.
However AI texts might represent something. They are much better than using a dice to concatenate characters to each other. But we don’t really know how to treat them. A first step would be to tag them as AI-generated and to avoid metaphors that lead to the assumption that they are the same as human texts.

Joseph Weizenbaum’s book »Computer Power and Human Reason« has the subtitle »From Judgement to Calculation.« Later he considered that the subtitle would have been the better title: »From Judgement to Calculation« This was a warning.

Feedback – please leave a message

I'd be happy to receive and read your feedback via_

à propos