We don't need to explain here the fundamental limitations of computerised processes: you can read them in the "Imaginary Friends" series. We need to remind you of only three things:

1. Computers are limited by their technology which, no matter how "AI" evangelists dress it up can only, ever, be binary: yes/no, on/off, etc. Fuzzy logic is a yes/no state that has not made a decision yet.
2. Computers cannot think and cannot understand context. This is a consequence of their binary nature. The appearance of thinking or rationalising is fake, created by whoever wrote the software.
3. Computers can store data but they cannot "learn".

We have demonstrated failure about almost two years ago (here and here) and we are told that tech has improved exponentially since then but while that may be true in relation to e.g. the clarity of images, it is demonstrably not true, or not true enough, in terms of the process leading to the results we expect. 

This was the challenge we set this week:

How unreliable is "artificial intelligence" and "generative AI" and do machines learn, identify and use data in a reliable fashion?

The answer is no, but the message is not getting through. See if this helps:

The instructions (it's artificial to call it a "prompt") involved that the image had to include a BT Trimphone in two tone green. 

The phone in the image isn't a Trimphone and it's not two tone green. There's nothing ground-breaking about such errors but...

This is Google's Gemini. Does Google not know what a Trimphone is and is it aware of items on two colours? 

The answer is that Google does. A Google search for BT Trimphone in two tone green produces multiple examples. 

It is clear that the data used to create the image is not the same as the data available. Therefore the research is selective and incomplete. 

There is something else: the girl in the image has been generated from the same description that we have used many times before but in recent weeks, Gemini (and other similar software) has changed her in subtle but meaningful ways. 

So asking for something today and again and again will, over time, produce different results If a user is looking for something authoritative, this is problematic. If the user is looking for an opinion, then it is clear that opinion will change over time. If the user is looking for an opinion, then it is clear that opinion will change over time, not because the opinion is being revised but because the computer is selecting different data upon which to generate an opinion.

This is in addition to the fact that the software regenerates from scratch each time. 

Are we really satisfied that for important stuff, not for our marketing images, good enough is acceptable?

What this teaches us

This reinforces the fact that computers cannot think and cannot rationalise nor recognise and react to context. 

But most importantly it shows us that the old adage that results out are only as good as the data in remains true but amplified by the fact that the data in is only as good as the sources the software searches and the algorithms that the data is subjected to. 

Incomplete and inaccurate data is an ever-present risk in all forms of research: that's why both academia and science - along with good management practice - says that results must be tested.

As business culture has moved towards "agile" and "fail fast" the culture of testing results before deploying them has been eroded.

This feeds directly into the results from computer-generated images and text: we absolutely cannot make the assumption that what the computer produces is comprehensive and correct.

And that means that we cannot rely on it. 

So fine, use it where the results are inconsequential or as a research assistant to produce something to work upon but do not, ever, regard its results as conclusive, especially where someone's livelihood, freedom or even life depends on it. 

It's all about thinking and machines can't do that.

Turing said, in terms, that for machines to be able to follow and interpret instructions, first they must understand English.

He was English. He spent a lot of time in the USA. So there's the first problem. But Turing had an advantage over us today: he learned English in a far more structured way: "verbs are ing words. Nouns do not end in ing. If there's no ing, it's not a verb." That lesson, with others has been lost leading to imprecision in language. So, aside from diversity, we also have the loss of structure, upon which machines rely for precision because computers do precision well, they are binary. They do lack of precision extremely badly.

Turing was a mathematician and probability (which is uncertainty in a contained environment) and the certainty and predictability of numbers was at the core of his work, along with some seriously geeky philosophers whose problems originated in maths. That's maths, not math. The 19th Century Germans (in England) and thinkers like Bertrand Russell were almost his contemporaries.

But thousands of years earlier, the concept of monism, that there is only one thing and all else (which are not things) are within that thing, was, in some respects, at the heart of the debates about what algorithmic analysis can be used for. Scrap that.

What is algorithmic analysis good for?

Cue the music "AI, What is it good for? Absolutely nothing."

That's a wild exaggeration: algorithmic analysis is one of the building blocks by which everything operates.

Everything.

"I'm a baby. I'm hungry. If I cry I get food; if I don't cry, I don't."

In the human world, we operate on graduated information: how hot is too hot?

But just as monism makes little or no sense except as a theoretical concept, so the idea of an electronic device that can, independently, make value judgements is senseless.

The world needs algorithms: it does not need facile, fake and flawed tech and the fanciful, even fictitious, stories its supporters tell.

If only people who bandy the term "AI" about would bother to learn more than the buzzwords and then to rush to jump aboard a bandwagon before it crashes we might not be facing the kind of uncertain future that we would face if monism were right and the thing were a squash ball. But we are.

In a world where people say "get real" then do the opposite, we need the balance redressed because it's been expensive to get into the mess that is looming: it will be exponentially more so to get out of it.