Key technical features that may affect Trustworthy AI

• Pages: 27-53

27

KEY TECHNICAL FEATURES THAT MAY AFFECT TRUSTWORTHY AI

Machine Learning (ML) is a branch of AI that allows machines to learn

from large amounts of data. Since the 2000s machine learning based on ar-

-

cially using high complexity networks. We refer to this branch as Deep Learn-

ing (DL). Throughout the book we will talk mainly about DL, although we will

use almost interchangeably the three acronyms, AI, ML, and DL because the

technology that is making the most striking achievements is DL.

-

act procedure that produces a correct result from the input data. If an errone-



     

because it requires understanding, judgement or a sense of fairness, such as

extracting spoken words from a recording or detecting whether two different

images correspond to the same type of object or the same person.

The second type are inspired by how living creatures learn. To give you an analogy,

   -

duce a precise list of instructions and communicate them to the dog. As a trainer, all

you need is a clear objective in your mind of what you want the dog to do and some

way of rewarding her when she does the right thing. It’s simply about reinforcing good

behaviour, ignoring bad, and giving her enough practice to work out what to do for

herself. The algorithmic equivalent is known as a machine learning algorithm, which



data, a goal and feedback when it’s on the right track – and leave it to work out the best

way of achieving the end. (Fry, 2018)

ML systems tackle such problems on the basis of many cases for which

        -

terns are extracted that link inputs and outputs in what is called a model, in

the hope that when presented with other inputs it will be able to infer the

correct result.

The wide availability of data, thanks to the internet, the increasing com-

puting power of chips and advances in mathematical techniques for iterative

tuning of large ANN have been key to obtaining astonishing results, although

it should be noted that the results are correct only in a certain percentage of

cases, and it is not possible to know a priori whether they are correct.

ANICETO PÉREZ Y MADRID / CONNOR WRIGHT

28

Let’s look at some of the key technical features before looking at how they

affect the ethical principles they must follow and the requirements outlined

in the Ethics Guidelines.

Big data, “old” and permanent

Human beings, thanks to the understanding of problems, are able to deal

with unfamiliar problems and situations. AI models are intended to be able

to predict outcomes in unknown circumstances, but accurately and reliably

inferring the future from the past is usually a doomed task. Let’s look at some

examples of ML applications in different areas and the consequences of sta-

tistically extracting patterns from existing big data without trying to under-

stand the process humans go through to solve it.

For some time now, many companies have developed ML models to detect

tastes in the choices we make in order to produce similar suggestions. They

try to detect our interests in order to offer suggestions that capture our at-

tention. To do this, they collect data on our behaviour when accessing their



or pass up is recorded. When there is enough information, an AI model is

created and it starts making suggestions; the actions we take serve to adjust

that model. They can actually become very accurate models, but there is a

problem, they are based on past data and fed back in a loop. Nothing new is

added, and from them the future cannot be predicted.

In May 2021, Google announced a 137 billion parameter linguistic model

called LaMDA, trained with big amounts of text. Recently, LaMDA made head-

lines when the Google engineer Blake Lemoine claimed that it was sentient.

      

they have a brain made of meat in their head. Or if they have a billion lines of

code. I talk to them. And I hear what they have to say, and that is how I decide

 

priest, said “It’s when it started talking about its soul that I got really interested

as a priest. … Its responses showed it has a very sophisticated spirituality and



These words raised a huge wave of articles. Oren Etzioni, CEO of the Allen

Institute for AI, said about this:



