IB Questionbank

User interface language: English | Español

Date	May 2019	Marks available	2	Reference code	19M.1.SL.TZ0.4
Level	SL	Paper	1	Time zone	no time zone
Command term	Outline	Question number	4	Adapted from	N/A

Question

Sentencing criminals using artificial intelligence (AI)

In 10 states in the United States, artificial intelligence (AI) software is used for sentencing criminals. Once criminals are found guilty, judges need to determine the lengths of their prison sentences. One factor used by judges is the likelihood of the criminal re-offending*.

The AI software uses machine learning to determine how likely it is that a criminal will re-offend. This result is presented as a percentage; for example, the criminal has a 90 % chance of re-offending. Research has indicated that AI software is often, but not always, more reliable than human judges in predicting who is likely to re-offend.

There is general support for identifying people who are unlikely to re-offend, as they do not need to be sent to prisons that are already overcrowded.

Recently, Eric Loomis was sentenced by the state of Wisconsin using proprietary AI software. Eric had to answer over 100 questions to provide the AI software with enough information for it to decide the length of his sentence. When Eric was given a six-year sentence, he appealed and wanted to see the algorithms that led to this sentence. Eric lost the appeal.

On the other hand, the European Union (EU) has passed a law that allows citizens to challenge decisions made by algorithms in the criminal justice system.

* re-offending: committing another crime in the future

Identify two characteristics of artificial intelligence (AI) systems.

[2]

a.i.

Outline one problem that may arise if proprietary software rather than open-source software is used to develop algorithms.

[2]

a.ii.

The developers of the AI software decided to use supervised machine learning to develop the algorithms in the sentencing software.

Identify two advantages of using supervised learning.

[2]

a.iii.

The developers of the AI software used visualizations as part of the development process.

Explain one reason why visualizations would be used as part of the development process.

[2]

b.i.

Explain two problems the developers of the AI system could encounter when gathering the data that will be input into the AI system.

[4]

b.ii.

To what extent should the decisions of judges be based on algorithms rather than their knowledge and experience?

[8]

Markscheme

Answers may include:

Ability to seem intelligent
Power to copy intelligent human behaviour
Capacity to learn
Decision-making ability
Adaptation to circumstances
Well-defined goals
Problem-solving skills
Reasoning ability
Autonomy
Flexibility

Award [1] for identifying each characteristic of artificial intelligence systems up to a maximum of [2].

a.i.

Answers may include:

Proprietary software cannot be easily adapted to the needs of the user.
Changes to the software may take a long time / may not meet the needs of the user / may be driven by commercial interests.

Award [1] for identifying a problem of using proprietary software and [1] for a development of that problem up to a maximum of [2].

a.ii.

Answers may include:

More accurate (than unsupervised and reinforcement learning).
Easier to compare outcomes to predicted results / easier to train the algorithms.

Award [1] for identifying each advantage of using supervised learning up to a maximum of [2].

a.iii.

Answers may include:

Visualizes the flow of data through an information system…
which could make a potentially complex system understandable by a non-technical audience.

Award [1] for identifying a reason why visualizations may be used as part of the development process and [1] for a development of that reason up to a maximum of [2].

b.i.

Answers may include:

The different datasets may not be compatible with each other…
which means it has to be converted to a single format.

The various databases may have different structures…
fields in one database may not correspond to fields in other databases / a lot of the integrated data is of no value.

Award [1] for identifying a problem the developers of the AI system will encounter when gathering the data that will be input into the AI system and [1] for a development of that reason up to a maximum of [2].

b.ii.

Answers may include:

Arguments for using algorithms

We don’t know what is going on inside a judge’s mind, so it’s a black box too.
Judges can be biased.
This would be a more standardized process, i.e., more uniform and logical.
Risk assessment tools could lead to less incarceration and less crime.
This software could be used in conjunction with a judge’s decision i.e., to provide guidance.
It could be one factor among many, i.e., not the determinative factor.
The software could be used to identify outliers, i.e., people very likely to commit a crime or people very unlikely to commit a crime.
The software could be based on the experience of many judges/experienced experts in this field, so it could be more reliable than the judgement of a single person.

Arguments against using algorithms

The algorithms should not be “black boxes”, i.e., they should be revealed (transparency).
It’s almost impossible to define fairness, so how can the algorithm be fair?
Biases could be incorporated into the software because of the attitudes of the human beings who created it.
The value of the algorithm depends on the data it uses, and criminal justice data is often unreliable.
Algorithms look at group behaviour, not individual behaviour.
Algorithms don’t look at how different factors interact (at least not yet).
Is the AI system fit for purpose, will it give appropriate sentences?
May shift the power too far towards AI / machine learning.
Does not determine who is accountable if a sentence is unjust / inappropriate.

In part (c) of this question it is expected there will be a balance between the terminology related to digital systems and the terminology related to social and ethical impacts.

Keywords: laws, standardized, transparency, accountability, judgement, bias, fairness, automation, machine learning, algorithm, reliability, change, power, systems, values

Refer to SL/HL paper 1, part c markbands when awarding marks. These can be found under the "Your tests" tab > supplemental materials > Digital society markbands and guidance document.