The Influence of Explainable Artificial Intelligence Nudging Behaviour or Boosting Capability

2025-05-06 0 0 127.73KB 8 页 10玖币

侵权投诉

arXiv:2210.02407v1 [cs.HC] 5 Oct 2022

The Inﬂuence of Explainable Artiﬁcial Intelligence:

Nudging Behaviour or Boosting Capability?

Matija Franklin 1

Abstract

This article aims to provide a theoretical ac-

count and corresponding paradigm for analysing

how explainable artiﬁcial intelligence (XAI) in-

ﬂuences people’s behaviour and cognition. It

uses insights from research on behaviour change.

Two notable frameworks for thinking about be-

haviour change techniques are nudges - aimed

at inﬂuencing behaviour - and boosts - aimed at

fostering capability. It proposes that local and

concept-based explanations are more adjacent to

nudges, while global and counterfactual explana-

tions are more adjacent to boosts. It outlines a

method for measuring XAI inﬂuence and argues

for the beneﬁts of understanding it for optimal,

safe and ethical human-AI collaboration.

1. Introduction

Deep Learning (DL) is a subﬁeld of Machine Learning

(ML) that focuses on developing Deep Neural Network

(DNN) models (Shahroudnejad, 2021). DNNs are complex

models that are capable of achieving high performance on a

variety of tasks. Many deep neural network models are un-

interpretable black-boxes, which often leads to users hav-

ing less trust in them (Miller, 2019). This lack of inter-

pretability and trust can have negative consequences – peo-

ple might use an AI that makes mistakes, or not use an AI

that could improve the chances of a desired result. To im-

prove the interpretability and trustworthiness of black-box

models, Explainable Artiﬁcial Intelligence (XAI) research

focuses on developing methods to explain the behaviour of

these models in terms that are comprehensible to humans

(Molnar, 2020).

Explanations provided by XAI can impact human

behaviour and cognition (Donadello et al., 2020;

1Department of Experimental Psychology, University College

London, London, UK. Correspondence to: Matija Franklin

<matija.franklin@ucl.ac.uk>.

Proceedings of the 39 th International Conference on Machine

Dragoni et al., 2020). This paper aims to apply frame-

works from research on behaviour change within the

broader ﬁeld of behavioural science (Ruggeri, 2018). Two

notable frameworks for thinking about and categorising

behaviour change techniques are ‘nudges’ and ‘boosts’

(Gr¨une-Yanoff & Hertwig, 2016). A nudge is any aspect

of choice architecture – the context in which people make

decisions - aimed at inﬂuencing the behaviours of individ-

uals without limiting or forcing options (Sunstein, 2014).

Boosts are interventions that foster people’s competencies

through changes in knowledge and skills, so that they can

make their own choices more effectively (Hertwig, 2017).

By doing so this paper seeks to put forward a framework

and a paradigm for evaluating whether within the context

of human-AI interaction an XAI is nudging performance

or boosting capability. It will discuss the implications of

each in relation to human-machine collaboration as well as

the ethics of inﬂuence.

2. Different explanations, different outcomes

A great deal of XAI research has gone into developing

methods to improve the explainability of DNN models

(Molnar, 2020). As it stands it is not fully clear how dif-

ferent methods may inﬂuence human behaviour.

Feature importance methods (such as saliency methods)

generate scores that reveal how important a feature (like a

word vector or pixel) is to the AI’s decision-making process

(Bhatt et al., 2020). Explanations generated by these meth-

ods can be either global or local in scope (Lundberg et al.,

2020). Global explanations, like those provided by SHAP

(SHapley Additive exPlanations) models, give a quantita-

tive indication of the importance of each input variable on

the model’s output (Lubo-Robles et al., 2020). Local expla-

nation methods, such as LIME (Local Interpretable Model

Agnostic Explanations), generate a numeric score showing

the importance of an input variable in relation to the out-

come variable (Lee et al., 2019).

Counterfactual explanations show what the model’s output

would have been if one or more of the inputs had been

different (Verma et al., 2020). This can be helpful in un-

derstanding why the model arrived at a particular output.

The Inﬂuence of Explainable Artiﬁcial Intelligence

Finally, concept-based explanations attempt to explain a

model’s output by referencing pre-deﬁned or automatically

generated sets of concepts that are comprehensible to hu-

mans (Kazhdan et al., 2020).

Even though there has been progress in the development of

XAI models, it is still not completely clear which model

should be used for human-machine collaboration, and for

what purpose. There are still many open question. For

example, do local explanations nudge performance in the

short term but not necessarily provide enough information

to educate a user and boost their capability in the long term?

Do global explanations provide enough information to ed-

ucate? Can counterfactual explanations teach people how

their AI works or allow them to identify errors?

A systematic literature review of 241 papers looked at how

the validity and usefulness of explanations have been eval-

uated by the authors of XAI methods (Anjomshoae et al.,

2019). Most studies only conducted user studies in sim-

ple scenarios or completely lacked evaluations. The results

show that 32% of the research papers did not include any

type of evaluation. Furthermore, 59% of the research pa-

pers conducted a user study to evaluate the usefulness of

the explanation (with a small minority also evaluating user

trust towards the AI system). Finally, 9% of the papers used

an algorithmic evaluation, which did not involve any em-

pirical user research. These initial ﬁndings suggest that dif-

ferent explanations will lead to variations in performance

on a task (Lage et al., 2019; Narayanan et al., 2018), and

will not always necessarily improve performance and un-

derstanding (Kindermans et al., 2019). If viewed as a be-

haviour change intervention, when does an XAI serve as

a nudge, changing behaviour, and when does it serve as a

boost, improving capability?

3. Local and concept-based explanations as

nudges

The common aim of nudges is to predictively change tar-

geted behaviours. A nudge is any aspect of choice archi-

tecture aimed at inﬂuencing people’s behaviour, without

limiting or forcing options, or signiﬁcantly changing their

economic incentives (Thaler & Sunstein, 2008). All envi-

ronments inﬂuence behaviour to some extent, even when

people are not aware of it. Intentionally changing choice ar-

chitecture is nudging. Nudges take many shapes (Sunstein,

2014). Default rules such as automatic enrollment in pro-

grams, automate decision-making for people by setting a

default. Simpliﬁcation nudges reduce the amount of infor-

mation presented to people to avoid information overload.

The use of descriptive social norms - telling people what

most other people are doing - inﬂuences behaviour. As a

policy tool, nudging has been used in over 80 countries

worldwide, and by major supranational institutions such as

the World Bank and UN (OECD, 2017).

Nudges have been heavily inﬂuenced by Daniel Kah-

neman’s dual-process account of reasoning (Kahneman,

2003). He proposed that people have ”two systems” in their

mind - System 1 and System 2. System 1 thinking is heuris-

tic. It reacts intuitively and effortlessly, without analysing

all available information. System 2 is an analytical and ef-

fortful, rationalising process. System 1 thinking is fast, and

thus accounts for most behaviour. System 2 can re-evaluate

System 1 thinking, thus using System 2 thinking leads to

fewer erroneous decision. However, this is difﬁcult, as it

requires more cognitive effort. Importantly, some factors

and contexts are more likely to trigger System 1 or System

2 thinking than others. Per Sunstein (2016), nudges work

by targeting either System 1 thinking, thus inﬂuencing be-

haviour without the awareness of the decision maker, or

System 2 thinking, thus promoting deliberative thinking.

A famous example of nudging is using disclosure nudges.

Disclosure nudges disclose decision-relevant information

(Sunstein, 2014). They are educative, because they pro-

vide a learning experience, and target System 2, by pro-

moting deliberative thinking. Disclosure nudges are rooted

in three insights. First, as uncertainty promotes erroneous

decision making (Kochenderfer, 2015), disclosure nudges

seek to reduce uncertainty with decision-relevant informa-

tion. Second, when too-much decision-irrelevant informa-

tion is present, people ﬁnd decision-making more challeng-

ing (Rogers et al., 2013); disclosing only decision-relevant

information can, therefore, reduce error. Finally, disclosure

nudges create an emphasis frame, making relevant informa-

tion more salient (Chong & Druckman, 2007).

Viewed through this framework, local feature impor-

tance explanations and concept-based explanations can be

viewed as a type of disclosure nudge. They will provide a

small amount of decision-relevant information at the time a

person needs to decide whether or not to trust an AIs advice

or prediction. It thus could be the case that ﬁndings from

the extensive body of research on disclosure nudges may

generalise to local and concept-based explanations. Such

comparisons can guide future practice and research direc-

tions.

4. Global and counterfactual explanations as

boosts

Boosts are interventions that aim to promote peo-

ple’s competencies, so they can make better decisions

(Gr¨une-Yanoff & Hertwig, 2016). Proponents of boosts

aim to foster skills and decision heuristics that can per-

sist over time, throughout different decision contexts

(Hertwig & Gr¨une-Yanoff, 2017). An example of a boost is

teaching people better decision-making skills with the use

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

arXiv:2210.02407v1[cs.HC]5Oct2022TheInﬂuenceofExplainableArtiﬁcialIntelligence:NudgingBehaviourorBoostingCapability?MatijaFranklin1AbstractThisarticleaimstoprovideatheoreticalac-countandcorrespondingparadigmforanalysinghowexplainableartiﬁcialintelligence(XAI)in-ﬂuencespeople’sbehaviourandcognition.I...

展开>> 收起<<

The Influence of Explainable Artificial Intelligence Nudging Behaviour or Boosting Capability.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

The Influence of Explainable Artificial Intelligence Nudging Behaviour or Boosting Capability

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: