Socially-Guided Machine Learning for

Self-Organised Community Empowerment

Asimina Mertzani

and Jeremy Pitt

Electrical and Electronic Engineering Dept., Imperial College London, London, U.K.

{asimina.mertzani20, j.pitt}@imperial.ac.uk

Keywords:

Human-Computer Interaction, Power-Sensitive Design, Generative AI, Self-Governance, Innovation.

Abstract:

Two key features of self-organising socio-technical systems are, ﬁrstly, the interaction of humans with AI,and

secondly, the collective determination of social arrangements. However, this presents the risk of an inequitable

distribution of power: either by translating or reinforcing existing power asymmetries directly into digital sys-

tems, unintended concession of power by human to computational, or arrogation of power by using AI as a

proxy. In this paper, based on a deﬁnition of empowerment, we implement a socially-guided machine-learning

system which integrates multi-agent system, generative AI and user-centred visualisation. The system is eval-

uated through proof-of-concept demonstrations showing how it could assist users in understanding the impact

of social arrangements and so empower communities with choice, control and innovation. The signiﬁcance of

this work is to show how, through the synergy of human expertise, generative AI and (multi-)agent-based sim-

ulations, it might be possible to enhance human creativity to imagine original social arrangements, visualise

their impact on community empowerment, and maintain an equitable distribution of power.

1 INTRODUCTION

The essence of community empowerment is that

those affected by mutually-agreed, and voluntarily-

complied with, social arrangements (rules, proce-

dures, structures, etc.) should participate in the selec-

tion, modiﬁcation and application of those arrange-

ments. However, humans, in general, have limited ex-

perience of and expertise in the practise of determina-

tion on issues of public interest. Also, there are intri-

cate interplays between community experience, social

arrangement and task complexity (Rychwalska et al.,

2021). Those in combination with the increasing hy-

bridisation (i.e. communities involving seamless in-

teractions between humans and AI) (Sarkadi, 2024)

might produce either of two possible outcomes. The

transition could be harmonious, and productive, even

if in unexpected ways (Metz, 2016), or it might be

harmful to individuals or damaging to the social fab-

ric in intended or unintended ways (Robbins, 2019).

For example, the role of technology in repro-

ducing a kind of feudalism has been observed (e.g.

(Zarkadakis, 2020)). This ‘techno-feudalism’ can be

attributed, in part, to an inequitable distribution of

https://orcid.org/0000-0002-6084-9212

https://orcid.org/0000-0003-4312-8904

power. This distribution may be a product of an inad-

vertent concession by “dumbing down” in the face of

a supposedly superior intelligence (Robbins, 2022).

However, it could also be a deliberate arrogation of

power by using AI as a proxy for reproducing, re-

inforcing and amplifying extant power imbalances in

socio-technical systems (Lewis et al., 2021).

This paper aims to avoid such harmful outcomes,

while also providing beneﬁcial ones, by offering a

novel tool for effective self-governance through the

co-production between humans and AI. This applies

the socially-guided machine learning methodology

(SGML) (Thomaz and Breazeal, 2006) to combine

codiﬁed social knowledge and human expertise with

Generative AI (GenAI) and multi-agent simulation

(MAS) in a system for opportunistic self-organisation

of innovative social arrangements for community em-

powerment. This derives from the use of GenAI

for unexpected linkage of diverse knowledge (Metz,

2016), and MAS for unexpected emergence of pro-

social behaviours (Mertzani et al., 2022).

Accordingly, this paper is structured as follows.

Section 2 elaborates on the background and motiva-

tion, with respect to power and empowerment, col-

lective deliberation, and SGML. Section 3 gives an

overview of the system for community empower-

ment; while Section 4 details the system implemen-

Mertzani, A. and Pitt, J.

Socially-Guided Machine Learning for Self-Organised Community Empowerment.

DOI: 10.5220/0013081900003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 2, pages 24-35

ISBN: 978-989-758-737-5; ISSN: 2184-433X

tation. This is evaluated through proof-of-concept

demonstrations in Section 5 showing how the system

could assist users in understanding the impact of so-

cial arrangements and so empower communities with

choice, control and innovation. After a comparative

discussion of related and further work in Section 6,

we conclude in Section 7 with the contributions to

interoceptive awareness and power-sensitive design,

through which it might be possible enhance human

creativity to imagine original social arrangements,

visualise their impact on community empowerment,

and maintain an equitable distribution of power.

2 BACKGROUND AND

MOTIVATION

2.1 Motivational Example

Open-plan ofﬁces are environments where multiple

individuals share a common space, while the produc-

tivity of those individuals can be affected by the be-

haviour of others. Previous work has proposed the

use of an anonymous online system for ﬂagging vio-

lations of the social norms to restore human relation-

ships and improve co-working conditions (Santos and

Pitt, 2014). This was implemented in an “affective

conditioning system” which combined elements of

normative, affective, and adaptive computing to sup-

port self-regulation of a co-working space. Some ex-

periments showed that psychological theories of for-

giveness could be used to restore a homeostatic equi-

librium after a normative violation, but was also de-

pendent on pre-existing pro-social relationships.

With the further development of Internet of Things

(IOT) and AI, the ambient environment of an open-

plan ofﬁce has become a socio-technical situation, in

which humans and AI co-exist. This means that hu-

mans and AI interact to make decisions relevant to the

self-governance of the co-working space. Such deci-

sions affect the social arrangements (SAs) which aim

to satisfy the individual preferences of the humans

while considering the requirements of the technical

system (e.g. minimised energy consumption). Differ-

ent SAs might be agreed based on the characteristics

of the individuals, the capabilities of the technology,

and the nature as well as the conditions of the envi-

ronment in which the space is situated.

In a simple scenario, humans might agree on a

ﬁxed set of rules concerning the behaviour of the indi-

viduals and the operation of technology. For instance,

they might agree to have their phones on silent and

do not have meetings in the shared space to decrease

noise levels, and to set the temperature equal to the

average preferred temperature of the individuals shar-

ing the space, and the air-conditioning (A/C) should

be turned on from nine to ﬁve. Accordingly, the air-

conditioning system would be set on the ﬁxed temper-

ature operating during the agreed times, while people

would put their phones on silent by the time they get

in and book a meeting room for having meetings.

However, individuals might realise that the agreed

temperature should be adjusted to the seasonal vari-

ations, so they might propose a new rule, that is the

change of temperature every month. It might then be

observed that the air-conditioning is on while nobody

is in the ofﬁce. This might result in a new SA that

detects motion in the space and turning on or off the

A/C as appropriate. Later, they might realise that not

all the people are concurrently in the space. Conse-

quently, they might propose another SA that considers

the preferences of the humans present at one time.

Overall, though, this scenario demonstrates that

there is a transition underway, as physical spaces

are increasingly saturated with sensors, and devices,

which can exhibit some form of intelligence. More-

over, the interaction between human intelligence and

this computational intelligence is dynamic rather than

static, focuses on peer deliberation rather than query-

answer, and involves co-production rather than provi-

sion. These features impact on issues of power, em-

powerment, and the self-determination of SAs.

2.2 Power and Empowerment

The previous section grounded the problem in a sce-

nario in which humans and AI co-exist in the same en-

vironment, highlighting the need for innovating social

arrangements for empowering communities to self-

determine their self-organisation. The primary mo-

tivation for this work is to achieve an equitable distri-

bution of power in contemporary socio-technical sys-

tems. However, to assess equity, we need to deﬁne the

‘measurable’ form(s) of power and empowerment.

Using modal logic, a formal characterisation of in-

stitutionalised power was given in (Jones and Sergot,

1996). This formalised the idea that an agent, occu-

pying a designated role in an institution, could create

facts of conventional signiﬁcance by the performance

of speciﬁc acts (e.g. a speech act), which “counted-

as” if the institution itself had done it. The equi-

table distribution of institutionalised power among the

agents in a MAS, was a key feature of a framework for

procedural justice speciﬁed in (Pitt et al., 2013).

This characterisation of “power” is precise and

even countable, but it is too narrow in the context

of general SAs. However, the terms power and em-

powerment have been studied and deﬁned in many

Socially-Guided Machine Learning for Self-Organised Community Empowerment

contexts, and not without contention (Adams, 2008).

Summarily, there are many different aspects, includ-

ing: sovereign power, interpreted as the monopoly

on violence, information and charisma (Graeber and

Wengrow, 2021); constitutional power, relating to the

creation and framing of SAs and deﬁning citizen-

ship; economic power, as the accumulation of scarce

resources and the leverage that provides; and even

raw compute power – machine learning algorithms

are computationally intensive, and democratising this

technology is currently infeasible.

For the purposes of this paper, we focus on two

other aspects of power. The ﬁrst aspect is a subjective

measurement of individual empowerment, whereby

‘intelligent’ and ‘reﬂective’ components in a socio-

technical system have the cognitive capacity to repre-

sent and reason with respect to ﬁve cognitive dimen-

sions of individual empowerment (referred as cogni-

tive DoEs)(Wach et al., 2016).These dimensions are

each individual’s sense of self-determination itself,

and an awareness of their competence in, inﬂuence

on, knowledge of, and meaning of this process, as

speciﬁed in (Mertzani and Pitt, 2024).

The second aspect is an objective measurement of

collective empowerment, referred to herein as com-

munity health. This is assessed by the following six

properties of community health: inclusivity, trans-

parency, diversity, equality, accountability and satis-

faction (as a form of interactional justice, i.e. how

well individuals feel they have been treated).

2.3 Empowerment of Deliberation

“Social arrangements”, informally introduced in

(Graeber and Wengrow, 2021), is an umbrella term

for any type of socially-constructed rule-based sys-

tem mutually agreed by members of a group, to vol-

untarily regulate their behaviour and hold themselves

accountable to one another. In this paper, we identify

SAs as the structures, rules and processes used in a

community to self-organise its empowerment, includ-

ing its polity (i.e. relationships to external actors).

For example, an SA focused on improving the

cognitive dimension of knowledge could be “Organ-

ise a monthly discussion panel related to democratic

decision-making”. In a human community the extent

of their knowledge could be determined (e.g. by sur-

vey) before and after introducing this SA, although it

might also have an effect on other dimensions.

However, this raises two questions: ﬁrstly, from

where do new SAs originate; and secondly, how can

the effect of an SA be evaluated? Ideally, the an-

swer to the ﬁrst question would be endogenous: it

would come from within the community itself. In

practice, given people’s limited experience and exper-

tise in such matters, some guidance is likely to prove

necessary. We propose that this task is very well-

suited to GenAI, in part because the models would

have been trained on large volumes of data from a va-

riety of sources, but also, as previously mentioned,

because of GenAI’s capacity to ﬁnd unexpected link-

ages in data, which can inspire human creativity and

innovation. Moreover, although the use of GenAI is

of contention, it has been shown to support human

creativity and brainstorming (Bouschery et al., 2023;

Memmert and Tavanapour, 2023).

The answer to the second question lies in mod-

elling and simulation. With a good model of the com-

munity, we could better understand the effect of the

SA. For this purpose, MAS are an appropriate tool:

we can design and implement a MAS that, to the ex-

tent that it reliably models the community, can be an-

imated and used to evaluate the effect of an SA on the

community through local interactions and social in-

ﬂuence. As previously mentioned, a key feature is the

emergence of unexpected pro-social behaviours.

2.4 Socially-Guided Machine Learning

In a variation of the model-view-controller pattern,

we use a MAS model, a visualisation of the empow-

erment of the agents, and two controllers: the human

user and GenAI, as illustrated in Figure 1. Either

the human user or the GenAI can recommend alter-

native SAs: the effect of these new SAs is simulated

in the MAS, and the impact on community empower-

ment is visualised for human ‘consumption’. The co-

production of SAs between aims to conﬁne the weak-

nesses of GenAI (e.g. biases, hallucinations), while

beneﬁt from its strengths to support human creativity.

Interactive Interface

Human

GenAI API

MAS

Figure 1: Socially-Guided Machine Learning.

Hence, using Socially-Guided Machine Learning

(SGML) methodology (Thomaz and Breazeal, 2006),

we iterate ﬁrst through a phase in which we run the

MAS and visualise its ﬁnal state to the user, and sec-

ond a phase in which the user evaluates that state

and proposes a change (with or without consulting

GenAI). This change is applied to the system and

leads to the next iteration.

As a result, because this is essentially a non-

deterministic cybernetic system whose outputs are its

own inputs, what happens to the community is more

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

signiﬁcant in determining its ﬁnal state than the start-

ing conditions. Resetting (or bootstrapping) the sys-

tem allows exploration of multiple different iterations

of proposed SAs, enabling evaluation of compara-

tive performance and long-term impact with differ-

ent combinations of GenAI and MAS behaviour, as

well as the opportunity for potentially unbounded co-

production of new SAs.

3 SYSTEM OVERVIEW

3.1 System Interface

The interface of the system aims to facilitate the in-

teraction between the user, GenAI, and MAS. Specif-

ically, it supports the following user activities:

• input of the SAs they want to test in the MAS

• get inspiration of new SAs by querying the GenAI

• access the results of applying an SA to the MAS

through visualisations (e.g. graphs)

Speciﬁcally, an indicative example of visualisation,

enabling the users access the effects of the iterative

application of an SA to the MAS is shown in Figure 2.

These spider plots show the extend in which each cog-

nitive DoE and community health property is fulﬁlled,

ranging from 0% to 100%. This interface is suitable

for being accessed via a web browser, while running

in a remote server, according to a client-server model,

however the current version of it is hosted locally.

Figure 2: Example Visualisation.

3.2 Operational Cycle

The control ﬂow of the system is described in Algo-

rithm 1 and in Figure 4. The modules speciﬁed are

analysed in the next Section. The text in green corre-

sponds to user input, text in blue is a displayed output,

and text in red is a module call. The operational cy-

cle starts with initialisation of the MAS, initialisation

of the interaction with the GenAI API, and an initial

simulation run of the MAS for a predeﬁned number of

rounds m, for which no SA is proposed or used. After

the completion of the ﬁrst epoch, dimensions of the

empowerment and community health properties are

displayed by the visualiser.

Algorithm 1: Control Flow.

1: Let t = 0, epochs = 0, rounds = 0, reset = false

2: System Initialiser. (Section 4.2)

3: for rounds = m do

4: Baseline Scenario Executor. (Section 4.4)

5: Parameter Collector.

6: rounds = rounds + 1, t = t + 1.

7: end for

8: Visualisation → Visualiser (Section 4.11)

9: rounds = 0, epochs = epochs + 1

10: while t < T and not(reset) do

11: Input1 → Ask if need for change of SA.

12: if Input1 ̸= No then

13: Need Detector. (Section 4.5)

14: Output1 → Need from Need Detector.

15: if Input1 == Human then

16: Input2 → Input of SA from User.

17: else if Input1 == GenAI then

18: GenAI Messenger. (Section 4.6)

19: Output2 → SA from GenAI Messenger.

20: Input2 → Ask for validation or modiﬁcation.

21: end if

22: SA Analyser. (Section 4.7)

23: t

= t.

24: end if

25: for rounds = m do

26: Baseline Scenario Executor. (Section 4.4)

27: SA Effect Calculator. (Section 4.8)

28: DoE Calculator. (Section 4.9)

29: Opinion Formation Executor. (Section 4.10)

30: Parameter Collector.

31: t = t +1, rounds = rounds + 1.

32: end for

33: Output3 → Informative Message from SA Effect

Calculator.

34: Visualisation → Visualiser (Section 4.11)

35: epochs = epochs + 1, rounds = 0.

36: Input3 → reset = User decides to bootstrap.

37: end while

The Program queries the user if s/he wants to

change the SA, and replying in the afﬁrmative, s/he

has also to decide if the new SA should be user- or

GenAI-generated. The system determines and dis-

plays what parameter is under-valued, and either asks

the user to input a new SA or queries the GenAI; in

the latter case, the user can accept or modify GenAI’s

answer. In either case, the outcome is a new SA; if

the user did not want to change the SA, the ‘new’ SA

in the next epoch is the same as the old SA.

The new SA is applied in another epoch of the

MAS, and its effects after each round are stored in

memory. After the completion of the epoch (m rounds

of the MAS), the impact of the SA in each cognitive

DoE and community health property are presented to

the user by the visualiser.

Finally, the user is asked to consider bootstrapping

the system, in which case the next cycle will start

from the initialisation stage, or if not, to proceed to

the next epoch, which then returns control to the orig-

inal change query. Resetting the system to its original

starting conditions, and re-running, enables the user,

and the system, to learn what SA, and what order of

SAs, have what impact in which situations.

Socially-Guided Machine Learning for Self-Organised Community Empowerment

3.3 Illustrative Walkthrough

Figure 3 provides an illustrative walkthrough of the

system operation: on the left, the visualisation of em-

powerment and community health; and on the right,

the user-system dialogue.

In the ﬁrst ‘row’, the user chooses to input his/her

own SA; the system recommends an SA to improve

the knowledge DoE, and the user inputs a new SA. In

the second ‘row’, the user chooses to query GenAI;

the system recommends an SA to improve the inﬂu-

ence DoE, and queries GenAI accordingly. The an-

swer produced can be accepted or modiﬁed (in this

case it is accepted). In the third row, the user de-

cides to stick with the current SA. In each case, the

impact of the SA on the MAS is computed and new

level of empowerment, in terms of cognitive DoEs

and community health properties, is demonstrated by

the change in shape of the spider plots.

Figure 3: Illustrative Visualisation and Dialogue.

4 SYSTEM IMPLEMENTATION

This section summarises the system’s implementa-

tion, giving a detailed speciﬁcation of each module,

and describing their information processing.

4.1 System Architecture

The Program is composed of ten modules:

• the System Initialiser, which instatiates the MAS

and the independent parameters.

• the MAS Simulator, which generates the MAS;

• the Baseline Scenario Executor, which runs the

MAS for an epoch (m rounds) without any SA;

• the Need Detector, which evaluates the state of the

MAS and determines the need for empowerment;

• the GenAI Messenger, which sends and receives

prompts by GenAI API;

System Initialiser

Baseline Scenario Executor

SA Effect Calculator

DoE Effect Calculator

Actor

Parameter Collector

Visualisation

iterative

application

in MAS

User observes the graphs

with the effects of SA

Need Detector

GenAI Messenger

Visualisation

User modiﬁes or keeps

generated SA

Actor

User inputs if there is a

need for change of SA

Actor

User inputs if they want to use their

own SA or co-create it with GenAI

Change

No Change

Actor

User observes the estimated need

User

GenAI

Interactive

Interface

Interactive

Interface

GenAI API

Actor

User observes the generated SA

Interactive

Interface

Baseline Scenario Executor

Parameter Collector

iterative

application

in MAS

Opinion Formation Executor

Actor

User observes the graphs

with the effects of SA

Actor

User inputs if they consider appropriate

bootstapping the sytem or not

Interactive

Interface

NoYes

User inputs an SA

Actor

Figure 4: System’s Architecture.

• the SA Analyser, which extracts the information

from the GenAI response or the input from the

user, and maps it to DoEs;

• the SA Effect Calculator, which grounds an SA to

the current round;

• the DoE Calculator, which calculates the cogni-

tive DoEs and community health properties for

the current round;

• the Opinion Formation Executor, in which agents

interact and compute the updated cognitive DoEs

and community health properties; and

• the Visualiser, which shows the empowerment af-

ter the iterative application of the new SA.

The way the modules process information is as

follows. System Initialiser assigns values to the pa-

rameters according to the inputs of the user. MAS

simulator generates the MAS according to the inputs

from the Initialiser and outputs an instance of itself

(e.g. a MAS comprising x agents etc.). Baseline Sce-

nario Executor receives the MAS instance and calcu-

lates agents’ cognitive DoEs for m rounds without us-

ing any SA. For instance, for each agent a in round r,

it calculates each cognitive DoE i, which would corre-

spond to cd

a,i,base

= y (base is used to differentiate the

baseline value of the DoE from the one after adding

the contribution from the SA) aggregates them, and

forms the corresponding CD

i,base

= Y .

Need Detector receives CD

i,base

and computes the

need, e.g. need

= inﬂuence. This together with the

known SAs are the inputs of the GenAI Messenger,

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

which interacts with GenAI API and outputs a new

SA, e.g. ‘Foster a culture of open communication’.

The new SA is sent to the SA Analyser, which outputs

a bonus vector corresponding to the mapping of SAs

to DoEs, e.g. bonus =[self-determination:1, compe-

tence:0, inﬂuence:2, etc.,]. SA Effect Calculator ini-

tialises the distribution of the effectiveness of an SA

over time and outputs the current effectiveness, for in-

stance effectiveness of SA i in round r is e f

Furthermore, the outputs of the Baseline Scenario

Executor, the SA Analyser, and the SA Effect Calcu-

lator become the inputs to the DoE Calculator which

calculates the agents’ current DoEs. For instance,

for the given scenario, its output would correspond

to cd

i,a

= Z and CH

= W . The agents’ cognitive

DoEs are sent to the Opinion Formation Calculator

who lets agents’ interact and choose to use their cd

i,a

or neighbours cd

i,a,nei

cognitive DoEs and form their

ﬁnal cd

i,a, f

. As such, the output of that module be-

comes the aggregate value of the agents’ cognitive

DoEs CD

i,a

and together with the CH

are sent to

the Visualiser which outputs a graph similar to that

of Figure 2. For brevity, when we refer to the DoEs of

all the agents we omit a and i. Overall, the modules

are summarised in Table 1, while the system’s archi-

tecture is given by Figure 4.

Table 1: Modules and Experimental Parameters.

Modules

Name Input Output

System Initialiser independent parameters Instantiated independent parameters

MAS Simulator Instantiated independent parameters MAS instance

Baseline Scenario Executor MAS instance cd

a,base

,CD

base

Need Detector CD

base

need

GenAI Messenger need

, KB new SA

SA Analyser new SA bonus

SA Effect Calculator Gamma params., t

, t e f

DoE Calculator cd

base

, bonus,e f

, CH

Opinion Formation Executor cd

own

, SN,cr,sc CD

Visualiser CD

, CH

Graphs (e.g. Fig2)

Independent and Exp. Determined parameters

Symbol Description In/ExD Value

N number of agents In 100

T,m maximum duration in epochs and rounds in an epoch In 50,25

init

initial knowledge base of SAs In

cgmax,dev max ﬁxed agents’ cognitive DoE and allowed deviation ExD 10,-

−d, d agents’ cognitive DoE allowed deviation interval ExD [−1,1]

k, θ, loc shape, scale and time shift of Gamma distr. ExD 5, 10, -

loc

,loc

time shift ﬁxed and random part ExD 1,-

loc

min

,loc

max

lower/upper bounds of time shift random part ExD [-5,5]

mul PDF multiplier of Gamma ExD -

mul

,mul

mult. ﬁxed and random parts ExD 50,-

mul

min

,mul

max

lower/upper bounds of mult. random part ExD [-25,25]

bonus

max

max. reinforcement to a DoE ExD 10

τ critical health message threshold ExD 45

c reinforcement of credence/conﬁdence in a round ExD 0.01

a,n,init

, sc

a,init

initial credence of agent a in n/self-conﬁdence of a ExD 0.5,0.5

4.2 System Initialiser

The initialisation of the system includes the instatia-

tion of the MAS and the deﬁnition of the experimen-

tal parameters. Table 1 gives an overview of the in-

dependent and experimentally determined parameters

and presents the values assigned to them in the exper-

iments below.

4.3 MAS Deﬁnition

The MAS corresponds to a self-organising institution

comprising N agents, having a knowledge base KB,

a list with the ﬁve cognitive DoEs CD (where CD

is the collective value of them in time t), and one

having the collective DoE (corresponding to commu-

nity health properties). Each agent a is initialised

with a ﬁxed individual value for each cognitive DoE

(randomly elected from the interval [0,cgmax],

where cgmax is an experimentally determined param-

eter) and in each round this value can deviate from

the ﬁxed by a random value dev from the interval

[−d, d], d > 0, e.g. cd

= cd

+ dev. This is to gen-

erate some agents being ‘experts’, i.e. having higher

cognitive DoEs and consequently being more empow-

ered. Also, each agent a has a social network, and has

credence cr

a,n

to each of their neighbours n, but also

self-conﬁdence sc

, and agents are initialised to have

equal self-conﬁdence and credence to all neighbours.

Also, Table 2 provides an example of KB.

4.4 Baseline Scenario Executor

In the baseline, no SAs known,so the agents itera-

tively form their cd

base

and the total value of each

DoE i is given by Equation 1:

i,base

∑

a∈N

a,i

cgmax ∗ N

∗ 100 (1)

where N is the number of agents, and cd

a,i

is agent’s

a value of each cognitive DoE i in round r. How-

ever, the collective DoEs are zero in the baseline, i.e.

base

= 0. This is to reﬂect the lack of community

health when there is no knowledge about SAs.

4.5 Need Detector

Need Detector senses the MAS and detects the current

need in terms of SAs. Therefore, it constitutes an in-

ternal mechanism of interoceptive awareness (Pitt and

Nowak, 2014) developed in the MAS, which com-

pares the cognitive DoEs and highlights a ‘threat to

the body politic’. Speciﬁcally, it receives the past cog-

nitive DoEs, it calculates the rate of change of each

i,rate

, and outputs the need

which corresponds to

the cognitive DoE i that has the greatest negative rate

of change in the current epoch t, given by Equation 2:

need

= argmax

∑

e=1

(CD

− CD

e−1

)

(2)

4.6 GenAI Messenger

The GenAI Messenger receives the current need need

and the KB, and composes a message in which it spec-

Socially-Guided Machine Learning for Self-Organised Community Empowerment

iﬁes the need and the known SAs for that need (i.e. the

corresponding row in the KB-Table). For instance, if

the need is ‘self-determination’ and the KB is that of

Table 2 an indicative interaction between the system

and the API is:

Message: “The population needs a social ar-

rangement to improve the individuals’ cognitive di-

mension of self-determination, and already knows

the following: ‘Promote active engagement’. Can

you give another one?”

Response: “Empower individuals to have a voice

and take ownership in shaping their communities.”

The message is sent to the GPT-4o mini API (OpenAI,

2024) and the reply corresponds to the new SA, which

is shown to the user for modiﬁcation or approval.

4.7 SA Analyser

The SA Analyser receives the input from the user or

the (user-modiﬁed or not) response from GenAI, ei-

ther of which corresponds to a new SA, and outputs

a vector, named bonus, with the expected impact of a

new SA to the eleven DoEs. To convert the SA to a

vector, we deﬁne a vocabulary comprising terms (e.g.

words and phrases) that are mapped to DoEs. So, each

time a new SA is given, it is analysed, as described

below, and the terms in it are used to deﬁne the rein-

forcement, which is then speciﬁed in the bonus vector.

Table 2: Example KB and Part of Vocabulary.

Cognitive DoE SAs

Self-determination Promote active engagement.

Competence Empowerment through education and support.

Promoting self-advocacy and decision-making skills.

Inﬂuence Foster a culture of open communication.

Knowledge -

Meaning Encourage civic education and run several assessments

Conduct surveys to gather input and address any discrepancies.

Part of Vocabulary

Term S-D C Inﬂ K M Incl D A Eq T

Active Participation Y N N N N N N N N N

Open Communication N N Y N N Y N N N Y

Ownership Y N N N Y N Y Y N N

An indicative vocabulary is given by Table 2 and

the full vocabulary is available here. The ﬁrst column

presents the terms, and the other columns correspond

to the eleven DoEs, where the letter ‘Y’ denotes that

the term affects the DoE while ‘N’ shows that it does

not. The steps are the following:

1. the bonus vector assigns a zero to each DoE

2. the punctuation marks of the new SA are removed

and the text is converted to lowercase

3. the SA is parsed and the analyser is looking for

key-terms that are included in the vocabulary

4. if a keyword is detected, then the analyser looks

for ‘Y’ in the vocabulary and increases the value

of that DoE in the bonus.

For instance, if the new SA suggests to ‘Foster a cul-

ture of open communication, collaboration, critical

thinking, and ethical behaviour to engage individuals

in meaningful contributions to collective knowledge,

and decision-making’, then the bonus would be:

bonus = [self-determination: 3, competence: 2, inﬂuence:

3, knowledge: 3, meaning: 4, inclusivity: 3, transparency:

2, diversity: 0, equality: 1, accountability: 1]

4.8 SA Effect Calculator

By analogy to a community in which humans need

time to process and engage with changes, the sys-

tem is designed so that each time a new SA is ap-

plied in the MAS, the population requires some sys-

tem time to understand it, engage with it and derive

the beneﬁts, as discussed in (Kahneman, 2011). Also,

after some time, the population or the environment

might change, and that SA might not be relevant any-

more. So, we set the effectiveness of an SA to follow

a Gamma distribution, as this is also used to model

the waiting time for a drug to reach its maximum ef-

fect in the body. Moreover, since different SAs re-

quire a different amount of time to be accepted, and

others might be more or less effective, we generate a

new Gamma distribution for each new SA, with the

following properties.

The shape k and the scale θ of the Gamma dis-

tribution of each SA are experimentally determined

parameters. Moreover, the distribution is shifted on

time by loc, which is a parameter having a ﬁxed

loc

and a random loc

D which is sampled from

a uniform distribution deﬁned in the [loc

min

, loc

max

]

interval, and is equal to loc = loc

+ loc

, where

loc

∼ Uniform(loc

min

, loc

max

), and loc

, loc

min

and

loc

max

are experimental parameters. Also, to make

the probability density function (PDF) of the Gamma

distribution to take values from minus one to one, in

the speciﬁed interval, its value is multiplied by mul

which has a ﬁxed mul

and a random mul

D deﬁned

similarly with the DoEs of the shift.

As such, this module takes as input the experimen-

tal parameters k, θ, loc, loc

min

, loc

max

, mul, mul

min

and mul

max

, the current t and the number of rounds

since the SA application t

, and calculates the effec-

tiveness e f

t−t

of SA i in time t. This is equal to

the value of the probability density function (PDF) of

Gamma at the current time t minus the time that the

new SA was ﬁrst applied t

, multiplied by mul, and

shifted by loc, given by Equation 3:

t−t

= mul ∗ f (t −t

;k, θ) + loc (3)

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

This module has an additional control developing a

second layer of interoceptive awareness (in the human

computer interaction), on top of the one developed in

the Need Detector (in the MAS). This calculates the

gradient of the distribution of the effectiveness on the

current time instance after the application of the SA

(t − t

) and makes the user aware of whether it has

reached its maximum effectiveness (negative gradi-

ent) or not (positive gradient).

So, depending on the sign of ∇ef

t−t

the mes-

sages ‘Maximum effectiveness has been reached’ and

‘Wait more for maximum effectiveness’ are shown to

the user. Furthermore, if the SA has reached maxi-

mum effectiveness and any of the DoEs is lower than

a threshold τ, a warning message saying ‘Community

is disempowered, consider a change!’ is shown.

4.9 DoE Calculator

With input the baseline cognitive DoEs, the bonus for

the SA i, and its effectiveness e f

, the DoE Calculator

outputs the agents’ DoEs. The value of an agent’s a

cognitive DoEs are cd

a,i

, the populations’ cognitive

DoE CD

equals the average of cd

a,i

,∀a ∈ A, and the

collective DoEs are CH

, given by:

a,i

a,i,base

∗ bonus ∗ e f

bonus

max

, CH

bonus ∗ e f

bonus

max

where bonus is the bonus of the current SA and

bonus

max

is an experimental parameter that gives the

maximum bonus that an SA can cause to a DoE.

4.10 Opinion Formation Executor

The Opinion Formation Executor is responsible for

the interaction of the agents and the formation of their

ﬁnal opinion in terms of cognitive DoEs. In particu-

lar, agents interact with their social network and they

can choose to use their own cognitive DoEs or to ask a

neighbour. This is to make MAS simulate a commu-

nity, in which agents interact and can choose to use

their own opinion or ask a source to optimise their de-

cision making, as described in (Nowak et al., 2019).

As such, an agent a has their cd

and selects one

of the most credited neighbours n for their cd

(corre-

sponding to cd

a,nei

). Based on cr

a,n

and sc

, a uses the

own or asked cognitive DoEs in their ﬁnal cd

a, f

, given

by Equation 4, while the population’s ﬁnal cognitive

DoEs CD correspond to their mean:

a, f

(

a,nei

, if cr

a,n

≥ sc

, if cr

a,n

< sc

(4)

Then, a updates the credence to n and self-conﬁdence

based on whether the average value of n’s cognitive

DoEs is greater than a’s. The amount of the reinforce-

ment of those per round is deﬁned by an experimen-

tally determined parameter called c, and the credence

reinforcement is given by Equation 5, while the oppo-

site applies for the self-conﬁdence:

a,n

(

a,n

∗ (1 + c), if

∑

i∈CD

i,a,nei

≥

∑

i∈CD

i,a

a,n

∗ (1 − c), otherwise

(5)

Note that the opinion formation affects only the

agents’ cognitive DoEs and not the community health

properties. This is because agents individually can

optimise their cognitive capacity, but the community

health is an emergent property that is the outcome of

cognitive empowerment and well-being. Therefore, it

cannot be optimised by simply asking a source.

4.11 Visualiser

After running the MAS for m rounds, the system has

to inform the user regarding the effects of an SA to

it. Therefore, the Visualiser receives the DoEs and

generates some informative graphs. An indicative ex-

amples is that given by Figure 2.

5 EVALUATION AS PROOF OF

CONCEPT

The system developed is non-deterministic, implying

that the sequence of events is more signiﬁcant than

the starting conditions, so different order of inputs,

can result in different outputs. However, this section

gives proof of concept examples which demonstrate

the development of interoceptive awareness, the va-

riety of SA effects, and the role of the user and the

emergence of expertise through bootstrapping.

The MAS deﬁned for that experimentation has

the minimum amount of characteristics to showcase

that it can be relevant to multiple different applica-

tions. Speciﬁcally, it comprises N agents initialised

with randomised values of DoEs forming a Klem-

Eguiluz social network, which resembles real-life net-

works (Prettejohn et al., 2011), and the values of the

other related parameters are given by Table 1. A more

elaborate deﬁnition of the MAS in a speciﬁc context

would increase the leverage of the users.

5.1 Interoceptive Awareness

To facilitate the empowerment of communities

through the self-determination of SAs, the system is

designed to have mechanisms that inform the users

Socially-Guided Machine Learning for Self-Organised Community Empowerment

and enable them develop interoceptive awareness

with respect to the system’s health (second level in-

teroceptive awareness mechanism).

Figure 5 showcases this property of the system,

where the user initially identiﬁes the need for change

and uses GenAI (ﬁrst grey box). The SA proposed is

‘establish clear communication channels, utilize col-

laborative tools, encourage knowledge sharing and

collaboration, and regularly review and update knowl-

edge repositories’, and is approved. This progres-

sively results in increased empowerment, mainly in

terms of knowledge, inﬂuence, diversity, inclusivity,

transparency, and satisfaction. However, after some

epochs, the SA is not effective anymore resulting in

decreasing the DoEs, shown in the spider plots. This

is observed by the system which displays a warning

(highlighted in red in the Figure).

Increasing Empowerment

Decreasing Empowerment

Increasing Empowerment

Figure 5: User’s Interoceptive Awareness.

Accordingly, the user proposes a change using

GenAI, and the new SA is ‘Promote active civic en-

gagement, and civic participation, and provide op-

portunities for meaningful and equal participation in

decision-making.’ which is approved and results in

increasing DoEs. Therefore, the system owns a mech-

anism of self-healing which is achieved through the

effective user-computer interaction. Speciﬁcally, the

system makes the user aware of the need, the user

acknowledges that and triggers a change, and that

change is applied to the system and enable its recov-

ery from a state of disempowerment.

5.2 Variety of SA Effects

As discussed above, different SAs result in different

outcomes in terms of DoEs. The upper part of Figure

6 shows the DoEs in the baseline compared to those

when applying two different SAs, which are ‘Estab-

lish clear communication channels, utilise collabora-

tive tools, and regularly update and share information

among team members.’ during round 600, and ‘Con-

duct regular surveys or feedback sessions to gather

input and address any discrepancies’ during 800.

Comparison of

Effects of SAs

and Baseline

Variety of SAs’ Effects and Human

Carelessness

Figure 6: Variety of SAs and Human Carelessness.

Additionally, the lower part of Figure 6 shows the

effect of the following SAs:

• Regularly review practices against core values and en-

courage open communication for dissent.

• Promote inclusive and transparent decision-making to

empower individuals to actively participate in shaping

their communities.

• Foster a culture of inclusivity, transparency, critical

thinking, and accountability.

• Establish clear communication channels, utilize collab-

orative tools, and encourage knowledge sharing.

Notice that different SAs address different needs and

therefore result in new DoEs, while all outperform the

baseline. For example, the ﬁrst SA increases mainly

meaning, diversity and inclusivity, while the fourth in-

creases inﬂuence, knowledge, and transparency. This

highlights how proposed SAs should be based on the

need of the system, which is also the reason why the

system detects the need and uses that to optimise its

decision-making with respect to SAs.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

5.3 Importance of Human Commitment

Focusing on the role of the user, the right side of the

second row of Figure 6 provides an example of in-

teraction in which the user is indifferent, and there-

fore cannot beneﬁt from the system’s suggestions. In

particular, despite the user opting to make a change,

their input is irrelevant (e.g.,“I do not know”). In the

next epochs, although the DoEs are decreased the user

does not intervene, e.g. they do not trigger a change

as highlighted in red boxes in the Figure (e.g. “no”).

Overall, the carelessness of the user and their indiffer-

ence to the warnings results in individual and collec-

tive disempowerment (reﬂected by low DoEs).

Figure 7: Human Impatience.

Another aspect for the effective operation of the

system is that the user is patient. Figure 7 shows a

scenario in which the user consecutively asks GenAI

to propose a new SA, and they do not wait for this to

be effective. Speciﬁcally, the AI generated SAs are

the following:

• Encourage individuals to engage in shaping their

communities through inclusive decision-making

processes;

• Support people to share perspectives, listen ac-

tively, seek evidence-based solutions, and uphold

ethical principles;

• Promote self-advocacy and decision-making.

Although the SAs can be beneﬁcial for the system, its

impact does not become apparent due to the user’s im-

patience (reﬂected by quick changes from SA to SA).

This way the population in MAS does not manage to

engage with the new SA, remaining disempowered.

Therefore, it is critical that the user follows the sug-

gestions of the system (i.e. message ‘Wait more for

maximum effectiveness’) and waits till the effects of

the change become measurable.

5.4 Opinion Formation in MAS

As discussed in the Opinion Formation Executor,

agents can decide to ask a source (a neighbouring

agent from the social network having higher cogni-

tive DoEs). Therefore, this highlights the emergence

of expertise and speciﬁcally shows how agents learn

to distinguish the ‘experts’ and seek for their opinion

(corresponding to cognitive DoEs).

This emergent property can be observed in Fig-

ure 8, where the ﬁrst column shows the amount of

the agents using their own (red) and their neighbours

(blue) cognitive DoEs to form their ﬁnal cognitive

DoE, and the second row shows the population’s cog-

nitive DoEs where they are using their own (red) com-

pared to the ones that they ﬁnally select (blue). There-

fore, the ﬁrst graph shows that agents in the MAS

learn to use their credited sources’ cognitive DoEs,

fact that results in higher cognitive DoEs reﬂected by

the graph in the second row. As such, it becomes ev-

ident that not only they ask their neighbours, but also

identify the experts, and overall optimise their com-

munity’ cognitive capacity leading to empowerment.

Selected Cognitive

DoEs

Cognitive DoEs

Own vs Final

Figure 8: Effect of Social Inﬂuence in Cognitive DoEs.

6 RELATED AND FURTHER

WORK

6.1 Related Work

Our work contributes to initiatives for support-

ing localised self-governance, evidence-based law-

making and enactment, and effective human-AI co-

production. This section discusses how it builds on

those attempts and combines them to offer a mecha-

nism that combines MAS, GenAI, and human users

to support community empowerment.

In the ﬁeld of political science, (Manville and

Ober, 2019)have called for a new “release” of democ-

Socially-Guided Machine Learning for Self-Organised Community Empowerment

racy, Democracy 4.0. They propose new architectures

of engagement, which includes not just more direct

participation and less “representative” approaches to

decision-making, but also channelling democratic en-

ergy and initiative into smaller scale, but more per-

sonally meaningful, forms of self-governance (cf.

(Bookchin, 2004)). This system complements their

call by providing a technology that could support this

localised self-governance by making accessible cause

and effect relations to the users.

In the domain of legal drafting, it is observed that

“traditional drafting methodologies produce laws that

do not work” (Seidman and Seidman, 2009). This

failure is attributed, in part, to the focus of legislative

scholarship, in the study of power, being on the pro-

cess of compromise between legislators on detailed

provisions before enactment. The four-step ILTAM

methodology is intended to assist law-makers en-

act evidence-based legislation that works as intended.

With our system, we are not only transferring power

from legislators to those affected by legislation, but

also providing evidence (using MAS) of the impact

of new SAs on the extent of that empowerment.

There is awareness of the need for a better un-

derstanding of human-AI co-production (Thomaz and

Breazeal, 2006), and hybrid systems integrating ma-

chine learning and machine reasoning (Kierner et al.,

2023). Our work contributes to both initiatives, in

the former by allowing the human and AI to work in

tandem on the selection and recommendation of new

SAs, and in the latter by enabling GenAI and MAS to

work in tandem on the application and effect of SAs.

This system also admits a potential for brainstorm-

ing and enrichment of human knowledge through re-

peated exploration of different event sequences being

applied to the same starting conditions.

6.2 Further Work

Although substantive results have been demonstrated

as a proof of concept, this system is at present a work

in progress, and there is a number of limitations to

overcome and improvements to be made in further re-

search and development. These include:

• in the SA Analyser, we need a substantive user

survey to establish a stronger correlation between

keywords and the DoEs;

• in the MAS, a method for mapping the features

and characteristics of a community to the MAS is

required;

• in the Need Detector, we need to develop a more

complex mechanism for interoceptive awareness

that takes into consideration multiple parameters;

• in the Visualiser, although some text entry (of SA)

is required, we should upgrade the text-based dia-

logue to improve the UX;

• generally, considering groups of users or experts

could improve the system’s effectiveness, while

further exploration on issues of model misalign-

ment, scale, and change on human behaviour due

to the interaction with the system is needed; and

• overall, we aim to package the system as a plug-in

for PlatformOcean (Pitt et al., 2021) and evaluate

performance in ﬁeld trials.

Following these developments, we would aim to de-

ploy and evaluate the system in a ﬁle trial, with sev-

eral possible applications, for example in deliberative

assembles for public policy, or co-housing for local

communal policy formation.

7 CONCLUSIONS

This paper has identiﬁed an inequitable distribution

of power as a fundamental challenge for the devel-

opment of socio-technical systems. Accordingly, it

has developed a system that helps people design their

communities better, using a model of equitable soci-

ety as a guide. This system, enables users to visualise

the extent of community empowerment in cognitive

and collective dimensions, and through a synthesis

of GenAI, MAS and self-organisation, to explore and

evaluate the impact of new SAs on their empower-

ment. While each individual module executes a rela-

tively simple process (and will be enhanced in further

work), the collective assembly of them forms an end-

to-end functional system which displays a rich variety

of behaviours and demonstrates the proof of concept.

The primary contributions are threefold. Firstly,

it implements anan innovative system for the self-

organisation of social arrangements and the visualisa-

tion of community empowerment. Secondly, it offers

a demonstration of power-sensitive design (Mertzani

and Pitt, 2024), an instance of value-sensitive design

in which the qualitative human values being targeted

are power and empowerment. Thirdly, it provides

a demonstration of institutional interoceptive aware-

ness as a community detects and responds to a situa-

tion that is affecting their empowerment.

The signiﬁcance of this work is that it proposes a

multi-component system which combines human so-

cial knowledge and expertise, with the creative capac-

ity of GenAI, stemming from the unexpected linkage

of diverse knowledge, and the capability to observe

emergence of agent-based simulations in MAS. The

system assists humans to think out-of-the-box and

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

envision future trajectories of alternative solutions,

which enables them to effectively self-determine their

social arrangements and maintain an equitable distri-

bution of power. This ‘tool’ could support deliber-

ative assemblies, such as humans sharing an ofﬁce,

or stakeholders of a co-housing project, to shape their

social arrangements such that they improve their lives.

REFERENCES

Adams, R. (2008). Empowerment, participation and social

work. New York, NY: Palgrave Macmillan.

Bookchin, M. (2004). Post-Sarcity Anaychism. AK Press.

Bouschery, S. G., Blazevic, V., and Piller, F. T. (2023).

Augmenting human innovation teams with artiﬁcial

intelligence: Exploring transformer-based language

models. Journal of Product Innovation Management,

40(2):139–153.

Graeber, D. and Wengrow, D. (2021). The Dawn of Ev-

erything: A New History of Humanity. London, UK:

Allen Lane.

Jones, A. and Sergot, M. (1996). A formal characterisa-

tion of institutionalised power. Journal of the IGPL,

4(3):427–443.

Kahneman, D. (2011). Thinking, fast and slow. New York,

NY: Farrar, Straus and Giroux.

Kierner, S., Kucharski, J., and Kierner, Z. (2023). Tax-

onomy of hybrid architectures involving rule-based

reasoning and machine learning in clinical decision

systems: A scoping review. J Biomed Inform.,

144:104428.

Lewis, P. R., Marsh, S., and Pitt, J. (2021). AI vs ‘AI’:

Synthetic minds or speech acts. IEEE Tech. Soc. Mag.,

40(2):6–13.

Manville, B. and Ober, J. (2019). In search of Democracy

4.0: Is democracy as we know it destined to die? IEEE

Tech. Soc. Mag., 38(1):32–42.

Memmert, L. and Tavanapour, N. (2023). Towards human-

ai-collaboration in brainstorming: Empirical insights

into the perception of working with a generative AI.

In Aanestad, M., Klein, S., Tarafdar, M., Han, S.,

Laumer, S., and Ramos, I., editors, 31st European

Conference on Information Systems (ECIS).

Mertzani, A. and Pitt, J. (2024). Power-sensitive design us-

ing higher-order cybernetics patterns. In Artiﬁcial Life

Conference Proceedings 36, volume 2024, page 126.

Cambridge, MA: MIT Press.

Mertzani, A., Pitt, J., Nowak, A., and Michalak, T. (2022).

Expertise, social inﬂuence, and knowledge aggrega-

tion in distributed information processing. Artiﬁcial

Life, 29(1):37–65.

Metz, C. (2016). In two moves, AlphaGo and Lee Sedol

redeﬁned the future. In Wired (2016), available at

https://tinyurl.com/twomoves.

Nowak, A., Vallacher, R., Rychwalska, A., Roszczynska,

M., Ziembowicz, K., Biesaga, M., and Kacprzyk, M.

(2019). Target in control: Social inﬂuence as dis-

tributed information processing. Springer.

OpenAI (2024). ChatGPT API. https://chat.openai.com.

Pitt, J., Busquets, D., and Riveret, R. (2013). Procedu-

ral justice and ‘ﬁtness for purpose’ of self-organising

electronic institutions. In Proc. PRIMA, volume 8291

of LNCS, pages 260–275. Cham, CH: Springer.

Pitt, J. and Nowak, A. (2014). Collective awareness and the

new institution science. In Pitt, J., editor, The Com-

puter After Me, chapter 12, pages 207–218. London,

UK: ICPress.

Pitt, S., Lacey, M., Scaife, E., and Pitt, J. (2021). No app

is an island: Collective action and sustainable devel-

opment goal-sensitive design. International Journal

of Interactive Multimedia and Artiﬁcial Intelligence,

6:24–33.

Prettejohn, B., Berryman, M., and Mcdonnell, M. (2011).

Methods for generating complex networks with se-

lected structural properties for simulations: A review

and tutorial for neuroscientists. Frontiers in computa-

tional neuroscience, 5:11.

Robbins, J. (2019). If technology is a parasite masquerading

as a symbiont, are we the host? IEEE Tech. Soc. Mag.,

38(3):24–33.

Robbins, J. (2022). The intelligence factor: Technology and

the missing link. IEEE Tech. Soc. Mag., 41(1):82–93.

Rychwalska, A., Roszczy

nska-Kurasi

nska, M., Ziembow-

icz, K., and Pitt, J. (2021). Fitness for purpose in on-

line communities: Community complexity framework

for diagnosis and design of socio-technical systems.

Front. Psychol., page 12:739415.

Santos, M. S. and Pitt, J. (2014). Emotions and norms in

shared spaces. In Balke, T., Dignum, F., van Riems-

dijk, M. B., and Chopra, A. K., editors, Coordination,

Organizations, Institutions, and Norms in Agent Sys-

tems IX, pages 157–176. Cham, CH: Springer.

Sarkadi, S. (2024). Self-governing hybrid societies and de-

ception. ACM Trans. Auton. Adapt. Syst., 19(2):9:1–

24.

Seidman, A. and Seidman, R. (2009). ILTAM: Draft-

ing evidence-based legislation for democratic social

change. BUL Rev., 89:435.

Thomaz, A. and Breazeal, C. (2006). Reinforcement learn-

ing with human teachers: Evidence of feedback and

guidance with implications for learning performance.

In AAAI Conf. on Artiﬁcial Intelligence.

Wach, F., Karbach, J., Rufﬁng, S., Br

unken, R., and

Spinath., F. (2016). University students’ satisfaction

with their academic studies: Personality and motiva-

tion matter. Frontiers in Psychology, 7:1–55.

Zarkadakis, G. (2020). Cyber Republic: Reinventing

Democracy in the Age of Intelligent Machines. Cam-

bridge, MA: MIT Press.

Socially-Guided Machine Learning for Self-Organised Community Empowerment