AwarePrompt: Using Diffusion Models to Create Methods for Measuring

Value-Aware AI Architectures

Kinga Ciupinska

1 a

, Serena Marchesi

1 b

, Giulio Antonio Abbo

2 c

Tony Belpaeme

2 d

and Agnieszka Wykowska

1 e

Social Cognition in Human-Robot Interaction (S4HRI), Italian Institute of Technology, Genova, Italy

IDLab-AIRO, Ghent University, imec, Belgium

Keywords:

Awareness, Generative AI, Diffusion Models, Value-Aware AI, Insight, Neural Correlates.

Abstract:

The integration of diffusion models (DMs) into generative AI systems presents an approach with implications

for ethical and moral AI development and our understanding of human-AI interaction. This study explores

the intersection of generative AI, human values, and neuroscience, emphasizing the signiﬁcance of value-

awareness in AI systems. The methodology involves a behavioral experiment to evaluate the accuracy of

DM-generated visual stimuli in capturing human values and related keywords. Results indicate promising

match rates, marking stride in aligning AI systems with ethical and moral considerations. Additionally, the

study introduces a criterion for selecting stimuli based on an “Aha” moment, setting the stage for an EEG

experiment to explore the neural correlates associated with becoming aware of a value. This multidisciplinary

study is a step toward the development of procedures to evaluate the effectiveness of Value-Aware AI systems

in enhancing the perceived ethical and moral agency.

1 INTRODUCTION

Artiﬁcial Intelligence (AI) has become an integral

part of our daily lives, with applications ranging from

virtual assistants and recommendation systems to lan-

guage generation. As AI systems, particularly gen-

erative language models, evolve and become more

sophisticated, the importance of integrating value-

awareness into their design and functionality becomes

paramount. Value-awareness refers to the incorpo-

ration of ethical considerations, cultural sensitivity,

and a deep understanding of human values into AI

systems. In the context of generative AI, which in-

volves the creation of text, the implications of value-

awareness are far-reaching and crucial for responsible

and ethical AI development.

Generative AI, particularly in language models

like OpenAI’s GPT-3, has demonstrated remarkable

capabilities in understanding and generating text. Dif-

https://orcid.org/0000-0002-9909-4400

https://orcid.org/0000-0001-9931-156X

https://orcid.org/0000-0001-6301-0028

https://orcid.org/0000-0001-5207-7745

https://orcid.org/0000-0003-3323-7357

fusion Models (DMs) for computer vision are trained

to iteratively remove noise from an image that was

blurred. The trained models can then be used for

multiple applications including image generation, im-

age manipulation (adding or removing elements from

the image, changing the lighting, etc.) and image

improvement – denoising, increasing the resolution

(Croitoru et al., 2023; Liuand et al., 2022). Access

to these models is increasing constantly, and more

people can use them for a wide range of tasks: from

improving a rushed selﬁe to generating graphics for

slides and publications, from new kinds of digital art,

to very convincing images to show side by side a

fake news article generated using a Large Language

Model.

Given the strong impact on humans’ lives, the role

of generative AI in decision-making processes has

prompted increased scrutiny regarding the alignment

of these systems with human values. While generative

AI has the potential to acquire additional knowledge

from the real world, incorporating human value judg-

ments into these systems remains a signiﬁcant chal-

lenge, and there are few speciﬁc strategies identiﬁed

for addressing this issue (Hendrycks et al., 2021). To

effectively collaborate with people and navigate hu-

1436

Ciupinska, K., Marchesi, S., Abbo, G., Belpaeme, T. and Wykowska, A.

AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures.

DOI: 10.5220/0012596400003636

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

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 1436-1443

ISBN: 978-989-758-680-4; ISSN: 2184-433X

man environments, AI systems must have the abil-

ity to understand, interpret, and predict human de-

cisions. Many human decisions involve moral con-

siderations such as concern for harm, justice, fairness

(Turiel, 1983), or broader issues of interdependent ra-

tional choice (Hayakawa, 2000), reﬂecting the intri-

cate interplay between individual ethical frameworks

and societal values that shape the complex landscape

of human behavior.

The adaptive capacity of the human moral mind

allows individuals to cooperate for mutual beneﬁt in

changing circumstances and emerging opportunities

to both help and harm (Tomasello and Vaish, 2013).

However, predicting human moral judgment poses a

formidable challenge for AI systems, especially when

it comes to responding sensibly to novel situations

that do not arise during training (Hendrycks et al.,

2021). This ﬂexibility, central to human moral cog-

nition, poses a particularly complex challenge for AI

systems.

As we have already mentioned, the ethical impli-

cations of algorithmic decision-making and the syn-

chronization of artiﬁcial intelligence systems with hu-

man values have become major focal points in the

ﬁeld of AI. Therefore, in recent years, several at-

tempts have been made to create moral AI systems

(Charisi et al., 2017; Gonzalez Fabre et al., 2021).

These methodologies aim to effectively communicate

moral and ethical values in a way that both machines

and humans can understand. Therefore, drawing in-

spiration from the VALUENET dataset, a comprehen-

sive repository of human-driven dialogues centering

on values (Qiu et al., 2022), we embark on a jour-

ney to visually articulate complex ethical and moral

concepts. The chosen value-related words, meticu-

lously curated from this dataset, form the foundation

for generating a diverse array of images through DMs.

By employing DMs, we can systematically generate

visual stimuli that encapsulate a spectrum of ethical

and moral values. This process is vital for evaluating

the ﬁdelity of generative AI systems in reﬂecting the

nuances of human values.

The ﬁrst goal of using DMs in our research was

to assess how well AI systems represent human val-

ues and related words. Through a careful selection

of input data and ﬁne-tuning of model parameters,

we were able to generate diverse visual stimuli that

encapsulate a wide range of ethical considerations.

These stimuli were then presented to human subjects

for evaluation, allowing us to gauge the alignment be-

tween the generated content and human values. This

iterative process provides valuable insights into the

strengths and limitations of generative AI systems in

capturing the complexity of ethical and moral frame-

works.

Building on the success of DMs in generating

ethically relevant visual stimuli, the second goal in-

volves leveraging these stimuli to study the neural

correlates of becoming aware of a value. The “aha

moment” or insight research, which focuses on sud-

den realizations and shifts in awareness (Sprugnoli

et al., 2017), can be applied to the domain of ethical

decision-making. By carefully selecting stimuli that

evoke ethical insights, we can further conduct EEG

experiments to identify the neural processes associ-

ated with the recognition and internalization of val-

ues.

Therefore, the implications of our research extend

to the realm of human-AI interactions and the devel-

opment of value-aware AI architectures. By analyz-

ing the effectiveness of visual stimuli generated by

DMs, we aim to contribute to the understanding of

an AI ethical framework that responds to human per-

spectives and sensitivities. This work focuses on de-

veloping three important points:

1. Advancements in Value Representation This

research is expected to contribute to the understand-

ing and development of methods for representing

moral and ethical values in AI systems. The com-

bination of linguistic and visual stimuli generated by

DMs provides a comprehensive approach to encap-

sulate the diverse range of human values and related

keywords.

2. Validation of Representations The image

recognition experiment serves as a critical validation

step, ensuring that the generated visual stimuli appro-

priately convey the intended values. This step is vital

for the practical application of AI systems in contexts

where ethical and moral decision-making is essential.

3. Neural Correlates of Becoming Aware of

Value Such AI-generated stimuli will serve as tools

that allow for systematically exploring the neural un-

derpinnings of understanding ethical and moral val-

ues. The conclusions obtained from this research

will not only contribute to understanding how ethics

and moral-related values are processed at the neural

level but also will be used to develop behavioral and

neuroscience (EEG) studies to assess the success of

value-aware AI systems in increasing perceived moral

agency.

2 METHODOLOGY

2.1 Participants

In order to test how well people will recognize values

represented by stimuli generated using DMs, a total

AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures

1437

of 36 participants (12 for each of 3 versions, see Sec-

tion 2.2.3) completed a pilot behavioral experiment

via Proliﬁc (20 males, age 20-54, M = 33.33, SD =

9.69). All participants gave informed consent prior to

the experimental session and were compensated for

their time (9 EUR per session).

2.2 Materials and Experimental

Paradigm

2.2.1 Selection of Values

The selection of values was based on the VALUENET

dataset (Qiu et al., 2022), a dataset designed for hu-

man value-driven dialogue systems. The dataset en-

compasses a wide range of moral and ethical val-

ues expressed through human conversations. From

this dataset, a list of words representing key values

was curated, forming the basis for generating stim-

uli through DMs. Our glossary contains all the words

that were listed in Figure 1 of Qiu and collaborators

(2022). To keep our experiment as consistent as pos-

sible, all words were transformed to their noun form

(i.e. if a word was an adjective in the original text,

we changed it to a noun). The study was conducted

on people whose native language is Italian. The word

list was translated by people whose native language is

also Italian and English is their second language.

2.2.2 Generation of Visual Stimuli Using DMs

The images are generated using Stable Diffusion

XL (Podell et al., 2023), a text-to-image model.

In particular, the setup comprises a base model,

stable-diffusion-xl-base-1.0, which takes a

text prompt and produces an intermediary out-

put. This is then elaborated by a reﬁner model,

stable-diffusion-xl-refiner-1.0, to obtain the

ﬁnal image. In addition, the model allows to spec-

ify negative prompts, containing a description of what

should be avoided in the picture.

After several tests, the ﬁnal prompt consists of two

parts: a noun (the name of the value) and a deﬁni-

tion taken from the Oxford Dictionary. The choice of

including the deﬁnition was dictated by two reasons:

to help disambiguate those terms that have multiple

meanings, and to add context to the model prompt. In-

deed, text-to-image models work best when provided

with a clear description of the desired image. In this

case, adding the deﬁnition of the terms gives higher

quality results compared to using only one word. An

example of a prompt is the following: “Wisdom: ca-

pacity of judging rightly in matters relating to life

and conduct; soundness of judgement in the choice

of means and ends.”

However, a test run over a subset of the values re-

vealed three main issues with the images. First, this

method produced images containing text, and (sec-

ondly) the style of the images resembled the etch-

ings commonly found in books. One explanation for

both of these issues is the formal register of the dic-

tionary descriptions, which often are associated with

these image features. The third problem is that, in

some cases, the images produced were chaotic and

unclear, whereas for the scope of this evaluation it is

best if the images have a single subject. For these

reasons, a negative prompt was added, which works

similarly to a normal prompt, except that it speciﬁes

what is not desired in the output image. The nega-

tive prompt contained two terms to reduce each of the

three problems, for a total of six words separated by a

comma. We obtained good results with the following

negative prompt: “text, letters, drawing, scan, com-

plex, chaos”.

The implementation of a negative prompt facili-

tated the generation of stimuli wherein each image

encapsulates a distinct scene as opposed to a compos-

ite amalgamation of diverse elements. The employ-

ment of this negative prompt was principally guided

by a strategic consideration for subsequent utilization

of these stimuli in EEG testing. The rationale under-

lying the adoption of stimuli featuring solitary scenes

stems from the endeavor to minimize potential con-

founds during EEG measurements. The incorporation

of intricate and disorderly visual stimuli can introduce

extraneous noise in the recorded signal, thus necessi-

tating the adoption of simpliﬁed and focused visual

stimuli to enhance signal ﬁdelity and interoperability.

For each of the values, three images were gener-

ated using 50 inference steps with an 80% ratio (40 for

the base model and 10 for the reﬁner), the guidance

scale used was 7, which is a commonly used value

to obtain quality results that do not stray away from

the prompt. Figure 1 shows an example of the im-

ages generated for the term wisdom, using the prompt

reported above

Figure 1: The three images generated for the value Wisdom.

All stimuli generated for this study are available at

https://doi.org/10.5281/zenodo.10516944.

AWAI 2024 - Special Session on AI with Awareness Inside

1438

2.2.3 Experimental Design

To validate the effectiveness of the generated stim-

uli in representing values, an image recognition ex-

periment was conducted. Our experimental proce-

dure was based on paradigms examining the Aha! ef-

fect (Sprugnoli et al., 2017). We decided to base our

paradigm on procedures examining the insight mo-

ment because we considered it an appropriate mea-

sure that would allow us to ﬁnd neural markers of be-

coming aware of a value in further EEG experiments.

To check whether DMs can adequately generate

images representing values and to select a set of ap-

propriate stimuli for further EEG experiments (de-

scribed in Section 4), a pilot behavioral experiment

was conducted. The experimental procedure was sim-

ilar to that adopted by previous studies (Zhao et al.,

2014; Sandk

uhler and Bhattacharya, 2008; Sheth

et al., 2009) (see in Figure 2).

The experimental procedure was written in Psy-

choPy version 03.02.2021 (Peirce, 2007). Since our

stimuli set contained visual representations of 162

values, we decided to split it into 3 smaller sets (54

stimuli in each). Stimuli for each set were selected

randomly. Each participant was assigned to one of the

three versions also randomly. At the beginning of the

experiment, participants were provided with training

consisting of 5 stimuli. On each trial, a white ﬁxa-

tion cross (10 x 10 pixels) was displayed centrally on

a gray background for a random duration between 2

and 3 s. Next, a set of 3 images was presented on the

screen for 60 s (coordinates were: -300, 0; 0, 0; 300,

0 pixels). Participants were required to come up with

an appropriate word that was represented by the stim-

uli. During this initial period, the participants were

instructed to press the space bar if they came up with

the solution before the time was up. Next, partici-

pants were asked to enter the word or leave the ﬁeld

blank if they could not come up with any idea. After

that, they were asked to provide the following sub-

jective ratings: (1) the Rating of Suddenness: the de-

gree of suddenness of the emergence of the answer

ranging from 1 (the task was solved step by step) to

4 (the answer came very sudden). (2) the Rating of

Figure 2: The ﬂow map of the experiment.

Conﬁdence: the degree of conﬁdence the participants

felt about the answer they reported before, which also

ranged from 1 (no conﬁdence) to 4 (full of conﬁ-

dence). (3) the Rating of Restructuring: the restruc-

turing process was described as rejecting the original

meaning of a word and reinterpreting it in a new way,

ranging from 1 (no restructuring) to 4 (full restructur-

ing). After these subjective ratings, the correct answer

was presented on the screen for 3 s. Next, partici-

pants were instructed to report their insight or “aha”

feeling described as the moment when they suddenly

understood or discovered something that was previ-

ously unclear or difﬁcult to understand. In the case of

this task, it means that they suddenly understood the

correct answer because they didn’t know/understand

it before, ranging from 0 (no insight feeling) to 1

(having insight feeling). Finally, they were asked to

indicate on 5-point Likert scale (Likert, 1932) how

strongly they agree that a given set of pictures rep-

resents a given word properly (from 1 - strongly dis-

agree to 5 - strongly agree). Before each rating and

the presentation of a correct answer, a blank was pre-

sented for a random duration between 500 and 1000

ms. The whole experimental session lasted approxi-

mately 1 hour.

Accuracy and the Likert scale (as an objective

and subjective measure, respectively) were used to

check whether DMs correctly represent human val-

ues (i.e., how accurately people recognize the value

represented and how they evaluate the appropriate-

ness of the images’ representation of this value). The

“Aha Rating” was used as a measure that allows the

selection of appropriate stimuli for the EEG experi-

ment, because in experiments examining insight, the

“Aha” needs to be a sudden awareness or an unex-

pected comprehension of an answer to a question, not

obvious before (Shen et al., 2013). The suddenness,

conﬁdence, and restructuring ratings were used as ad-

ditional measures to identify the classical features of

these insight problems (Zhao et al., 2014).

2.3 Data Analysis

Data analysis was conducted using R Studio version

2022.07.1. To check how well a given set of stimuli

represents value, we analyzed accuracy and rating on

the Likert scale. Accuracy was calculated based on

participants’ responses i.e., when a word indicated by

participants matched the word of a value the answer

was correct (1), otherwise (participant’s answer did

not match the word or no answer was provided) the

answer was incorrect (0). For subjective rating, we

dichotomized the 1-5 Likert scale scores as follows:

1, 2, and 3 ratings indicated low agreement that a set

AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures

1439

of stimuli represents a given value correctly (0), while

4 and 5 indicated high agreement (1).

The feeling of “Aha moment” indicated directly

by participants after seeing the correct answer was a

direct measure of the insight. Ratings equal to 0 in-

dicated no insight feeling, while answers equal to 1

showed having insight feeling. Additionally, previ-

ous research (Zhao et al., 2014; Sandk

uhler and Bhat-

tacharya, 2008; Sheth et al., 2009) deﬁned insight as a

solution accompanied by feelings of high suddenness,

high conﬁdence, and high restructuring. Based on

this operationalization, to assess the proportions of in-

sightful and non-insightful solutions we dichotomized

the 1-4 scale scores on each component, as follows: 1

and 2 ratings indicated low suddenness, conﬁdence

and restructuring (0), while 3 and 4 indicated high

scores (1). Thus, stimuli for which participants in-

dicated feelings of insight (i.e., average rating higher

than 0.5), and additionally high suddenness (i.e., aver-

age rating higher than 2), high conﬁdence (i.e., aver-

age rating higher than 2), and high restructuring (i.e.,

average rating higher than 2) will be selected as a set

for the further EEG experiment. Noninsightful stim-

uli were indicated by any other combinations of the

components’ levels, e.g., no insight feeling, low sud-

denness, low conﬁdence, and/or low restructuring.

3 RESULTS

3.1 Validation of the Representation of

the Value by Generated Stimuli

Data showed that the average objective accuracy rate

reached to 15% (SD = 0.16), and only 17 out of 162

(10%) stimuli sets were correctly recognized in more

than 50% of trials (M = 0.63, SD = 0.12). For the

subjective rating measured by the Likert scale, there

were 85 (52%) stimuli sets in which the rating values

were on average larger than 3 (M = 3.67, SD = 0.47;

meaning that participants agreed that a given set of

images accurately represents a value). See Table 1 for

the summary of the results.

Table 1: Statistics for assessing how well generated visual

stimuli sets represent values.

M SD

Accuracy (0/1) 0.63 0.12

Likert’s rating (1-5) 3.67 0.47

Table 2: Statistics of Aha! Ratings.

Insight No Insight

M SD M SD

Insight feeling (0/1) 0.69 0.11 0.34 0.07

Suddenness (1-4) 2.69 0.27 2.38 0.39

Conﬁdence (1-4) 2.59 0.24 2.49 0.49

Restructuring (1-4) 2.69 0.26 2.03 0.22

3.2 Validation of the Insight Feeling

Our data indicated that 76 stimuli sets (47%) were ac-

companied by the feeling of insight (M = 0.69, SD

= 0.11), and additionally high suddenness (M = 2.69,

SD = 0.27), high conﬁdence (M = 2.59, SD = 0.24),

high restructuring (M = 2.69, SD = 0.26). The re-

maining 86 stimuli sets (53%) were not accompanied

by the feeling of insight (M = 0.34, SD = 0.07). How-

ever, they were also accompanied by high suddenness

(M = 2.38, SD = 0.39), conﬁdence (M = 2.49, SD =

0.49), and restructuring (M = 2.03, SD = 0.22). See

Table 2 for the summary of the results and Figure 3

for results visualization.

The list of 76 values that will be used for the

EEG experiment looks like this: accomplishment,

antiquity, authority, beauty, brilliantness, challenge,

charity, Christian, classic, cleanliness, comfort, com-

munication, compassion, compatibility, completion,

conscience, courage, creation, devoutness, discipline,

divinity, equality, eternity, excitement, exercise, ex-

ploration, faithfulness, force, forgiveness, formality,

friendship, fun, generosity, gentleness, guard, health,

helpfulness, humanity, indulgence, intelligence, in-

tensity, interests, Islam, kindness, leadership, limit-

lessness, loyalty, manner, mercy, norms, order, or-

thodoxy, parents, participation, passion, peace, pious-

ness, principle, production, protection, regulation, re-

lax, republicanism, rich, rights, safekeeping, satisfac-

tion, self-reliance, sociality, sovereignty, spirituality,

strictness, support, unity, wisdom, work.

4 DISCUSSION

In this study, we based our experimental paradigm on

traditional ‘Aha! effect’ research and used diffusion

models to create visual stimuli representing words re-

lated to human values. The work presented focused

on two primary goals: evaluating how well generative

AI systems represent such moral and ethical-related

words and selecting stimuli suitable for studying the

neural correlates of becoming aware of a value.

AWAI 2024 - Special Session on AI with Awareness Inside

1440

Figure 3: Comparison of means on ratings measuring in-

sight feeling for stimuli accompanied by the feeling of Aha!

and not.

4.1 Alignment with Human Values

The results of the behavioral experiment shed light

on the effectiveness of DM-generated visual stimuli

in representing human values. Results showed that

our prompt did not generate stimuli that represent val-

ues that can be objectively recognized by participants.

However, the subjective measure of compliance (i.e.,

Likert scale) provides a comprehensive view, indi-

cating a noteworthy degree of accuracy in conveying

ethical and moral values by visual stimuli. While it

is clear that there is much work to be done, positive

compliance rates (measured by Likert scale) suggest

that DMs can effectively capture and communicate a

diverse range of values. The low accuracy of recog-

nizing a given value based on generated stimuli may

be the result of linguistic limitations, such as the exis-

tence of synonyms. The human language is very com-

plicated and we can often ﬁnd several words (Turney,

2008) for one expression. Therefore, it may be that

our subjects, even if they understood what concept

was represented by the series of DMs-generated stim-

uli, used a synonym for the correct word. Therefore,

one of our future plans will also be to check whether

subjects actually used synonyms to describe a given

set of stimuli. It should also be noted that the values

found in the glossary refer to abstract concepts that

are difﬁcult to grasp and visualize, and therefore it

may be difﬁcult to label them correctly.

It is crucial to highlight that a fundamental draw-

back exists in these models due to the initial bias

present in their training sets. The representation of

different values may still be somewhat homogeneous

in Western society. As noted by (Henrich et al., 2010)

a signiﬁcant portion of research in human psychology

operates under the assumption that fundamental cog-

nitive processes are universally shared, and ﬁndings

from one population can be universally applied. As

the authors note, most of the research results and as-

sumptions are based on the Western population.

In sum, behavioral ﬁndings suggest that the inabil-

ity to generate stimuli that participants could objec-

tively recognize as values prompts a critical reﬂection

on the methodology and the underlying assumptions.

It suggests that the chosen method or prompt might

not effectively capture or convey the nuanced and sub-

jective nature of human values. This recognition is

crucial for reﬁning future experimental designs and

prompts, taking into account the complex and mul-

tifaceted nature of values. While the negative result

may alter the anticipated trajectory of the study, it cat-

alyzes further exploration, reﬁnement, and adaptation

of methodologies.

Beyond their role in eliciting the ”Aha moment”

during EEG experiments, these images hold poten-

tial in realistic scenarios. They can serve as educa-

tional tools, enriching learning materials and presen-

tations to foster a more profound understanding of ab-

stract values. Moreover, the validated stimuli, espe-

cially those associated with the insight feeling, may

ﬁnd utility in psychological and therapeutic settings,

aiding introspection and discussions on personal val-

ues. In the realm of design and creativity, these im-

ages may inspire artistic endeavors, offering a visual

language to explore and communicate complex soci-

etal themes. However, the variability in participant re-

sponses underscores the need for a nuanced approach

in future applications, acknowledging diverse inter-

pretations to maximize impact across different con-

texts.

4.2 Potential for Value-Aware AI

Development

As we described earlier, in times of extraordinary

development of AI systems, their proper interaction

with human agents is extremely important. Our

study’s ﬁndings offer practical implications for the

ethical development of AI systems, as the ability of

generative AI to create visual stimuli that align with

human values is a crucial step towards responsible

AI. Firstly, by ensuring that AI-generated content re-

ﬂects human values, we mitigate the risk of uninten-

tionally perpetuating biases or promoting content that

may be ethically questionable (Chimbga, 2023). Sec-

ondly, value-aligned visual stimuli can enhance user

trust and acceptance of AI applications. Users are

more likely to engage positively with technology that

resonates with their values, fostering a sense of relia-

bility and ethical responsibility (Ma and Huo, 2023).

AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures

1441

Thirdly, in scenarios where AI systems interact with

users, generating content in line with human values

contributes to more ethical and respectful interactions

(Dignum, 2017).

Regardless of the application, AI must consider

societal values, ethical concerns, and moral consid-

erations (Dignum, 2017). This is mainly because AI

systems are tools under the control of human users,

their potential autonomy and learning capabilities ne-

cessitate a deliberate incorporation of accountability,

responsibility, and transparency principles in the de-

sign process (Charisi et al., 2017). We propose that

AI development should prioritize the consideration

of human values in AI systems because emphasizing

value-aware AI performance can lead to the explo-

ration of innovative techniques and applications.

The success and challenges observed in convey-

ing values-related words through visual stimuli em-

phasize the need for tailored approaches in designing

awareness architectures. Future architectures should

prioritize adaptability to individual cognitive pro-

cesses, acknowledging the nuanced ways individu-

als interpret and resonate with visual representations.

Incorporating mechanisms that account for the vari-

ability in cognitive responses can enhance the effec-

tiveness of awareness-building initiatives. Integrating

such insights into the design of awareness architec-

tures can potentiate their impact, fostering a more pro-

found and meaningful understanding of values in di-

verse contexts.

4.3 Insightful Selection for EEG Study

The identiﬁcation of values that evoke an “Aha!” mo-

ment or insight feeling, accompanied by high sud-

denness, conﬁdence, and restructuring, serves as a

pivotal aspect of this study. This insightful selec-

tion criterion serves as a deliberate and strategic foun-

dation for the subsequent EEG experiment, ensur-

ing that the neuroscientiﬁc investigation delves into

values-related words that evoke profound cognitive

processes. Speciﬁcally, the sudden recognition of

moral or ethical-related value, as indicated by the par-

ticipants’ heightened experiences of insight, becomes

a focal point for understanding the neural underpin-

nings of value perception. Thanks to this we will de-

velop experiments that examine how people perceive

AI systems as moral and intentional agents. Such an

approach will enable us to make an important step to-

ward the neuroscientiﬁc assessment of the success of

Value-Aware AI systems in increasing the perceived

moral and ethical agency.

5 CONCLUSION

In conclusion, this study represents a signiﬁcant step

forward in the convergence of generative AI, ethi-

cal considerations, and neuroscience. The integra-

tion of diffusion models as a means to generate vi-

sual stimuli for representing ethical and moral val-

ues not only demonstrates the feasibility of aligning

AI systems with human values but also opens up av-

enues for exploring the neural underpinnings of value

awareness. As the ﬁeld advances, the interplay be-

tween AI and human values, as studied through the

lens of DMs, contributes not only to a deeper under-

standing of the cognitive processes involved in ethical

decision-making but also will help us to test whether

and to what extent people perceive AI systems as ca-

pable of understanding and making moral and ethi-

cal values. This multidisciplinary approach positions

our study at the forefront of ethical AI development,

offering valuable insights for researchers, developers,

and policymakers navigating the complex intersection

of technology and human values.

ACKNOWLEDGEMENTS

This work has received support from the Euro-

pean Union under the European Innovation Council

(EIC) research and innovation programme, Project

VALAWAI, Grant Agreement number 101070930.

REFERENCES

Charisi, V., Dennis, L., Fisher, M., Lieck, R., Matthias,

A., Slavkovik, M., ..., and Yampolskiy, R. (2017).

Towards moral autonomous systems. arXiv preprint

arXiv:1703.04741.

Chimbga, B. (2023). Exploring the Ethical and Societal

Concerns of Generative AI in Internet of Things (IoT)

Environments. In Southern African Conference for

Artiﬁcial Intelligence Research (pp. 44-56). Cham:

Springer Nature Switzerland.

Croitoru, F. A., Hondru, V., Ionescu, R. T., and Shah, M.

(2023). Diffusion models in vision: A survey. IEEE

Transactions on Pattern Analysis and Machine Intelli-

gence.

Dignum, V. (2017). Responsible artiﬁcial intelligence: de-

signing AI for human values. ITU Journal: ICT Dis-

coveries, Special Issue No. 1.

Gonzalez Fabre, R., Camacho Ib

nez, J., and Tejedor Es-

cobar, P. (2021). Moral control and ownership in AI

systems. AI & SOCIETY, 36, 289-303.

Hayakawa, H. (2000). Bounded rationality, social and

cultural norms, and interdependence via reference

AWAI 2024 - Special Session on AI with Awareness Inside

1442

groups. Journal of Economic Behavior & Organiza-

tion, 43(1), 1-34.

Hendrycks, D., Carlini, N., Schulman, J., and Steinhardt,

J. (2021). Unsolved problems in ml safety. arXiv

preprint arXiv:2109.13916.

Henrich, J., Heine, S. J., and Norenzayan, A. (2010). Most

people are not WEIRD. Nature, 466(7302), 29-29.

Likert, R. (1932). A technique for the measurement of atti-

tudes. Archives of psychology.

Liuand, N., Liand, S., Duand, Y., Torralbaand, A., and

Tenenbaumand, J. B. (2022). Compositional visual

generation with composable diffusion models. Euro-

pean Conference on Computer Vision (pp. 423-439).

Ma, X. and Huo, Y. (2023). Are users willing to embrace

ChatGPT? Exploring the factors on the acceptance of

chatbots from the perspective of AIDUA framework.

Technology in Society, 75, 102362.

Peirce, J. W. (2007). PsychoPy—psychophysics software

in Python. Journal of neuroscience methods, 162(1-

2), 8-13.

Podell, D., English, Z., Lacey, K., Blattmann, A., Dock-

horn, T., M

uller, J., Penna, J., and Rombach, R.

(2023). SDXL: Improving Latent Diffusion Models

for High-Resolution Image Synthesis.

Qiu, L., Zhao, Y., Li, J., Lu, P., Peng, B., Gao, J., and Zhu,

S. C. (2022). Valuenet: A new dataset for human value

driven dialogue system. In Proceedings of the AAAI

Conference on Artiﬁcial Intelligence (Vol. 36, No. 10,

pp. 11183-11191).

Sandk

uhler, S. and Bhattacharya, J. (2008). Deconstructing

insight: EEG correlates of insightful problem solving.

PLoS one, 3(1), e1459.

Shen, W., Liu, C., Yuan, Y., Zhang, X., and Luo, J. (2013).

Temporal dynamics of mental impasses underlying

insight-like problem solving. Science China Life Sci-

ences, 56, 284-290.

Sheth, B. R., Sandk

uhler, S., and Bhattacharya, J. (2009).

Posterior beta and anterior gamma oscillations predict

cognitive insight. Journal of cognitive neuroscience,

21(7), 1269-1279.

Sprugnoli, G., Rossi, S., Emmendorfer, A., Rossi, A., Liew,

S. L., Tatti, E., ..., and Santarnecchi, E. (2017). Neural

correlates of Eureka moment. Intelligence, 62, 99-

118.

Tomasello, M. and Vaish, A. (2013). Origins of human co-

operation and morality. Annual review of psychology,

64, 231-255.

Turiel, E. (1983). The development of social knowl-

edge: Morality and convention. Cambridge Univer-

sity Press.

Turney, P. D. (2008). A uniform approach to analogies,

synonyms, antonyms, and associations. arXiv preprint

arXiv:0809.0124.

Zhao, Q., Li, Y., Shang, X., Zhou, Z., and Han, L. (2014).

Uniformity and nonuniformity of neural activities cor-

related to different insight problem solving. Neuro-

science, 270, 203-211.

AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures

1443