upload_otlcngnrsxjjhui6zuasjfa355.pdf

72

AI

-

ASSISTED DECISION

-

MAKING IN PUBLIC SECURITY

Opportunities, Vulnerabilities and Professional Training Requirements

Colonel (ret) Professor George-Marius ȚICAL, Ph.D



(Academy of Romanian Scientists, 3 Ilfov, 050044, Bucharest, Romania,

email: secretariat@aosr.ro)

Abstract:

The

integration

of

artificial

intelligence

into

public

security

is

changing

the

relationship

between

data,

analysis

and

professional

decision

-

making.

Algorithmic

systems

can

process

large

volumes

of

information,

identify

patterns

that

are

difficult

to

observe

and

support

risk

assessment

or

resource

allocation. However, the same capabilities can produce effects contrary to legality

and

effectiveness

when

data

are

incomplete,

models

are

opaque,

results

are

affected by bias or users place excessive trust in technological recommendations.

This article analyses the opportunities and vulnerabilities of AI

-

assisted decision

-

making in public security with reference to Regulation (EU) 2024/1689, Europol

and FRA documents, and UNICRI

–

INTERPOL instruments. It argues that human

oversight becomes effective only when the operator has the competence, authority

and time required to verify and, where appropriate, reject the algorithmic output.

As an original contribution, the article proposes the C.O.N.T.R.O.L. professional

training model, focused on understanding the system, verifying data provenance,

applying the relevant rules, testing the output, assuming responsibility, observing

effects and drawing on lessons learned.

Keywords:

artificial

intelligence;

public

security;

assisted

decision

-

making; human oversight; algorithmic bias; professional training.

DOI

10.

56082/

annalsarscimilit.2026.3.7

2

1. Introduction

The

digitalisation

of

public

order

and

safety

activities

no

longer

means

merely

computerising

records

or

rapidly

transmitting

data.

The

development

of

artificial

intelligence

systems

makes

it

possible

to

move

from tools that describe the existing situation to tools that classify, estimate

probabilities, identify links and formulate recommendations. In this context,

the decision

-

maker interacts

not

only

with raw information,

but

also

with

algorithmically processed outputs, often presented as a risk score, an alert, a

prioritisation or a course of action.

Public

security

is

a

field

in

which

the

promise

of

technology

is

particularly

attractive.

Institutions

must

respond

rapidly,

manage

limited



“Andrei Șaguna” University, Associate

d

Member of the

Academy of Romanian Scientists,

Entitled

Member

of

The

Academy

of

National

Security

Sciences,

email:

ticalgeorgem@gmail.com

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

73

resources,

correlate

heterogeneous

sources

and

make

decisions

under

uncertainty. Artificial

intelligence can substantially enhance organisations'

analytical capacity. Nevertheless, decisions in this field may affect liberty,

privacy, equality, the presumption of innocence and access to an effective

means

of

challenge.

For

this

reason,

effectiveness

cannot

be

the

sole

assessment criterion.

The issue cannot be reduced to a simple opposition between human

and

machine.

In

practice,

most

relevant

applications

are

sociotechnical

systems:

data

are

collected

by

individuals

and

institutions,

the

model

is

designed according to defined objectives, the output is interpreted within an

organisational context, and the consequences are borne by individuals and

communities.

An

error

may

originate

in

the

data,

the

model's

design,

the

way

the

output

is

presented

or

the

user's

conduct.

Responsibility

must

therefore be examined throughout the system's entire life cycle.

The

research

question

is:

under

what

conditions

can

artificial

intelligence support decision

-

making in public security without undermining

legality,

impartiality,

accountability

and

fundamental

rights?

The

article

seeks to identify the principal opportunities and vulnerabilities, clarify the

meaning of effective human oversight

and formulate requirements for the

initial and continuing training of personnel.

2. Conceptual and Methodological Framework

2.1. From Human Decision

-

Making to Assisted Decision

-

Making

AI

-

assisted

decision

-

making

must

be

distinguished

both

from

exclusively

human

decision

-

making

and

from

fully

automated

decision

-

making. In the first case, the algorithm plays no part in the reasoning. In the

second,

the

system's

output

produces

consequences

without

meaningful

human intervention. Assisted decision

-

making lies between these extremes:

the system formulates a prediction, classification or recommendation, while

the competent person evaluates the output against the other information and

assumes responsibility for the decision.

This

distinction

is

normative,

not

merely

technical.

The

formal

presence

of

a

person

in

the

workflow

does

not

establish

the

existence

of

human control. If the operator does not understand the system, lacks access

to the information needed for verification, is constrained by time or lacks

the authority to reject the recommendation, the intervention becomes a mere

confirmation. In such a situation, the decision is effectively automated, even

if a person's signature appears on the document. “Human judgment should

be used when setting metrics for specific trustworthy AI characteristics.”

1

National Institute of Standards and Technology,

Artificial Intelligence Risk Management

Framework

(AI RMF 1.0), NIST AI 100-1 (Gaithersburg, MD: NIST, 2023), p. 12.

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

74

The NIST standard thus supports a substantive, rather than merely

formal, understanding of the operator's intervention.

2.2. Research Methodology

The research uses legal and institutional analysis, supplemented by

scenario methodology

and risk analysis. The principal legal

framework

is

Regulation

(EU)

2024/1689

on

artificial

intelligence.

It

is

examined

in

conjunction

with

the

Council

of

Europe

Framework

Convention

on

Artificial Intelligence and Human Rights, Democracy and the Rule of Law,

the UNESCO Recommendation on the Ethics of Artificial Intelligence, and

the NIST AI Risk Management Framework.

The

operational

dimension

is

grounded

in

Europol

reports

on

artificial

intelligence

in

policing,

algorithmic

bias

and

technology

assessment

in

law

enforcement.

Reports

issued

by

the

European

Union

Agency

for

Fundamental

Rights

are

used

for

the

fundamental

-

rights

dimension.

Training

requirements

are

considered

against

the

UNICRI

–

INTERPOL Toolkit for Responsible AI Innovation in Law Enforcement.

The analysis does not assess a specific operational application and

does not use classified data or information from actual cases. Its conclusions

are conceptual and normative and seek to construct a framework applicable

to professional training and institutional assessment.

3. Applications and Opportunities in Public Security

3.1. Data Analysis and Pattern Identification

One of the major advantages of AI is its capacity to analyse volumes

of data that personnel could not examine in full within a useful timeframe.

Systems

can

correlate

persons,

locations,

times,

objects,

transactions

and

modi

operandi,

flagging

relationships

that

require

further

verification.

Europol

shows

that

these

capabilities

can

support

intelligence

analysis,

digital

investigations,

fraud

detection,

image

recognition

and

the

management

of

digital

evidence.

“Facial

recognition

technology

(FRT)

makes

it

possible

to

compare

digital

facial

images

in

order

to

determine

whether they belong to the same person.”

2

3

The

probabilistic

nature

of

the

match

nevertheless

requires

confirmation through independent information.

Operational value does not lie in turning correlation into certainty,

but in directing the analyst's attention. A useful system narrows the search

space

and

proposes

verifiable

hypotheses.

It

becomes

dangerous

when

2

Europol,

AI

Bias

in

Law

Enforcement:

A

Practical

Guide

(Luxembourg:

Publications

Office of the European Union, 2025), p. 8.

3

European

Union

Agency

for

Fundamental

Rights,

Facial

Recognition

Technology:

Fundamental

Rights

Considerations

in

the

Context

of

Law

Enforcement

(Luxembourg:

Publications Office of the European Union, 2019), p. 1.

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

75

correlation is treated as evidence or when a probability score is converted

into a presumption concerning a person.

3.2. Risk Assessment and Resource Allocation

Analytical

tools

can

support

patrol

planning,

the

identification

of

high

-

incidence

periods,

risk

analysis

at

public

events

and

resource

allocation. In emergency management, AI can combine meteorological data,

geographical information, traffic flows and field reports. The advantage lies

in the capacity to update the operational picture rapidly.

A prediction concerning a place or phenomenon must, however, be

distinguished from an assessment of individual risk. Area

-

based data may

reflect

not

only

the

actual

distribution

of

crime,

but

also

the

historical

intensity of police checks. Repeatedly directing resources to the same places

generates new recorded incidents, which subsequently reinforce the initial

prediction.

This

creates

a

feedback

loop

that

may

confuse

the

level

of

policing activity with the actual level of the phenomenon.

3.3. Reducing Repetitive Tasks

AI

can

be

used

for

transcription,

semantic

search,

document

classification, preliminary translation, anonymisation or the organisation of

materials. The time saved may allow personnel to focus on activities that

require professional judgment, empathy and human interaction. This benefit

is

real

only

if

generated

outputs

are

verified

and

automation

neither

introduces

non

-

existent

information

nor

alters

the

meaning

of

documents.

“Additionally,

AI

systems

can

automate

the

analysis

of

digital

evidence,

enabling investigators to quickly identify relevant clues from emails, online

conversations,

and

forensic

databases.”

(“Additionally,

AI

systems

can

automate the analysis of digital evidence, enabling investigators to quickly

identify

relevant

clues

from

emails,

online

conversations,

and

forensic

databases.”)

4

3.4. Training and Scenario Simulation

In

professional

training,

generative

systems

can

construct

dynamic

scenarios, adapt difficulty to the trainee's level and simulate the evolution of

operational

situations.

Their

use

can

develop

the

capacity

to

anticipate,

compare alternatives and provide reasons for a decision. In this context, AI

is

valuable

not

because

it

provides

the

“correct”

answer,

but

because

it

enables

a

wider

range

of

learning

situations

and

an

examination

of

the

consequences of different choices.

4

Țical

G.M.,

“

Artificial

Intelligence

and

Big

Data

Analysis

in

Crime

Prevention

and

Combat

”,

Annals

Series

on

Military

Sciences

17,

no.

1

(2025),

p.

48,

https://doi.org/10.56082/annalsarscimilit.2025.1.36.

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

76

Table 1. Relationship Between Opportunity, Vulnerability and Required Competence

Source: author's own elaboration based on Europol (2024, 2025), FRA (2022) and

UNICRI

–

INTERPOL (2024)

4. Vulnerabilities and Effects on Decision

-

Making

4.1. Data Quality

An artificial intelligence system cannot transcend the limitations of

the

information

on

which

it

was

built

and

the

data

supplied

during

use.

Information

may

be

incomplete,

outdated,

incorrectly

labelled

or

unrepresentative.

FRA

emphasises

that

poor

-

quality

data

may

produce

outputs

that

affect

the

right

to

non

-

discrimination

and

other

fundamental

rights.

In

public

security,

this

vulnerability

is

amplified

by

the

heterogeneous nature of the sources and by time pressure.

5

Historical

data

are

not

a

neutral

snapshot

of

reality.

They

reflect

institutional priorities, reporting practices and the distribution of checks. A

database of interventions shows where agencies intervened, not necessarily

where

all

offences

occurred.

If

this

distinction

is

not

understood,

the

algorithm may legitimise earlier practices and project them into the future.

4.2. Algorithmic Bias and Discrimination

Bias can arise at every stage of the life cycle: problem formulation,

data

selection,

definition

of

the

intended

outcome,

training,

validation,

deployment and interpretation. The Europol guide shows that historical and

representational biases may lead to unfair outcomes, while improper use and

excessive reliance on an output can turn a technical problem into a harmful

5

European Union Agency for Fundamental Rights,

Bias in Algorithms: Artificial

Intelligence and Discrimination

(Luxembourg: Publications Office of the European Union,

2022), pp. 12–15.

Opportunity

Associated vulnerability

Professional competence

Rapid data analysis

False correlations or

unrepresentative data

Verification of data source,

quality and relevance

Risk assessment

Turning probability into certainty

Probabilistic reasoning and

testing alternative hypotheses

Image recognition

False

-

positive identifications and

differences in accuracy

Independent confirmation and

understanding error rates

Document automation

Fabricated information,

omissions or loss of traceability

Editorial control, source citation

and responsibility for content

Resource allocation

Feedback loops and over

-

policing of certain communities

Impact assessment,

proportionality and continuous

monitoring

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

77

operational

action.

“AI

bias

can

occur

at

any

stage

of

the

system

life

cycle”

6

7

Verification therefore cannot be confined to the time of procurement

or initial validation.

Fairness

cannot

be

reduced

to

removing

sensitive

characteristics

from the model. Other variables may act as proxies for ethnic origin, social

status

or

place

of

residence.

Assessment

must

therefore

examine

the

system's

effects

on

different

groups

and

be

repeated

after

deployment.

“Biases in algorithmic systems can lead to discrimination”

8

The

FRA

formulation

warrants

a

legal

examination

of

effects,

distinct from the mere technical measurement of error.

4.3. Opacity and the Problem of Explainability

When presented with a score or recommendation, the user must be

able to understand at least the system's purpose, the relevant data, its known

limitations

and

the

meaning

of

the

output.

Explainability

does

not

necessarily require every user to have access to the source code; it requires

the

provision

of

the

information

needed

for

prudent

use

and

for

giving

reasons for the decision. An output that cannot be explained is difficult to

verify, challenge and defend in legal terms.

4.4. Automation Bias

Automation bias refers to a person's tendency to accept the system's

recommendation and disregard contradictory information. The risk is high

when technology is presented as objective, when the interface uses precise

scores

or

when

the

organisation

rewards

conformity

with

the

recommendation. Under such conditions, the expression “assisted decision

-

making” may conceal the effective transfer of reasoning to the system.

This phenomenon cannot be prevented by the generic statement that

“the final decision belongs to the human”. The user must know the tool's

limitations,

receive

information

about

uncertainty,

be

able

to

consult

the

sources

and

be

encouraged

to

formulate

alternative

hypotheses.

Reasoned

disagreement

with

the

system

should

be

treated

as

an

expression

of

competence, not as a departure from the technological workflow.

4.5. Cybersecurity and Technological Dependence

AI

systems

may

be

compromised

through

data

manipulation,

adversarial attacks, unauthorised access, supply

-

chain compromise or model

alteration.

In

public

security,

the

consequences

may

include

misdirecting

6

Europol,

AI Bias in Law Enforcement: A Practical Guide

(Luxembourg: Publications

Office of the European Union, 2025), pp. 9–11.

7

Europol

, AI Bias in Law Enforcement

, p. 9.

8

European Union Agency for Fundamental Rights,

Bias in Algorithms

, p. 23.

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

78

resources,

disclosing

sensitive

information

and

compromising

investigations.

Assessment

must

cover

robustness,

operational

continuity,

the

traceability

of

changes

and

the

possibility

of

reverting

to

alternative

procedures.

Dependence

on

the

provider

raises

distinct

issues:

access

to

documentation,

updates,

data

location,

auditing

and

termination

of

the

contract. An institution cannot outsource responsibility for the legality of a

decision merely because the tool belongs to a private company.

5.

The

European

Legal

Framework

and

the

Requirement

of

Human Control

5.1. The Risk

-

Based Approach

Regulation

(EU)

2024/1689

establishes

a

differentiated

approach

according

to

risk:

certain

practices

are

prohibited,

high

-

risk

systems

are

subject to extensive requirements, and other systems fall under transparency

obligations. In the field

of law enforcement, Annex

III includes uses that

may

have

significant

consequences

for

individuals,

while

Article

5

sets

limits on certain forms of individual prediction and biometric identification.

“AI

should

be

a

human

-

centric

technology.

It

should

serve

as

a

tool

for

people”

9

For high

-

risk systems, risk management, data governance, technical

documentation,

record

-

keeping,

transparency,

human

oversight,

accuracy,

robustness

and

cybersecurity

are

relevant.

These

obligations

indicate

that

the technical output cannot be separated from the organisational context in

which it is used.

5.2. Effective Human Oversight

Article

14

of

the

AI

Regulation

requires

high

-

risk

systems

to

be

designed and used in such a way that they can be effectively overseen by

natural persons. The overseer must understand the system's capabilities and

limitations,

identify

anomalies,

avoid

automatic

reliance

and

be

able

to

disregard, override or reverse the output. This obligation is complemented

by

the

requirements

applicable

to

institutional

deployers

and

by

the

fundamental

-

rights impact assessment.

Effective human control can be analysed through three cumulative

conditions.

The

first

is

competence:

the

person

must

understand

how

the

system is used and the risks involved. The second is authority: the person

must

be

able

to

reject

the

output

without

unjustified

organisational

consequences. The third is genuine opportunity: the person must have the

time,

data

and

tools

needed

for

verification.

The

absence

of

any

of

these

conditions turns oversight into a formal ritual.

9

European Parliament and Council of the European Union,

Regulation (EU) 2024/1689

,

recital 6, Official Journal of the European Union L 2024/1689 (July 12, 2024), p. 2.

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

79

5.3. Accountability and the Possibility of Challenge

The

use

of

an

algorithmic

recommendation

does

not

transfer

legal

responsibility

from

the

institution

or

the

competent

decision

-

maker

to

the

system. The decision must be reasoned on the basis of elements that can be

communicated and verified. The affected person must be able to understand

the significant role of the system and challenge the output, within the limits

established

by

law

and

consistently

with

the

protection

of

operational

methods.

The

Council

of

Europe

Framework

Convention

reinforces

this

perspective

through

the

principles

of

dignity,

individual

autonomy,

transparency,

oversight,

accountability

and

non

-

discrimination.

The

UNESCO

Recommendation

adds

the

need

for

public

and

professional

literacy,

as

well

as

ethical

impact

assessment

throughout

the

life

cycle.

“…there

should

be

adequate

transparency

and

oversight

requirements

tailored to the specific contexts and risks”

10

11

12

The standard rules out the uniform application of purely declaratory

safeguards.

6. Professional Training Requirements

6.1. AI Literacy

Article 4 of the AI Regulation makes artificial intelligence literacy

an organisational obligation. In public security institutions, training cannot

be

limited

to

a

general

presentation

of

the

technology

or

to

reading

instructions for use. Its level must be adapted to the role, the type of system,

the risk and the possible consequences for individuals.

Minimum

technical

competence

includes

understanding

the

difference between probability and certainty, the role of training data, model

limitations,

error

rates,

and

false

-

positive

or

false

-

negative

outputs.

The

operator need not become a programmer, but must be able to interpret the

output correctly and know which questions to ask. All of this “…requires

certain processes, an enabling culture, and the right people and expertise to

harness its potential effectively”.

13

Individual

training

must

therefore

be

supported

by

organisational

preparedness.

10

UNESCO,

Recommendation on the Ethics of Artificial Intelligence

(Paris: UNESCO,

2021), p. 17.

11

Council of Europe

, Framework Convention on Artificial Intelligence and Human Rights,

Democracy and the Rule of Law

, CETS No. 225 (Strasbourg, 2024), pp. 4–5.

12

Council of Europe, Framework Convention, art. 8, p. 4.

13

UNICRI and INTERPOL,

Toolkit for Responsible AI Innovation in Law Enforcement:

Organizational Readiness Assessment Questionnaire

(Turin/Lyon, 2024), p. 4.

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

80

6.2. Legal and Ethical Competences

Training

must

cover

data

protection,

non

-

discrimination,

the

conditions

for

using

high

-

risk

systems,

the

statement

of

reasons

for

a

decision,

traceability

and

accountability.

The

ethical

dimension

concerns

dignity, proportionality, the avoidance of stigmatisation and the preservation

of professional autonomy. This knowledge must be applied in scenarios, not

presented exclusively in theoretical terms.

6.3. Critical Thinking and Probabilistic Reasoning

A competent user asks questions about the output: what data were

used, how current they are, what information is missing, what the error rate

is, whether alternative explanations exist and whether the decision can be

reasoned

without

merely

invoking

the

score.

Training

must

develop

the

ability to disagree with the system on the basis of verifiable arguments.

6.4. Scenario

-

Based Training and Competence Assessment

Scenarios

may

include

correct

recommendations,

erroneous

recommendations

and

ambiguous

situations.

The

trainee

is

required

to

identify the necessary information,

verify the output and give reasons

for

accepting

or

rejecting

it.

Assessment

must

measure

not

only

speed

or

agreement with the recommendation, but also the quality of reasoning, the

identification of uncertainty and respect for rights.

The

2026

INTERPOL

training

catalogue

includes

the

T.R.A.I.L.

programme

–

Training on Responsible AI Innovation in Law Enforcement

–

confirming

that

the

responsible

use

of

AI

has

become

a

distinct

field

of

professional

training.

The

UNICRI

–

INTERPOL

Toolkit

recommends

developing competences, organisational culture and assessment procedures

before systems are introduced into routine activities. It also lists “Training

on Responsible AI Innovation in Law Enforcement

–

T.R.A.I.L. (EN)

–

one

week”

14

15

16

7.

The

C.O.N.T.R.O.L.

Model

–

A

Proposal

for

Professional

Training

To

translate

the

principles

into

an

easy

-

to

-

use

instrument,

the

C.O.N.T.R.O.L. model is proposed. It can serve as the structure of a training

module,

a

pre

-

decision

analytical

framework

or

a

tool

for

assessing

a

situation in which an algorithmic recommendation plays a significant role.

14

UNICRI and INTERPOL,

Organizational Readiness Assessment Questionnaire

, pp. 4–

17.

15

INTERPOL,

INTERPOL

I-Learn

Catalogue

(May

2026),

p.

14,

available

at

https://-

www.interpol.int/content/download/24489/file/INTERPOL%20I-Learn%20Catalogue%-

20May%202026.pdf, accessed on September 15, 2026.

16

Ibidem

, p. 14.

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

81

Element

Content

Control question

C

–

Comprehension

of the system

Purpose, operation, limitations and

intended use.

What can and cannot this system

do?

O

–

Origin of the

data

Source, currency, quality and

representativeness of the data.

What information is the output

based on?

N

–

Norms

applicable

Legality, necessity, proportionality

and fundamental rights.

Is the use permitted and justified

in the specific situation?

T

–

Testing the

output

Comparison with other sources

and formulation of alternative

hypotheses.

What elements confirm or refute

the recommendation?

R

–

Responsibility

of the human

Reasoning and assumption of the

decision by the competent person.

Can I explain and defend this

decision without invoking the

authority of the algorithm?

O

–

Observation of

effects

Monitoring errors, incidents and

differentiated effects.

What consequences has use of the

system produced?

L

–

Lessons learned

Reporting, reviewing procedures

and updating training.

What must be changed for future

uses?

Table 2. The C.O.N.T.R.O.L. Model

Source: author's own elaboration

The

model

does

not

replace

institutional

procedures

or

legal

assessment. Its role is to create a cognitive routine through which the user

avoids automatically accepting a recommendation. Its application must be

adapted

to

the

type

of

system

and

the

level

of

risk.

In

situations

with

significant

effects

on

individuals,

the

answers

must

be

documented

proportionately.

8. Empirical Research Proposal

To

test

the

hypothesis

concerning

automation

bias,

a

small

-

scale

experiment may be conducted with students or trainees from public

-

order

educational institutions. Participants are assigned to two groups and receive

the same hypothetical situation. The control group analyses the data without

an algorithmic recommendation, while the experimental group additionally

receives

a

recommendation

presented

as

having

been

generated

by

an

AI

system.

The

recommendation

must

contain

a

detectable

error

or

an

insufficiently

substantiated

conclusion,

without

misleading

participants

about

real

risks. The acceptance rate, the amount

of information

verified,

identification of the error, quality of reasoning, decision time and declared

level of confidence may be measured. After the exercise, its purpose must

be explained and the mechanism of automation bias discussed.

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

82

In a second stage, participants complete a training module based on

the

C.O.N.T.R.O.L.

model

and

then

solve

a

comparable

situation.

The

difference

between

the

two

stages

may

indicate

whether

training

reduces

uncritical

acceptance

and

improves

the

statement

of

reasons.

The

results

should not be interpreted as an assessment of general professional aptitude,

but as an indicator of a cognitive vulnerability specific to human

–

algorithm

interaction.

9. Discussion

Introducing

AI

into

public

security

is

not

merely

a

technological

project, but

an institutional change. A high

-

performing system used in

an

organisation without procedures, competences and mechanisms of challenge

may

produce

poorer

results

than

a

transparent

traditional

method.

Assessment

must

include

fitness

for

purpose,

the

impact

on

rights,

data

quality, staff capacity and the possibility of monitoring. “While AI brings

undeniable

benefits

to

security,

its

use

must

be

accompanied

by

strict

measures to protect citizens' rights and prevent abuses.” (“While AI brings

undeniable

benefits

to

security,

its

use

must

be

accompanied

by

strict

measures to protect citizens' rights and prevent abuses.”)

17

There is a risk that the requirement of human control will be used to

legitimise systems without changing actual working conditions. If a person

must

validate

hundreds

of

alerts,

if

the

explanation

is

inaccessible

or

if

departure

from

a

recommendation

must

be

justified

disproportionately,

control

is

not

effective.

Institutions

must

assess

not

only

the

technical

interface, but also workload, incentives and the distribution of authority.

Professional

training

must

be

coordinated

with

procurement

and

governance. Staff vigilance cannot compensate for a system that is unfit for

purpose or built on unsuitable data. Conversely, even the best system cannot

eliminate

the

need

for

contextual

reasoning.

Responsibility

is

distributed

across

the

process,

but

must

not

be

diluted:

providers

are

responsible

for

their

obligations,

management

for

the

decision

to

adopt

the

system

and

organise control, and the user for the specific manner in which the output is

integrated into the decision.

In public

-

order education, AI must be used both as an object of study

and as a learning tool. Trainees must understand the technology they will

encounter in professional practice and must also practise its responsible use.

The

aim

is

not

to

cultivate

obedience

to

the

system,

but

to

develop

the

capacity for critical human

–

AI collaboration.

17

Țical GM,

“Artificial Intelligence and Big Data Analysis in Crime Prevention and

Combat”,

p. 50.

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

83

10. Conclusions

Artificial intelligence can improve analysis, response speed and the

use

of

resources

in

public

security.

It

can

identify

patterns,

organise

information and generate scenarios that a decision

-

maker could not produce

with the same speed. These benefits do not, however, turn an algorithmic

output into truth or justify replacing professional judgment.

Vulnerabilities

arise

from

data

quality,

bias,

opacity,

attacks,

dependence

on

the

provider

and,

above

all,

the

way

in

which

people

interpret the output. Automation bias demonstrates that the formal presence

of a human does not guarantee control. Oversight is effective only when the

person has the competence, authority and genuine opportunity to verify the

output.

Professional

training

must

integrate

technical,

legal,

ethical,

operational

and critical

-

thinking

competences. The C.O.N.T.R.O.L. model

offers

a

structure

applicable

to

training

and

practice,

directing

the

user

towards

understanding

the

system,

verifying

the

data,

applying

the

rules,

testing

the output, assuming

responsibility, observing effects

and drawing

on lessons learned.

In

public

security,

an

artificial

intelligence

system

is

useful

not

when

it

decides in place of the professional, but when it helps the professional make

a better

-

informed decision without diminishing judgment, independence or

responsibility.

BIBLIOGRAPHY

COUNCIL OF EUROPE.

Framework Convention on Artificial Intelligence

and Human Rights, Democracy and the Rule of Law

.

CETS No.

225. Strasbourg, 2024, available at

https://rm.coe.int/1680afae3c;

EUROPEAN

DATA

PROTECTION

SUPERVISOR.

TechDispatch:

Human

Oversight

of

Automated

Decision

-

Making

.

Brussels,

2025, available at

https://www.edps.europa.eu/system/files/2025

-

09/25

-

09

-

15_tech

dispatch

-

human

-

oversight_en.pdf;

EUROPEAN UNION AGENCY FOR FUNDAMENTAL RIGHTS.

Facial

Recognition Technology: Fundamental Rights Considerations in

the

Context

of

Law

Enforcement

.

Luxembourg:

Publications

Offi

ce

of

the

European

Union,

2019,

available

at

https://

fra.

-

europa.

eu/sites/default/files/fra_uploads/fra

-

2019

-

facial

-

recognition

-

technology

-

focus

-

paper

-

1_en.pdf;

EUROPEAN UNION AGENCY FOR FUNDAMENTAL RIGHTS.

Getting

the Future Right: Artificial Intelligence and Fundamental Rights.

Luxembourg: Publications Off

ice of the European Union, 2020,

AI-ASSISTED DECISION-MAKING IN PUBLIC SECURITY

84

available

at

https://fra.europa.eu/sites/

default/files/fra_uploads/

-

fra

-

2020

-

artificial

-

intelligence_en.pdf;

EUROPEAN UNION AGENCY FOR FUNDAMENTAL RIGHTS.

Bias in

Algorithms:

Artificial

Intelligence

and

Discrimination.

Luxembourg: Publications Off

ice of the European Union, 2022,

available

at

https://fra.europa.eu/sites/default/

files/fra_uploads/

-

fra

-

2022

-

bias

-

in

-

algorithms_en.pdf;

EUROPEAN

UNION

AGENCY

FOR

FUNDAMENTAL

RIGHTS.

Assessing

High

-

Risk

Artificial

Intelligence:

Fundamental

Rights

Risks.

Luxembourg: Publications Off

ice of the European Union,

2025,

available

at

https://fra.europa.eu/sites/default/

files/fra_

-

uploads/

fra

-

2025

-

assessing

-

high

-

risk

-

ai

-

fundamental

-

rights

-

risks_en.pdf;

EUROPOL.

AI

and

Policing:

The

Benefits

and

Challenges

of

Artificial

Intelligence

for

Law

Enforcement.

Luxembourg:

Publications

Office

of

the

European

Union,

2024.

https://www.europol.

-

europa.eu/publication

-

even

ts/main

-

reports/ai

-

and

-

policing;

EUROPOL.

AI Bias in Law Enforcement: A Practical Guide

. Luxembourg:

Publications

Off

ice

of

the

European

Union,

2025,

available

at

https://www.europol.europa.eu/cms/sites/default/files/documents/

AI_bias_in_law_en

forcement_

-

_practical_guide.pdf;

EUROPOL.

Assessing

Technologies

in

Law

Enforcement

.

Luxembourg:

Publications

Office

of

the

E

uropean

Union,

2025,

available

at

https://www.europol.europa.eu/cms/sites/default/files/documents/

Assessing

-

Tech

nologies

-

in

-

Law

-

enforcement.pdf;

HOBSON,

ZOË,

JULIA

A.

YESBERG,

BEN

BRADFORD,

and

JONATHAN

JACKSON.

“Artificial

Fairness?

Trust

in

Algorithmic

Police

Decision

-

Making.”

Journal

of

Experimental

Criminology 19 (2023): 1027

–

60. https://doi

.org/10.1007/s11292

-

021

-

09484

-

9;

INTERPOL.

INTERPOL

I

-

Learn

Catalogue

.

May

2026,

availble

at

https://www.interpol.int/content/download/24489/file/INTERPOL

%20I

-

Learn%20Catalogue%20May%202026.

pdf;

NATIONAL

INSTITUTE

OF

STANDARDS

AND

TECHNOLOGY.

Artificial

Intelligence

Risk

Management

Framework

(AI

RMF

1.0). NIST AI 100

-

1

. Gaithersburg, MD: NIST, 2023, available at

https://nvlpubs.nist.go

v/nistpubs/ai/NIST.AI.100

-

1.pdf;

EUROPEAN

PARLIAMENT

AND

COUNCIL

OF

THE

EUROPEAN

UNION.

Regulation

(EU)

2024/1689

of

13

June

2024

Laying

Down

Harmonised

Rules

on

Artificial

Intelligence

.

Official

Journal

of

the

European

Union

L

2024/1689,

available

at

https://eur

-

lex.europa.eu/legal

-

content/RO/TXT/PDF/

?uri=OJ:L_

-

202401689;

Colonel (ret) Professor George-Marius ȚICAL, Ph.D

85

UNESCO.

Recommendation on the Ethics of Artificial Intelligence

. Paris:

UNESCO,

2021,

available

at

https://unesdoc.unesco.org/

-

ark:/48223/pf0000381137;

UNICRI.

Decoding

Transparency:

How

to

Foster

Public

Trust

in

Responsible AI Innovation in Law Enforcement.

Turin: UNICRI,

2026,

available

at

https://unicri.org/sites/default/files/2026

-

04/

-

AI4

Citizens

-

Decoding

-

Transparency

-

How

-

to

-

Foster

-

Public

-

Trust

-

in

-

Responsible

-

AI

-

Inno

vation

-

in

-

Law

-

Enforcement_1.pdf;

UNICRI

and

INTERPOL.

Toolkit

for

Responsible

AI

Innovation

in

Law

Enforcement:

Organizational

Readiness

Assessment

Questionnaire

.

Turin/Lyon,

2024,

available

at

https://www.

-

interpol.int/content/

download/20933/file/AI_Toolkit_

_Organizati

onal_Readiness_Asse

ssment_Questionnaire%5B1%5D.pdf;

UNICRI

and

INTERPOL.

Toolkit

for

Responsible

AI

Innovation

in

Law

Enforcement:

Principles

for

Responsible

AI

Innovation.

Turin/Lyon,

2024,

available

at

https://unicri.org/sites/

default/

-

files/

2023

-

06/02_Principles%20for%

20Responding%

20AI%

-

20Innovation.pdf;

UNICRI

and

INTERPOL.

Toolkit

for

Responsible

AI

Innovation

in

Law

Enforcement:

Technical

Reference

Book.

Turin/Lyon,

2025,

available

at

https://unicri.org/sites/default/files/2025

-

03/Tech%

-

20Ref_Book_Mar25.pdf.