Comparing human-led reflective dialogue and AI-based tactical guidance as innovative training approaches in soccer: effects on tactical performance and problem-solving

Abstract

Background

The integration of human-led reflective dialogue and artificial intelligence represents two emerging, contrasting approaches to tactical training in soccer. However, the comparative effectiveness of these approaches within applied training environments remains insufficiently examined.

Aim

This study evaluated the differential effects of structured human idea-debate sessions versus AI-generated tactical guidance on tactical performance and problem-solving competencies during soccer training.

Methods

Eighty-four male amateur soccer players (age 17.1 ± 0.7 years) were randomly allocated to three groups (n = 28 each): Human Idea-Debate (HID), employing a play–discuss–play training model in which players collectively analyzed tactical problems and formulated solutions; AI-Ideation Delivery (AIID), implementing ChatGPT-generated tactical action plans as an AI-based training intervention; and a passive Control Group (CG). The intervention comprised eight small-sided game sessions (7v7 format, 50 × 40 m pitch), delivered twice per week over four weeks. Tactical performance was assessed using the Team Sport Assessment Procedure, while problem-solving competencies were evaluated using the validated Arabic Problem-Solving Inventory administered before and after the intervention.

Results

Significant group effects were observed for Received Ball (F = 4.02, p = 0.022, η²p = 0.043), with HID showing a 9.4% improvement compared with 7.0% in AIID and 3.8% in CG from Session 1 to Session 8. Performance Score demonstrated superior gains in HID (F = 4.12, p = 0.021, η²p = 0.045), increasing by 6.2% versus 4.1% in AIID. The Efficiency Index exhibited the largest improvement in HID (+ 70.5%), compared with + 18.7% in AIID. Problem-Solving Confidence differed significantly between groups (F = 14.82, p < 0.001, η²p = 0.162), with HID consistently outperforming both AIID and CG across all cognitive subscales.

Conclusion

Human-led reflective dialogue represents a more effective, innovative training approach than AI-generated tactical prescriptions for enhancing tactical performance and problem-solving in soccer. These findings highlight the critical role of social interaction, collective reasoning, and reflective learning processes in supporting adaptive decision-making and performance development in applied soccer training contexts.

Introduction

Tactical intelligence is a fundamental determinant of competitive success in soccer, requiring players to interpret dynamic game situations, coordinate collective actions, and execute contextually appropriate solutions under temporal and spatial constraints [1, 2]. This complexity demands rapid perception-action coupling and adaptive decision-making. Efficient inter-player communication is equally required, particularly in match-like training environments such as small-sided games (SSGs) [3]. Contemporary research grounded in ecological dynamics emphasizes that tactical team coordination emerges through continuous interactions between players and their environment rather than isolated individual cognition [4]. Team decisions arise from self-organized synergies shaped by shared affordances and dynamic coupling between performers and their environment. Ribeiro et al. [5] demonstrate that effective tactical development requires harnessing both top-down tactical structures and bottom-up adaptive interactions, promoting flexible tactical behavior through complementary global-to-local and local-to-global self-organizing processes. Empirical evidence in soccer indicates that progressively increasing tactical complexity during SSG training initially perturbs but ultimately enhances the synchronization of collective behaviors, illustrating how constraint manipulation facilitates the emergence of functional coordination patterns [6, 7].

Structured reflective dialogue has emerged as a theoretically grounded pedagogical strategy for tactical development in team sports. The debate-of-ideas methodology, in which players collectively identify tactical problems, negotiate solutions, and implement strategy during subsequent play, engages metacognitive processes that strengthen tactical awareness and decision-making autonomy [8]. Through guided discussion, players share situational interpretations and negotiate meanings. They co-construct tactical solutions grounded in direct field experience [9]. These communicative processes strengthen shared mental models, which function as organized knowledge structures that enable coordinated action selection, anticipatory positioning, and synchronized responses to opponents’ strategies [10]. Research demonstrates that cohesion, shared mental models, coordination, and collective efficacy form an interconnected network wherein each component reinforces the others, ultimately enhancing team functioning. Leadership behaviors that promote information exchange and perspective alignment accelerate the convergence of shared mental models, thereby elevating team performance over time [11]. The quality of wording and the organization of debate-of-ideas tasks significantly influence how learners articulate tactical solutions, negotiate meaning, and construct shared knowledge, thereby reinforcing cognitive engagement and collective tactical awareness [12]. However, group discussions may be influenced by cognitive biases, emotional states, hierarchical dynamics, and inconsistencies in player knowledge, which can limit the efficiency or accuracy of resulting action plans [13].

Parallel to pedagogical developments emphasizing collaborative learning, artificial intelligence has penetrated multiple domains of soccer performance analysis, offering new possibilities for tactical planning and decision support [14]. AI systems detect patterns in large datasets, model opponent behaviors, and generate action strategies with computational speed and analytical objectivity [15]. They also offer an important communicative function: converting complex tactical observations into structured, articulate recommendations that players may struggle to formulate independently [16–18]. Recent investigations highlight the utility of AI for tactical scouting, pattern recognition, and predictive modeling, thereby enabling data-driven recommendations for training and match preparation [19, 20]. AI-driven systems enhance player development by integrating machine learning models that detect performance patterns human observers often overlook [21]. These models enable more objective evaluations of technical ability, physical potential, and long-term developmental trajectories [22]. Algorithmic pattern recognition, game-state prediction, and advanced data mining reveal hidden tactical structures, optimize team shape, and support real-time strategic decision-making [19]. These technological capabilities not only help coaches understand complex tactical dynamics but also accelerate the creation of adaptive, data-informed strategies that enhance team performance [23]. Generative AI systems, particularly large language models, offer accessible platforms for delivering structured tactical guidance to athletes and coaches [15].

Despite increasing research on team cognition [24] and AI-assisted tactical instruction in sport contexts [21], direct comparisons between human-led dialogue and AI-generated guidance specifically in soccer training remain scarce. Existing studies typically examine either collective tactical reasoning within teams or algorithm-based decision-support systems in isolation, without exploring how each method shapes tactical behaviors that emerge during training tasks. Although AI-generated insights offer analytical advantages, they can lack contextual nuance, underestimate affective dimensions, or produce recommendations misaligned with implicit knowledge developed during authentic gameplay. This gap is particularly evident in SSG training, where tactical decisions unfold dynamically, and the source of instructional guidance may differentially shape emergent collective behavior. In this context, evidence directly comparing human-led debate and AI-generated guidance on tactical execution, spatial occupation, and adaptive coordination remains scarce. This absence constrains the evidence base for optimal integration of AI tools within pedagogical coaching frameworks. Furthermore, the fundamental question of whether AI-generated tactical recommendations can replicate or exceed the developmental benefits of human collaborative dialogue remains empirically unaddressed. This knowledge gap constrains evidence-based decision-making regarding optimal integration of technological tools within pedagogical frameworks.

The present investigation addresses this gap by comparing the effects of structured human idea-debate sessions with those of AI-generated action plans on tactical behaviors and problem-solving skills during soccer training. We hypothesized that human-driven communication processes, particularly idea-debate interactions, would produce greater improvements in tactical behavior than AI-based guidance. This prediction derives from theoretical propositions that debate-of-ideas requires active reasoning, the negotiation of meaning, and the shared interpretation of game situations, thereby cultivating deeper tactical awareness and stronger shared mental representations. Dialogic interaction enhances adaptability as players collectively construct solutions in response to dynamic on-field problems.

Methods

Study design and experimental protocol

This investigation employed a three-arm, parallel-group, randomized controlled trial conducted during the in-season competitive period, over four weeks, with two sessions administered per week (eight sessions total). The intervention framework utilized small-sided games (SSG) as the primary pedagogical vehicle for tactical development [3]. Participants were randomly allocated to one of three experimental conditions using computer-generated randomization sequences (allocation ratio 1:1:1): Human Idea-Debate group (HID), AI-Ideation Delivery group (AIID), and Control Group (CG). The HID condition employed a structured play–discuss–play methodology in which players collectively analyzed tactical problems encountered during gameplay and collaboratively formulated corrective action plans before subsequent play phases [25]. The AIID condition utilized ChatGPT-4o (OpenAI, San Francisco, CA, USA; accessed August 2024) to generate algorithmic tactical solutions in response to player-identified on-field challenges. The CG received no structured reflective intervention and completed equivalent SSG exposure with passive rest intervals. Single-blind procedures were implemented, whereby outcome assessors remained unaware of group allocation during video-based tactical behavior coding (Fig. 1).

Fig. 1.

Schematic representation of the randomized controlled trial design: intervention protocol, group allocation, training structure, assessment procedures, and temporal progression across the eight-session small-sided games intervention

The research protocol received ethical approval from the Institutional Review Board of the Higher Institute of Sport and Physical Education of El Kef (ISSEP-KEF), University of Jendouba, Tunisia (approval code: C-0035/2023), ensuring compliance with the principles of the Declaration of Helsinki for human research [26]. All participants and their legal guardians provided written informed consent following a comprehensive briefing on study objectives, procedures, potential risks, and data confidentiality measures. Participation was voluntary, with the explicit right to withdraw at any time without penalty. Personal identifiers were removed from all datasets, and encrypted digital storage protocols were used to ensure data security throughout the study [27]. All training sessions were supervised by two UEFA B-licensed soccer coaches with at least 5 years of experience in structured SSG-based training. Outcome assessments were conducted by ISAK Level 2 certified anthropometrists ("Anthropometric assessment procedures" section) and trained observers certified in video-based tactical behavior coding (κ = 0.92; "Tactical behavior assessment" section) [28].

Participants and sample size determination

All participants were engaged in their clubs’ standard competitive training schedule of three sessions per week, with official matches on weekends. Eighty-four male amateur soccer players (age 17.1 ± 0.7 years; stature 177.5 ± 5.8 cm; body mass 71.2 ± 7.6 kg; body mass index 22.6 ± 1.9 kg·m⁻²; body fat 12.4 ± 3.2%; training experience 5.6 ± 1.8 years) were recruited from clubs competing in the North-West region of the Tunisian Championship. Required sample size was determined a priori using G*Power 3.1 software [29], with the following parameters: repeated-measures analysis of variance (ANOVA) design, medium effect size (f = 0.25), alpha level of 0.05, statistical power of 0.85, three measurement groups, eight repeated measurements, and assumed correlation among repeated measures of 0.50. This yielded a minimum total sample requirement of 78 participants, with 26 per group. The final allocation of 28 participants per group provided an additional safeguard against potential attrition.

Inclusion criteria required regular participation in competitive soccer training (minimum of three sessions per week) and the absence of conditions that would limit full participation. All participants had prior exposure to SSG-based training as part of their regular club training, as confirmed by coach attestation. Exclusion criteria encompassed recent musculoskeletal injury within six months, diagnosed neurological or cardiovascular disorders, current use of medications affecting physical or cognitive performance, and inability to comply with study protocols or attend all scheduled sessions. Prior to data collection, all participants completed two familiarization sessions: one dedicated to the SSG format and the associated reflective or AI-based intervention procedures, and one to the TSAP video-coding assessment and PSI administration.

Anthropometric assessment procedures

Pre-intervention anthropometric measurements were conducted in a climate-controlled laboratory at ISSEP-KEF in accordance with standardized protocols. Participants were instructed to avoid strenuous exercise, caffeine, and food for 2 h prior to the assessment. Stature was measured barefoot to the nearest 0.1 cm using a calibrated stadiometer (Seca 213, Hamburg, Germany). Body mass was measured to the nearest 0.1 kg using a digital scale (Tanita BC-601, Tokyo, Japan), and body mass index was calculated. Waist and hip circumferences were measured with a flexible anthropometric tape (Lufkin W606PM, Apex, NC, USA). Skinfold thickness was assessed at four sites (triceps, subscapular, suprailiac, thigh) using Harpenden calipers (Baty International, West Sussex, UK) according to International Society for the Advancement of Kinanthropometry protocols. All measurements were performed by ISAK Level 2-certified anthropometrists, with a technical error of measurement below 5% [28].

Training intervention protocol

The SSG intervention replaced one standard weekly training session in all three groups, ensuring equivalent total training volume. Session intensity was controlled via heart-rate monitoring, targeting 80–90% of age-predicted maximum HR across all conditions [30]. All training sessions were conducted under identical temporal and environmental conditions. Sessions occurred on outdoor natural grass pitches (68 × 105 m) between 16:00 and 18:00 h under ambient temperatures ranging from 18 °C to 26 °C. Each 65-minute session comprised a standardized 15-minute warm-up (coordinated passing exercises, four 3-minute intermittent sprint blocks, dynamic stretching), followed by the SSG intervention block (30 min total: two 10-minute bouts in 7v7 format on a 50 × 40 m pitch with a 10-minute inter-bout interval), and concluding with technical-tactical consolidation (10 min) and cool-down (10 min). Target exercise intensity was maintained at 80–90% of age-predicted maximum heart rate, consistent with validated SSG physiological demands [31].

During the 10-minute inter-bout rest period, intervention groups engaged in distinct reflective processes. The HID group used collaborative tactical problem-solving discussions to address session-specific objectives (Table 1). The AIID group followed a structured five-step protocol during each inter-bout interval. (1) The supervising coach verbally identified the primary tactical problem observed during Bout 1, based on live observation rather than video replay. (2) One designated player typed this standardized tactical challenge as a prompt into ChatGPT-4o (OpenAI, San Francisco, CA, USA; accessed via tablet device, August 2024). (3) The AI-generated response was displayed on the shared tablet screen, positioned so all group members could read it simultaneously. (4) Players read the recommendations individually and silently for approximately three minutes. (5) Players then implemented the AI-generated solutions during Bout 2 without group discussion. Players were explicitly instructed by the supervising coach not to discuss AI-generated content with teammates during the reading interval. The game itself was not filmed prior to prompting; the prompt reflected the coach’s real-time tactical observation rather than post-hoc video analysis. The complete session-by-session objectives, prompts, and AI-generated solutions for all three groups are summarized in Table 1. The CG completed passive rest without tactical intervention. Progressive tactical complexity was implemented across the eight-session mesocycle while maintaining consistent physical loading parameters.

Table 1.

Tactical objectives, inter-bout interval procedures, and group-specific instructional content across eight small-sided game sessions for human Idea-debate (HID), AI-Ideation delivery (AIID), and control (CG) groups

Session	Tactical Objective	HID Condition: Play-Discuss-Play		AIID Condition: AI-Ideation Delivery		CG
		Discussion Focus (Problems Identified During Bout 1)	Collaboratively Constructed Action Plans	Player Prompt (Typed into ChatGPT-4o)	AI-Generated Solutions
1	Possession and scanning	Players struggled to maintain possession under pressure; they often sought safe options rather than progressing the ball; poor scanning reduced passing options.	Constantly scan before receiving the ball; create passing triangles; encourage supporting runs to provide outlets; maintain team shape.	“We keep losing the ball under pressure.”	(1) Scan surroundings before receiving. (2) Move to create passing triangles. (3) Maintain spacing and team shape.	Passive rest
2	Counter-attack	Difficulty transitioning quickly after regaining possession; players’ delayed decision-making caused missed opportunities.	Identify triggers for immediate attack; assign specific players to exploit spaces; practice quick forward passes; reinforce speed of decision and movement off the ball.	“We are too slow to exploit turnovers.”	(1) Identify triggers for immediate attack. (2) Assign players to exploit space. (3) Quick forward passes and movement off the ball.	Passive rest
3	Overloads and role clarity	Confusion about responsibilities in offensive and defensive phases; numerical imbalance created vulnerabilities.	Clarify player roles for each phase; coordinate positional rotations; emphasize exploiting overloads; reinforce verbal and non-verbal communication to maintain balance.	“Roles are unclear, leaving gaps in defense.”	(1) Define roles in each phase. (2) Coordinate positional rotations. (3) Use verbal and non-verbal communication to maintain balance.	Passive rest
4	Progression to the target zone	Players struggled to break compact defensive lines; timing and weight of passes were often incorrect.	Use supporting players to create space; focus on timing and direction of passes; exploit wide channels; combine dribbles and passes strategically.	“We cannot break the compact defensive lines.”	(1) Exploit wide channels. (2) Use support players to create passing options. (3) Optimize timing and weight of passes.	Passive rest
5	Transitions and counter-press	Delayed reaction after losing possession; the team became disorganized, leaving gaps for opponents to exploit.	Immediate counter-press on loss; organize defensive lines quickly; communicate pressing responsibilities; assign recovery zones for faster team regrouping.	“We lose shape after a possession loss.”	(1) Counter-press immediately. (2) Reorganize defensive lines quickly. (3) Communicate pressing responsibilities. (4) Assign recovery zones.	Passive rest
6	Defensive reorganization	Players were slow to recover positions and mark opponents; defensive compactness was inconsistent.	Assign specific recovery zones; coordinate pressing triggers; communicate shifts in defensive line; practice synchronized movements to close gaps efficiently.	“We are slow to recover positions and mark opponents.”	(1) Assign recovery zones. (2) Synchronize defensive movements. (3) Communicate pressing and marking tasks.	Passive rest
7	Penetration and final pass	Hesitation in the attacking third under pressure; ineffective final passes and turnovers.	Identify penetrating lanes in advance; anticipate opponent positioning; maintain composure; increase decision-making speed; rehearse combinations to break defensive lines.	“We hesitate in the attacking third under pressure.”	(1) Identify penetrating lanes in advance. (2) Maintain composure under pressure. (3) Increase decision speed. (4) Rehearse combinations to break lines.	Passive rest
8	Build-up play	Difficulty implementing structured build-up from defense to attack; players relied on individual actions instead of collective organization.	Emphasize playing from the back; link defensive and midfield lines; organize passing sequences to progress the ball methodically; integrate learned tactical concepts; encourage constant scanning and communication.	“We struggle to organize structured attacks from defense.”	(1) Play from the back and link lines. (2) Organize passing sequences methodically. (3) Encourage scanning and communication. (4) Integrate all tactical concepts learned.	Passive rest

HID Human Idea-Debate group (play-discuss-play methodology: players collectively identified tactical problems during Bout 1 and constructed action plans prior to Bout 2), AIID AI-Ideation Delivery group (standardized five-step protocol: coach identified tactical problem; one player typed prompt into ChatGPT-4o via shared tablet; AI-generated response displayed to group; players read silently approximately 3 min; solutions implemented in Bout 2 without group discussion). CG Control Group (passive rest during inter-bout interval; no structured reflective or AI-based intervention). All three groups completed identical SSG exposure (two 10-minute bouts, 7v7 format, 50 × 40 m pitch). Intervention dates: Sessions 1–8, November 4–27, 2025 (in-season competitive period), twice per week

To ensure training load equivalence across the three clubs, the experimental SSG sessions replaced one regular weekly training session for all groups, maintaining the standard three-session-per-week schedule. Heart-rate monitoring confirmed that all groups trained at equivalent physiological intensities (80–90% of age-predicted maximum HR) throughout the intervention.

Outcome measures and psychometric properties

Tactical behavior assessment

Individual tactical performance was quantified using an adapted Team Sport Assessment Procedure (TSAP) derived from Grehaigne et al.’s study [32]. All SSG activity was recorded using a GoPro HERO12 Black camera (CHDHX-121-RW; 5.3 K resolution, 60 fps; GoPro Inc., San Mateo, CA, USA). The camera was positioned to capture the full pitch at all times. Trained observers coded the following variables for each player during both pre-intervention (Bout 1) and post-intervention (Bout 2) phases: Received Ball (RB), Conquered Ball (CB), Neutral Ball (NB), Lost Ball (LB), Offensive Ball (OB), and Successful Shot (SS). Composite indices included Volume of Play [VP = RB + CB], Efficiency Index [EI = (OB + SS) / (10 + LB)], and Performance Score [PS = (VP / 2) + (EI × 10)]. Inter-rater reliability was established through independent coding of 18% of the total footage by two assessors, yielding Cohen’s kappa of 0.92, indicating excellent agreement [33].

Problem-solving skills assessment

The validated Arabic version of the Problem-Solving Inventory (PSI) was administered immediately before and after each training session. This 35-item instrument utilizes a 6-point Likert scale (1 = strongly agree; 6 = strongly disagree) to evaluate three dimensions: Problem-Solving Confidence (PSC; 11 items), Approach–Avoidance Style (AAS; 16 items), and Personal Control (PC; 5 items). Higher scores indicate diminished problem-solving ability. The Arabic PSI demonstrates robust psychometric properties, with Cronbach’s alpha coefficients exceeding 0.80 across all subscales, and adequate construct validity in Arabic-speaking populations [34].

Statistical analysis

Data were analyzed using SPSS Statistics version 29.0 (IBM Corporation, Armonk, NY, USA). Descriptive statistics comprised means and standard deviations for all continuous variables. Normality was assessed using the Shapiro-Wilk test [35], and homogeneity of variance was verified via Levene’s test [36]. A three-way mixed-design ANOVA with one between-subjects factor (Group: HID, AIID, CG) and two within-subjects factors (Time: pre-session, post-session; Session: 1–8) examined main effects and interactions [37]. Effect sizes were interpreted using partial eta-squared (η²p): small (0.01), medium (0.06), and large (0.14) [38]. Significant omnibus effects were decomposed using Bonferroni-corrected pairwise comparisons [37]. Mauchly’s test assessed sphericity, with Greenhouse-Geisser corrections applied when ε < 0.75 [39]. Statistical significance was established at α = 0.05 (two-tailed). Percentage change calculations quantified pre-to-post alterations within sessions using the formula: [(post-score − pre-score) / pre-score] × 100. This trial was reported in accordance with the Consolidated Standards of Reporting Trials (CONSORT) statement for parallel-group randomized controlled trials. A completed CONSORT checklist is provided as supplementary material.

Results

Tactical performance outcomes

Descriptive patterns across intervention groups

Tactical performance variables exhibited distinct trajectories across the three experimental conditions during the eight-session intervention period (Table 2). The HID group showed the most pronounced improvements across all TSAP metrics. Received Ball scores increased from 196.2 ± 27.8 at Session 1 to 233.4 ± 35.9 at Session 8. Conquered Ball progressed from 9.3 ± 2.5 to 12.6 ± 3.6. Volume of Play advanced from 205.5 ± 39.3 to 245.9 ± 26.5. Neutral Ball increased from 135.2 ± 24.9 to 146.0 ± 28.4, while Lost Ball decreased from 22.6 ± 5.0 to 20.6 ± 4.3. Offensive Ball improved from 31.3 ± 6.7 to 36.0 ± 6.6. Successful Shot escalated from 17.1 ± 5.0 to 20.4 ± 5.5. The efficiency index increased substantially from 1.60 ± 0.5 to 2.9 ± 0.6. Performance Score increased from 120.1 ± 14.7 to 134.8 ± 16.2. The AIID group showed moderate improvements, with Received Ball increasing from 191.8 ± 29.4 to 225.9 ± 35.6 and Performance Score increasing from 117.9 ± 14.2 to 142.9 ± 18.0. The CG exhibited minimal alterations, with Received Ball changing from 187.3 ± 26.1 to 195.1 ± 28.1 and Performance Score from 116.9 ± 14.0 to 127.1 ± 16.1 (Table 2).

Table 2.

Descriptive statistics (M ± SD) for team sport assessment procedure variables measured pre-intervention (Bout 1) and post-intervention (Bout 2) across eight training sessions in Human Idea-Debate (HID), AI-Ideation Delivery (AIID), and Control (CG) Groups

Groups		HID (n = 28) M ± SD		AIID (n = 28) M ± SD		CG (n = 28) M ± SD
Session	Variable	Pre	Post	Pre	Post	Pre	Post
1	RB	196.2 ± 27.8	214.7 ± 32.2	191.8 ± 29.4	207.6 ± 31.7	187.3 ± 26.1	190.2 ± 22.2
	CB	9.3 ± 2.5	10.7 ± 2.9	9.0 ± 2.7	10.4 ± 3.1	9.1 ± 2.8	9.3 ± 2.9
	VP	205.5 ± 39.3	225.4 ± 62.1	200.8 ± 51.9	218.0 ± 41.8	196.4 ± 28.7	199.7 ± 79.5
	NB	135.2 ± 24.9	142.8 ± 26.1	132.1 ± 26.3	139.5 ± 27.6	130.3 ± 25.4	131.6 ± 25.8
	LB	22.6 ± 5.0	20.0 ± 4.7	22.0 ± 4.8	20.7 ± 5.1	22.9 ± 5.2	22.6 ± 5.1
	OB	31.3 ± 6.7	34.2 ± 6.5	30.9 ± 6.4	33.0 ± 6.1	30.6 ± 6.3	30.8 ± 6.1
	SS	17.1 ± 5.0	19.4 ± 5.1	16.5 ± 4.9	18.5 ± 5.2	16.1 ± 4.8	16.3 ± 4.7
	EI	1.60 ± 0.5	1.83 ± 0.7	1.54 ± 0.4	1.76 ± 0.3	1.50 ± 0.4	1.52 ± 0.4
	PS	120.1 ± 14.7	127.6 ± 15.9	117.9 ± 14.2	124.7 ± 15.5	116.9 ± 14.0	117.8 ± 14.2
2	RB	198.7 ± 40.1	217.5 ± 21.3	193.5 ± 58.6	210.3 ± 32.4	188.9 ± 27.6	191.1 ± 38.0
	CB	9.5 ± 2.6	10.9 ± 2.8	9.2 ± 3.1	10.6 ± 3.4	9.0 ± 2.7	9.2 ± 2.9
	VP	208.2 ± 49.9	228.4 ± 32.6	202.7 ± 30.8	220.9 ± 51.5	197.9 ± 28.5	200.3 ± 19.4
	NB	136.1 ± 25.3	144.0 ± 25.9	134.5 ± 25.1	141.8 ± 26.2	131.0 ± 24.7	132.1 ± 25.0
	LB	22.3 ± 4.8	19.7 ± 4.5	21.9 ± 4.5	20.3 ± 4.9	22.8 ± 4.9	22.5 ± 4.8
	OB	31.9 ± 6.5	34.7 ± 6.0	31.4 ± 6.1	33.9 ± 6.3	30.9 ± 6.2	31.0 ± 6.0
	SS	17.4 ± 5.2	19.7 ± 5.5	16.8 ± 5.1	18.9 ± 5.4	16.3 ± 4.6	16.4 ± 4.7
	EI	1.6 ± 0.2	1.8 ± 0.5	1.5 ± 0.4	1.7 ± 0.6	1.5 ± 0.4	1.5 ± 0.4
	PS	121.0 ± 14.3	129.2 ± 15.3	118.9 ± 13.8	126.0 ± 15.1	117.5 ± 13.9	118.3 ± 14.2
3	RB	200.4 ± 25.7	220.7 ± 32.6	195.1 ± 31.3	213.4 ± 43.9	189.5 ± 46.9	191.5 ± 57.5
	CB	9.6 ± 2.8	11.0 ± 2.7	9.4 ± 2.9	10.7 ± 3.0	9.2 ± 3.0	9.3 ± 2.8
	VP	210.0 ± 30.4	231.7 ± 34.1	204.5 ± 33.1	224.1 ± 34.7	198.7 ± 27.8	200.8 ± 28.1
	NB	137.9 ± 25.8	145.7 ± 26.1	136.3 ± 27.0	143.4 ± 26.8	131.4 ± 24.3	132.8 ± 25.1
	LB	22.0 ± 4.7	19.3 ± 4.6	21.7 ± 5.3	20.1 ± 5.2	22.6 ± 4.6	22.4 ± 4.7
	OB	32.2 ± 6.3	35.0 ± 6.7	31.9 ± 6.9	34.3 ± 6.4	31.1 ± 6.1	31.3 ± 6.0
	SS	17.7 ± 5.3	19.9 ± 5.5	17.0 ± 4.8	19.0 ± 5.0	16.4 ± 4.8	16.5 ± 4.7
	EI	1.6 ± 0.1	1.8 ± 0.8	1.5 ± 0.4	1.8 ± 0.5	1.5 ± 0.4	1.5 ± 0.4
	PS	121.9 ± 14.6	130.8 ± 15.8	119.8 ± 15.0	127.1 ± 15.4	117.8 ± 14.1	118.6 ± 14.0
4	RB	202.6 ± 28.9	223.1 ± 33.5	196.7 ± 32.2	215.9 ± 34.4	190.3 ± 27.2	192.4 ± 27.8
	CB	9.9 ± 2.7	11.3 ± 3.0	9.6 ± 3.0	10.9 ± 3.3	9.4 ± 2.9	9.6 ± 3.0
	VP	212.5 ± 31.0	234.4 ± 33.8	206.3 ± 33.9	226.8 ± 34.9	199.7 ± 29.6	202.0 ± 30.1
	NB	139.8 ± 26.7	146.9 ± 27.9	137.9 ± 27.5	144.7 ± 28.1	132.6 ± 25.7	133.9 ± 26.4
	LB	21.9 ± 4.8	18.8 ± 4.6	21.3 ± 5.1	19.8 ± 5.2	22.3 ± 5.0	22.1 ± 5.1
	OB	32.6 ± 6.2	35.4 ± 6.4	32.0 ± 6.5	34.6 ± 6.6	31.6 ± 6.0	31.8 ± 6.1
	SS	18.0 ± 5.1	20.1 ± 5.4	17.3 ± 5.0	19.3 ± 5.3	16.6 ± 4.9	16.8 ± 5.0
	EI	1.6 ± 0.7	1.9 ± 0.9	1.5 ± 0.2	1.8 ± 0.7	1.5 ± 0.9	1.6 ± 0.7
	PS	123.9 ± 15.0	132.7 ± 16.1	121.6 ± 15.6	129.4 ± 16.2	119.0 ± 14.5	120.2 ± 14.9
5	RB	204.1 ± 30.3	225.6 ± 34.1	198.9 ± 30.8	218.4 ± 33.0	191.1 ± 26.7	193.0 ± 27.0
	CB	10.2 ± 2.9	11.6 ± 3.3	9.8 ± 3.2	11.1 ± 3.5	9.6 ± 2.8	9.8 ± 2.9
	VP	214.3 ± 31.8	237.2 ± 34.5	208.7 ± 33.0	229.5 ± 34.6	200.7 ± 29.2	202.8 ± 30.0
	NB	141.0 ± 27.1	148.0 ± 27.9	139.1 ± 27.8	146.2 ± 28.7	133.4 ± 26.0	134.7 ± 26.6
	LB	21.6 ± 4.6	18.3 ± 4.5	20.9 ± 4.9	19.4 ± 5.0	22.1 ± 4.8	21.8 ± 4.9
	OB	33.9 ± 6.6	36.5 ± 6.8	33.2 ± 6.8	35.4 ± 6.7	32.3 ± 6.3	32.6 ± 6.4
	SS	18.6 ± 5.3	20.5 ± 5.6	17.9 ± 5.2	19.9 ± 5.5	16.9 ± 5.1	17.1 ± 5.2
	EI	1.9 ± 0.4	1.9 ± 0.6	1.6 ± 0.3	1.6 ± 0.8	1.5 ± 0.2	1.5 ± 0.1
	PS	126.4 ± 15.3	135.8 ± 16.7	123.8 ± 15.9	132.6 ± 16.6	120.5 ± 14.8	121.9 ± 15.1
6	RB	205.6 ± 29.1	227.9 ± 33.9	200.5 ± 29.7	220.7 ± 32.8	192.0 ± 27.3	193.5 ± 27.9
	CB	10.4 ± 2.7	11.9 ± 3.1	10.0 ± 2.8	11.4 ± 3.2	9.7 ± 2.9	9.9 ± 3.0
	VP	216.0 ± 30.2	239.8 ± 33.8	210.5 ± 32.1	232.1 ± 33.7	201.7 ± 29.0	203.4 ± 30.2
	NB	142.7 ± 27.3	150.5 ± 28.5	140.9 ± 28.2	148.6 ± 29.0	134.7 ± 26.4	135.9 ± 26.9
	LB	21.2 ± 4.5	18.1 ± 4.4	20.5 ± 4.7	19.0 ± 4.9	21.9 ± 4.6	21.6 ± 4.7
	OB	34.8 ± 6.4	37.2 ± 6.6	34.0 ± 6.6	36.3 ± 6.8	33.1 ± 6.1	33.4 ± 6.2
	SS	19.2 ± 5.2	21.2 ± 5.6	18.5 ± 5.1	20.4 ± 5.4	17.4 ± 5.0	17.7 ± 5.1
	EI	1.7 ± 0.5	1.9 ± 0.6	1.6 ± 0.5	1.8 ± 0.9	1.4 ± 0.1	1.5 ± 0.2
	PS	129.0 ± 15.7	139.5 ± 17.1	126.2 ± 16.1	135.3 ± 16.9	122.2 ± 15.0	123.8 ± 15.3
7	RB	207.4 ± 30.4	230.8 ± 35.2	202.1 ± 31.1	223.5 ± 35.0	192.8 ± 27.0	194.4 ± 28.2
	CB	10.7 ± 2.9	12.3 ± 3.5	10.2 ± 3.1	11.7 ± 3.6	9.9 ± 3.0	10.1 ± 3.1
	VP	218.1 ± 31.9	243.1 ± 35.8	212.3 ± 33.4	235.2 ± 35.4	202.7 ± 29.8	204.5 ± 31.0
	NB	144.2 ± 27.9	152.6 ± 29.8	142.1 ± 28.6	150.4 ± 30.0	135.6 ± 26.9	136.9 ± 27.6
	LB	20.9 ± 4.4	17.9 ± 4.3	20.1 ± 4.6	18.8 ± 4.7	21.5 ± 4.5	21.2 ± 4.6
	OB	35.5 ± 6.5	38.1 ± 7.0	34.6 ± 6.7	37.0 ± 6.9	33.6 ± 6.2	34.0 ± 6.4
	SS	19.8 ± 5.4	22.1 ± 5.8	19.0 ± 5.3	21.0 ± 5.7	17.8 ± 5.1	18.2 ± 5.2
	EI	1.7 ± 0.5	2.3 ± 0.6	1.6 ± 0.5	1.9 ± 0.2	1.58 ± 0.52	1.61 ± 0.53
	PS	131.9 ± 15.9	142.7 ± 17.9	128.8 ± 16.4	139.6 ± 17.6	123.9 ± 15.4	125.7 ± 15.9
8	RB	209.1 ± 31.2	233.4 ± 35.9	203.8 ± 32.0	225.9 ± 35.6	193.6 ± 27.5	195.1 ± 28.1
	CB	11.0 ± 3.0	12.6 ± 3.6	10.5 ± 3.2	12.0 ± 3.7	10.0 ± 3.1	10.2 ± 3.2
	VP	220.1 ± 52.2	245.9 ± 26.5	214.3 ± 34.1	237.9 ± 46.0	204.1 ± 30.3	206.1 ± 31.4
	NB	146.0 ± 28.4	154.2 ± 30.7	143.5 ± 29.1	151.6 ± 31.2	136.4 ± 27.4	138.0 ± 28.3
	LB	20.6 ± 4.3	17.6 ± 4.2	19.8 ± 4.5	18.5 ± 4.6	21.1 ± 4.4	20.8 ± 4.5
	OB	36.0 ± 6.6	38.9 ± 7.2	35.2 ± 6.9	37.8 ± 7.1	34.2 ± 6.3	34.6 ± 6.5
	SS	20.4 ± 5.5	22.9 ± 6.0	19.6 ± 5.4	21.6 ± 5.9	18.4 ± 5.2	18.9 ± 5.4
	EI	1.7 ± 0.9	2.9 ± 0.6	1.6 ± 0.5	1.9 ± 0.3	1.6 ± 0.3	1.6 ± 0.4
	PS	134.8 ± 16.2	146.9 ± 18.3	131.5 ± 16.8	142.9 ± 18.0	125.4 ± 15.7	127.1 ± 16.1

M mean, SD standard deviation, HID Human Idea-Debate group, AIID AI Idea-Debate group, CG Control Group, RB Received Ball, CB Conquered Ball, VP Volume of Play, NB Neutral Ball, LB Lost Ball, OB Offensive Ball, SS Successful Shot or Scoring attempts, EI Efficiency Index, PS Performance Score

Inferential statistical outcomes for tactical variables

A three-way repeated-measures ANOVA revealed significant main effects of group allocation across multiple tactical indices (Table 3). Significant between-group differences emerged for Received Ball (F = 4.02, p = 0.022, η²p = 0.043), Volume of Play (F = 3.97, p = 0.023, η²p = 0.042), Offensive Ball (F = 3.11, p = 0.048, η²p = 0.030), Successful Shot (F = 3.45, p = 0.034, η²p = 0.033), and Performance Score (F = 4.12, p = 0.021, η²p = 0.045). Session effects were statistically significant across nearly all variables (F = 7.46–16.11, p < 0.001–0.029, η²p = 0.124–0.265), indicating progressive improvement from Session 1 to Session 8. Time effects indicated consistent pre-to-post session gains (F = 4.52–19.02, p = 0.002–0.036, η²p = 0.021–0.281).

Table 3.

Three-way repeated-measures analysis of variance results for tactical performance variables: main effects, interaction effects, effect sizes, and Bonferroni-Corrected Post-Hoc Comparisons Across Group (HID, AIID, CG), Session (1–8), and Time (Pre-Post)

Variable	Source	F	P	Partial η²	Bonferroni Post Hoc
RB	Group	4.02	0.022	0.043	HID > CG (p = 0.020); AIID > CG (p = 0.045)
	Session	16.11	< 0.001	0.265	S1 < S8 (p = 0.001) ; S4 < S8 (p = 0.021)
	Time	19.02	0.002	0.281	Pre < Post (p = 0.001)
	Group × Session	1.91	0.044	0.033	HID S8 > CG S8 (p = 0.037)
	Group × Time	2.29	0.106	0.018	NS
	Session × Time	2.54	0.016	0.043	S1 Pre < S8 Post (p = 0.004)
	Group × Session × Time	1.18	0.287	0.021	NS
CB	Group	2.87	0.061	0.028	NS
	Session	12.94	0.004	0.223	S1 < S8 (p = 0.003)
	Time	17.05	< 0.001	0.254	Pre < Post (p = 0.001)
	Group × Session	1.62	0.092	0.028	NS
	Group × Time	1.95	0.146	0.018	NS
	Session × Time	2.21	0.039	0.036	S1 Pre < S8 Post (p = 0.012)
	Group × Session × Time	1.05	0.404	0.019	NS
VP	Group	3.97	0.023	0.042	HID > CG (p = 0.022)
	Session	15.88	< 0.001	0.263	S1 < S8 (p = 0.001)
	Time	18.74	< 0.001	0.274	Pre < Post (p = 0.001)
	Group × Session	1.88	0.047	0.032	HID S5 > CG S5 (p = 0.041)
	Group × Time	2.26	0.109	0.017	NS
	Session × Time	2.62	0.014	0.044	S1 Pre < S8 Post (p = 0.003)
	Group × Session × Time	1.19	0.282	0.021	NS
NB	Group	1.89	0.156	0.018	NS
	Session	7.46	< 0.001	0.124	S1 < S8 (p = 0.029)
	Time	4.52	0.036	0.021	Pre > Post (p = 0.036)
	Group × Session	1.08	0.385	0.019	NS
	Group × Time	0.91	0.406	0.009	NS
	Session × Time	1.42	0.199	0.024	NS
	Group × Session × Time	0.92	0.538	0.015	NS
LB	Group	1.12	0.331	0.011	NS
	Session	8.11	< 0.001	0.131	S1 > S8 (p = 0.022)
	Time	5.88	0.017	0.026	Pre > Post (p = 0.017)
	Group × Session	1.03	0.433	0.018	NS
	Group × Time	1.02	0.364	0.010	NS
	Session × Time	1.71	0.111	0.028	NS
	Group × Session × Time	0.97	0.497	0.016	NS
OB	Group	3.11	0.048	0.030	HID > CG (p = 0.043)
	Session	11.26	< 0.001	0.190	S1 < S8 (p = 0.002)
	Time	15.89	< 0.001	0.233	Pre < Post (p = 0.001)
	Group × Session	1.41	0.154	0.024	NS
	Group × Time	1.72	0.183	0.016	NS
	Session × Time	2.13	0.042	0.035	S1 Pre < S8 Post (p = 0.019)
	Group × Session × Time	1.04	0.422	0.018	NS
SS	Group	3.45	0.034	0.033	HID > CG (p = 0.040)
	Session	12.18	< 0.001	0.206	S1 < S8 (p = 0.001)
	Time	16.45	< 0.001	0.241	Pre < Post (p = 0.001)
	Group × Session	1.57	0.086	0.027	NS
	Group × Time	1.88	0.155	0.017	NS
	Session × Time	2.33	0.029	0.038	S1 Pre < S8 Post (p = 0.015)
	Group × Session × Time	1.07	0.399	0.019	NS
EI	Group	2.94	0.057	0.027	NS
	Session	10.12	< 0.001	0.176	S1 < S8 (p = 0.004)
	Time	14.23	< 0.001	0.214	Pre < Post (p = 0.002)
	Group × Session	1.32	0.180	0.023	NS
	Group × Time	1.61	0.204	0.015	NS
	Session × Time	1.98	0.060	0.033	NS
	Group × Session × Time	0.95	0.519	0.016	NS
PS	Group	4.12	0.021	0.045	HID > CG (p = 0.024)
	Session	15.37	< 0.001	0.256	S1 < S8 (p = 0.001)
	Time	18.45	< 0.001	0.276	Pre < Post (p = 0.001)
	Group × Session	1.87	0.048	0.032	HID S8 > CG S8 (p = 0.038)
	Group × Time	2.31	0.104	0.017	NS
	Session × Time	2.68	0.013	0.045	S1 Pre < S8 Post (p = 0.010)
	Group × Session × Time	1.21	0.274	0.022	NS

F F-Value, P P-Value; η² Partial eta square (Effect Size); HID Human Idea-Debate group, AIID AI Idea-Debate group, CG Control Group, RB Received Ball, CB Conquered Ball, VP Volume of Play, NB Neutral Ball, LB Lost Ball, OB Offensive Ball, SS Successful Shot or Scoring attempts, EI Efficiency Index, PS Performance Score, S1 Session 1, S4 Session 4, S8 Session 8, NS No Significant difference; Statistical significance was set at α = 0.05 (two-tailed). Values at p < 0.001 are indicated accordingly

Interaction effects provided additional insight into the dynamics of the intervention. Group × Session interactions were significant for Received Ball, Volume of Play, and Performance Score, indicating that HID improvements intensified in later sessions. Session × Time interactions for Received Ball, Conquered Ball, Volume of Play, Offensive Ball, and Successful Shot indicated that pre-to-post gains accumulated across the intervention period. Conversely, the Conquered Ball and Efficiency Index did not reach statistical significance in group comparisons (p = 0.061 and p = 0.057, respectively). Neutral Ball and Lost Ball showed non-significant group effects (p = 0.156 and p = 0.331, respectively).

Bonferroni-corrected post-hoc comparisons confirmed HID superiority over CG for Received Ball (p = 0.020), with AIID also exceeding CG (p = 0.045). The most significant between-group differences emerged at Session 8 for Received Ball (HID > CG, p = 0.037). Volume of Play demonstrated HID advantage over CG at Session 5 (p = 0.041). Offensive Ball, Successful Shot, and Performance Score each showed significant HID superiority relative to CG (p = 0.043, 0.040, and 0.024, respectively). Session-wise comparisons indicated that Session 8 values significantly exceeded those in Session 1 for Received Ball, Conquered Ball, Volume of Play, Offensive Ball, Successful Shot, Efficiency Index, and Performance Score (p = 0.001–0.029).

Percentage change analyses revealed that the Efficiency Index showed the most pronounced improvement trajectory in the HID group, increasing from 14.3% in Session 1 to 70.5% in Session 8 (Fig. 2). Other tactical variables exhibited more modest pre-to-post session changes, ranging from 6.2% to 14.5% across the intervention period. The AIID group displayed attenuated but comparable patterns, whereas CG changes remained consistently below 3% for all variables.

Fig. 2.

Percentage change in team sport assessment procedure variables from pre-intervention (Bout 1) to post-intervention (Bout 2) across eight training sessions for Human Idea-Debate (HID), AI-Ideation Delivery (AIID), and Control (CG) Groups. The y-axis represents the percentage change from pre-session baseline values calculated as [(post-score − pre-score) / pre-score] × 100

Problem-solving competencies

Problem-solving inventory descriptive trends

Problem-solving variables demonstrated a systematic improvement over the intervention period, with group-specific patterns (Table 4). For Problem-Solving Confidence, HID pre-session scores progressed from 21.0 ± 4.0 at Session 1 to 29.9 ± 3.1 at Session 8, with corresponding post-session values of 20.5 ± 3.9 to 29.0 ± 3.0. Approach-Avoidance Style pre-session scores advanced from 17.6 ± 5.8 to 25.3 ± 2.9, with post-session scores of 17.2 ± 3.7 to 24.6 ± 2.6. Personal Control pre-session values increased from 13.9 ± 3.0 to 19.7 ± 2.3, with post-session scores of 13.5 ± 2.9 to 19.1 ± 2.2. The AIID group exhibited intermediate gains, with pre-session Problem-Solving Confidence scores ranging from 20.3 ± 4.2 to 27.0 ± 3.5. The CG demonstrated minimal variation across all subscales throughout the intervention period.

Table 4.

Descriptive statistics (M ± SD) for problem-solving inventory subscales (problem-solving confidence, approach-avoidance style, personal control) measured pre-session and post-session across eight training sessions in Human Idea-Debate (HID), AI-Ideation Delivery (AIID), and Control (CG) Groups

Variable (PSI)		PSC		AAS		PC
		MSD		MSD		MSD
Session	Group	Pre	Post	Pre	Post	Pre	Post
1	HID	21.0 ± 4.0	20.5 ± 3.9	17.6 ± 5.8	17.2 ± 3.7	13.9 ± 3.0	13.5 ± 2.9
	AIID	20.3 ± 4.2	19.8 ± 4.1	17.0 ± 3.8	16.6 ± 3.7	13.6 ± 3.1	13.2 ± 3.0
	CG	19.8 ± 4.1	19.3 ± 4.0	16.8 ± 3.6	16.5 ± 3.5	13.5 ± 3.2	13.3 ± 3.1
2	HID	22.3 ± 3.9	21.7 ± 3.8	18.5 ± 3.6	18.0 ± 3.4	14.9 ± 2.9	14.5 ± 2.8
	AIID	21.5 ± 4.1	21.0 ± 4.0	17.7 ± 3.7	17.3 ± 3.6	14.1 ± 3.0	13.7 ± 2.9
	CG	20.0 ± 4.0	19.5 ± 4.1	16.9 ± 3.6	16.6 ± 3.4	13.6 ± 3.2	13.3 ± 3.0
3	HID	23.6 ± 3.8	22.9 ± 3.7	19.6 ± 3.5	18.9 ± 3.3	15.7 ± 2.8	15.2 ± 2.7
	AIID	22.8 ± 3.9	22.1 ± 3.9	18.5 ± 4.6	18.0 ± 3.5	14.8 ± 2.9	14.3 ± 2.8
	CG	20.2 ± 3.9	19.7 ± 3.9	17.0 ± 3.5	16.7 ± 3.6	13.7 ± 3.3	13.4 ± 3.1
4	HID	25.0 ± 3.6	24.3 ± 3.5	20.8 ± 3.3	20.2 ± 3.1	16.6 ± 2.7	16.1 ± 2.6
	AIID	23.6 ± 3.8	22.9 ± 3.8	19.4 ± 3.5	18.8 ± 3.4	15.4 ± 2.8	14.9 ± 2.7
	CG	20.4 ± 3.8	19.9 ± 3.8	17.2 ± 3.5	16.9 ± 3.5	13.9 ± 3.1	13.6 ± 3.0
5	HID	26.4 ± 3.5	25.6 ± 3.4	22.0 ± 3.2	21.3 ± 2.9	17.4 ± 2.6	16.9 ± 2.5
	AIID	24.5 ± 3.7	23.8 ± 3.7	20.3 ± 3.4	19.7 ± 3.3	16.2 ± 2.7	15.8 ± 2.6
	CG	20.7 ± 3.9	20.2 ± 3.9	17.4 ± 3.4	17.1 ± 3.4	14.0 ± 3.0	13.7 ± 3.0
6	HID	27.6 ± 3.3	26.8 ± 3.2	23.2 ± 3.1	22.5 ± 2.8	18.2 ± 2.5	17.6 ± 2.4
	AIID	25.3 ± 3.6	24.5 ± 3.6	21.1 ± 3.3	20.5 ± 3.2	16.8 ± 2.6	16.3 ± 2.5
	CG	20.8 ± 3.8	20.3 ± 3.7	17.6 ± 3.3	17.3 ± 3.3	14.1 ± 3.1	13.8 ± 3.1
7	HID	28.8 ± 3.2	27.9 ± 3.1	24.3 ± 3.0	23.6 ± 2.7	19.0 ± 2.4	18.4 ± 2.3
	AIID	26.2 ± 3.6	25.3 ± 3.5	21.9 ± 3.2	21.3 ± 3.1	17.5 ± 2.6	17.0 ± 2.5
	CG	21.0 ± 3.8	20.5 ± 3.8	17.8 ± 3.3	17.5 ± 3.4	14.3 ± 3.0	14.0 ± 3.0
8	HID	29.9 ± 3.1	29.0 ± 3.0	25.3 ± 2.9	24.6 ± 2.6	19.7 ± 2.3	19.1 ± 2.2
	AIID	27.0 ± 3.5	26.2 ± 3.4	22.6 ± 3.1	22.0 ± 3.0	18.0 ± 2.5	17.5 ± 2.4
	CG	21.2 ± 3.7	20.8 ± 3.7	18.0 ± 3.2	17.7 ± 3.3	14.4 ± 3.0	14.1 ± 2.9

M mean, SD standard deviation, HID Human Idea-Debate group, AIID AI Idea-Debate group, CG Control Group, PSI Problem-Solving Inventory, PSC Problem-Solving Confidence, AAS Approach–Avoidance Style, PC Personal Control

Statistical analysis of problem-solving variables

Three-way repeated-measures ANOVA identified significant main effects across all problem-solving dimensions (Table 5). Problem-Solving Confidence exhibited robust group effects (F = 14.82, p < 0.001, η²p = 0.162), pronounced session effects (F = 41.35, p < 0.001, η²p = 0.438), and significant time effects (F = 9.21, p = 0.003, η²p = 0.061). Approach-Avoidance Style demonstrated significant group differentiation (F = 10.41, p < 0.001, η²p = 0.121), session progression (F = 32.14, p < 0.001, η²p = 0.398), and pre-to-post reductions (F = 7.82, p = 0.006, η²p = 0.054). Personal Control showed group effects (F = 8.64, p < 0.001, η²p = 0.103), session improvements (F = 27.39, p < 0.001, η²p = 0.374), and time-related changes (F = 6.51, p = 0.011, η²p = 0.047).

Table 5.

Three-way repeated-measures analysis of variance results for problem-solving variables: main effects, interaction effects, effect sizes, and Bonferroni-Corrected Post-Hoc Comparisons Across Group (HID, AIID, CG), Session (1–8), and Time (Pre-Post)

Variable	Source	F	P	Partial η²	Bonferroni Post Hoc
PSC	Group	14.82	< 0.001	0.162	HID > AIID (p = 0.031); HID > CG (p < 0.001); AIID > CG (p = 0.042)
	Session	41.35	< 0.001	0.438	S1 < S5 (p < 0.001); S1 < S6 (p < 0.001); S1 < S7 (p < 0.001); S1 < S8 (p < 0.001); S8 highest (p < 0.001)
	Time	9.21	0.003	0.061	Pre > Post (small reduction) (p-values per group: HID p = 0.012; AIID p = 0.045; CG p = 0.038)
	Group × Session	3.94	< 0.001	0.089	HID progression > AIID/CG in S5–S8 (p = 0.021–0.035)
	Group × Time	4.28	0.016	0.034	HID reduction < AIID/CG (p = 0.019–0.041)
	Session × Time	5.67	< 0.001	0.072	Reduction stronger in S6–S8 vs. S1–S4 (p = 0.014–0.029)
	Group × Session × Time	2.41	0.009	0.055	HID effect strongest in S6–S8 (p = 0.021–0.037)
AAS	Group	10.41	< 0.001	0.121	HID > AIID (p = 0.044); HID > CG (p < 0.001); AIID > CG (p = 0.049)
	Session	32.14	< 0.001	0.398	S1 < S4 (p < 0.001); S1 < S5 (p < 0.001); S1 < S6 (p < 0.001); S1 < S7 (p < 0.001); S1 < S8 (p < 0.001); S8 highest (p < 0.001)
	Time	7.82	0.006	0.054	Pre > Post (p-values per group: HID p = 0.021; AIID p = 0.037; CG p = 0.044)
	Group × Session	2.98	0.002	0.067	HID > AIID/CG in S4–S8 (p = 0.025–0.039)
	Group × Time	3.55	0.032	0.029	HID drop smaller vs. AIID/CG (p = 0.028–0.041)
	Session × Time	4.90	< 0.001	0.063	Post decrease increases after S4 (S5–S8, p = 0.018–0.036)
	Group × Session × Time	1.98	0.041	0.038	HID interaction significant in S6–S8 (p = 0.022–0.039)
PC	Group	8.64	< 0.001	0.103	HID > AIID (p = 0.050); HID > CG (p < 0.001); AIID > CG (p = 0.040)
	Session	27.39	< 0.001	0.374	S1 < S5 (p < 0.001); S1 < S6 (p < 0.001); S1 < S7 (p < 0.001); S1 < S8 (p < 0.001); S8 highest (p < 0.001)
	Time	6.51	0.011	0.047	Pre > Post (p-values per group: HID p = 0.028; AIID p = 0.041; CG p = 0.045)
	Group × Session	2.20	0.009	0.054	HID improved progressively S4–S8 (p = 0.031–0.038)
	Group × Time	3.18	0.043	0.027	HID decline minimal vs. AIID/CG (p = 0.034–0.042)
	Session × Time	4.37	< 0.001	0.057	Reduction increases S5–S8 vs. S1–S4 (p = 0.017–0.033)
	Group × Session × Time	1.67	0.072	0.032	Trend toward significance; HID S6–S8 vs. others (p = 0.045–0.050)

F F-Value, P P-Value, η² Partial eta square (Effect Size); HID Human Idea-Debate group; AIID AI Idea-Debate group; CG Control Group; PSC Problem-Solving Confidence, AAS Approach–Avoidance Style, PC Personal Control, S1 Session 1, S4 Session 4, S5 Session 5, S6 Session 6, S7 Session 7, S8 Session 8, Statistical significance was set at α = 0.05 (two-tailed). Values at p < 0.001 are indicated accordingly

Bonferroni post-hoc analyses confirmed consistent HID superiority. For Problem-Solving Confidence, HID exceeded both AIID (p = 0.031) and CG (p < 0.001), with AIID also surpassing CG (p = 0.042). Session comparisons revealed that Sessions 5–8 significantly exceeded Session 1 (p < 0.001). Approach-Avoidance Style exhibited similar hierarchical patterns with HID > AIID (p = 0.044) and HID > CG (p < 0.001). Personal Control demonstrated HID advantages over AIID (p = 0.050) and CG (p < 0.001).

Percentage change calculations indicated systematic pre-to-post-session reductions across all problem-solving subscales, consistent with the PSI’s inverse-scoring structure (Fig. 3). The HID group exhibited the largest reductions, with Problem-Solving Confidence decreasing from − 2.38% in Session 1 to − 3.13% in Session 7. The AIID group demonstrated intermediate reductions, whereas CG changes remained minimal throughout the intervention period.

Fig. 3.

Percentage change in problem-solving inventory subscale scores (problem-solving confidence, approach-avoidance style, personal control) from pre-session to post-session assessments across eight training sessions for Human Idea-Debate (HID), AI-Ideation Delivery (AIID), and Control (CG) Groups

Discussion

The present investigation demonstrates that structured, human-led reflective dialogue is a superior pedagogical approach for enhancing tactical performance and problem-solving competencies in soccer training compared with AI-generated tactical prescriptions or conventional practice methods. Both intervention groups outperformed the control condition across multiple performance domains. However, the HID approach yielded the greatest improvements, as evidenced by statistically significant advantages in Received Ball, Performance Score, and Problem-Solving Confidence. The Efficiency Index showed a particularly dramatic increase in the HID group, rising 70.5% from Session 1 to Session 8, compared with modest gains in the AIID condition and negligible changes in the control conditions. These findings indicate that engaging players in collective tactical analysis, solution articulation, and collaborative reapplication during play produces stronger effects on both tactical execution and cognitive engagement than implementing externally generated algorithmic recommendations.

The superiority of human-mediated dialogue can be explained by the cognitive and social mechanisms activated during collective reflection in soccer-specific contexts [24]. In soccer, shared training through structured verbal exchanges strengthens representations of tactical knowledge and accelerates the convergence of shared mental models [40, 41]. Requiring players to articulate perceptual cues, justify tactical choices, and negotiate shared solutions deepens situational understanding and enhances the capacity to anticipate opponent actions — competencies that are particularly critical in high-speed, space-constrained match environments [ref]. This active reasoning process enhances anticipation of play sequences and coordination of collective actions. It also supports adaptation to emerging constraints during dynamic game situations [18, 40, 42]. The observed Group × Session interactions, wherein HID advantages intensified from Session 5 onward (p = 0.021–0.041), suggest that collaborative sense-making requires temporal scaffolding to achieve maximal efficacy. In soccer, players progressively internalize collaborative reasoning patterns and develop a shared tactical vocabulary, a process that requires repeated exposure to structured discussion cycles before maximal transfer to on-field behavior is achieved [43]. This temporal dimension reflects the gradual construction of collective tactical intelligence that transcends individual cognitive limitations.

In contrast, although AI-generated action plans provided structured, analytically coherent tactical guidance, their prescriptive nature appears to have constrained players’ opportunities to engage in the cognitive struggle necessary for deep tactical learning [44]. The moderate improvements observed in the AIID group suggest that algorithmic recommendations enhance tactical clarity and organizational coherence without replicating the metacognitive and collaborative benefits of human-led discussion. From a pedagogical standpoint, AI systems address the analytical dimension of tactical development but cannot scaffold the iterative reasoning, collective negotiation of meaning, or social accountability that drive deep tactical learning [45, 46]. AI-generated outputs are prescriptive and static relative to players’ evolving affective and cognitive states during training, limiting their capacity to prompt the metacognitive engagement required for durable tactical adaptation [47]. Human dialogue, by contrast, incorporates relational, emotional, and metacognitive elements absent from algorithmic outputs, including idea validation, social ownership of solutions, and sustained psychological investment across training phases [17]. Furthermore, the effectiveness of AI-generated recommendations depends critically on athletes’ pre-existing capacity to interpret, contextualize, and apply data-driven instructions within dynamic game environments [46].

The present findings align with two key empirical contributions. Ashford et al. [42]. Conducted a systematic review demonstrating that player decision-making quality in team sports is determined by perceptual-cognitive skills, cue identification, knowledge retrieval, and option generation, which are cultivated through iterative, reflective practice rather than passive reception of tactical instructions. Steiner et al. [48] showed that interpersonal coordination in interactive team sports depends on the integration of shared mental models with context-specific affordance perception, a process requiring repeated social interaction to consolidate. Together, these findings support the interpretation that the HID-induced improvements observed in the present study arose from the progressive development of precisely these competencies. Research on interactive learning processes emphasizes that peer dialogue promotes deeper cognitive elaboration and facilitates the construction of shared tactical frameworks through the iterative refinement of strategic concepts [18]. The debate-of-ideas methodology specifically promotes active reasoning and exploration of multiple solution pathways, thereby strengthening adaptive decision-making capabilities beyond domain-specific tactical knowledge [43]. Studies of technologically assisted coaching indicate that AI-based tools can support performance through precise, data-driven recommendations; however, their pedagogical impact is constrained by athletes’ pre-existing tactical literacy, digital fluency, and capacity to translate algorithmic prescriptions into context-appropriate on-field actions [15, 22]. When players lack the conceptual frameworks necessary to critically evaluate and adapt AI outputs to dynamic game situations, prescriptive recommendations risk being applied rigidly or disregarded, reducing their practical value [46]. This dependency on individual cognitive readiness represents a fundamental pedagogical limitation of AI-based tactical guidance relative to collaborative dialogue, in which meaning is co-constructed at a level calibrated to the group’s current tactical understanding [17]. The current results reinforce these patterns by demonstrating that although AI instruction improves tactical execution, its effects remain attenuated relative to human-mediated discussions, which engage broader cognitive, social, and interpretive capacities essential for dynamic soccer contexts.

The dramatic improvement in the Efficiency Index within the HID group merits particular attention. This composite metric, reflecting the ratio of offensive effectiveness to ball losses, captures the quality of tactical decision-making under competitive pressure. The 70.5% enhancement substantially exceeds the moderate gains observed in the AIID group and the negligible changes observed in the control group. This finding suggests that collaborative dialogue specifically enhances players’ capacity to identify and execute high-value tactical actions while minimizing errors [8, 40, 41]. Theoretical explanations invoke the development of anticipatory schemas through repeated collective analysis of tactical cues, opponent positioning, and spatial affordances. Players who systematically discuss tactical problems develop refined perceptual attunement to critical environmental information, thereby enabling more effective real-time decision-making in subsequent phases of play [43].

The significant improvements across all PSI subscales (Problem-Solving Confidence, Approach-Avoidance Style, Personal Control) provide convergent evidence for the cognitive benefits of structured reflective dialogue. The inverse scoring structure of the PSI means that observed pre-to-post session reductions indicate enhanced cognitive self-efficacy and more adaptive problem-solving orientations. These changes extend beyond domain-specific tactical knowledge to encompass general cognitive competencies applicable across diverse problem-solving contexts [49]. The cultivation of these transferable skills represents a substantial pedagogical advantage, suggesting that collaborative tactical learning serves dual functions: immediate enhancement of sport-specific performance and broader development of cognitive capabilities. The minimal improvements in the control group confirm that tactical learning requires explicit reflection, verbalization of strategic concepts, and social negotiation of meaning [40]. Collaborative dialogue specifically enables players to articulate game perceptions, challenge interpretive assumptions, and co-construct tactical frameworks grounded in shared on-field experience, processes documented in soccer contexts to improve game performance and collective involvement [43]. Unguided exposure permits habitual response patterns to persist unchallenged, precluding the adaptive recalibration that characterizes expert team cognition [42].

Several methodological constraints warrant acknowledgment. The four-week intervention period may not capture long-term development or durability of tactical and cognitive gains. The sample comprised exclusively male adolescent players from a restricted competitive level and geographic region, limiting generalizability to female athletes, elite professionals, and culturally diverse populations. AI-generated recommendations depended on model accuracy and could not reflect real-time tactical variability in authentic match conditions. Individual differences in tactical literacy, verbal fluency, and digital fluency, not systematically assessed, may have differentially moderated responsiveness to each intervention modality. The single-blind design mitigated observer bias (κ = 0.92) but could not eliminate participant expectancy effects [50]. Physiological load equivalence was verified through heart-rate monitoring, though potential differential cognitive demands across conditions remain unquantified. The AIID delivery relied on a single shared screen, and although players were instructed not to discuss AI-generated content, informal inter-player communication during the reading interval cannot be excluded. This potential social contamination of the AIID condition may have marginally attenuated the contrast with the HID group and should be controlled through individual-device delivery or physical separation in future studies [22].

Practical Implications

The theoretical and practical implications of these findings warrant careful consideration. Conceptually, the results highlight the central role of socially mediated cognition in shaping tactical intelligence, demonstrating that collaborative reasoning processes contribute more effectively to adaptive decision-making than externally delivered tactical inputs. The HID approach offers coaches a cost-effective, high-impact way to embed reflective dialogue into training routines, enabling players to co-construct tactical solutions, strengthen decision-making autonomy, and deepen game understanding. The positive but comparatively smaller effects in the AIID group indicate that AI-based support can serve valuable complementary functions in structuring tactical feedback and enhancing analytical depth when integrated alongside human facilitation rather than as a replacement.

Coaches integrating AI-based tools into tactical training should account for several documented limitations. AI-generated recommendations are static, context-insensitive, and contingent on the quality of player-supplied problem descriptions; they cannot adapt in real time to players’ cognitive or affective states [46]. Players with limited tactical literacy or digital fluency may be unable to meaningfully interpret or apply AI outputs, reducing their practical benefit [51]. Coaches should therefore use AI guidance as an analytical supplement within human-facilitated debriefing rather than as a primary instructional mechanism. Structured validation sessions, in which players critically compare AI-generated solutions against their own game perceptions, may help bridge the gap between algorithmic prescriptions and contextually grounded tactical understanding [8, 18]. Athletes should be explicitly informed that AI recommendations represent probabilistic generalizations and require situational judgment to apply appropriately in dynamic match conditions [46].

Future research directions

Extended intervention durations should be examined to determine whether cognitive and tactical benefits intensify, plateau, or attenuate over prolonged training periods. Hybrid models integrating AI-generated analytical outputs into HID sessions warrant investigation to determine whether combining approaches leverages their respective strengths. Studies involving female players, elite professionals, and prepubescent cohorts are necessary to evaluate cross-population applicability. Advanced AI systems with real-time adaptive responsiveness should be compared against static large-language-model guidance. Qualitative analysis of HID communication dynamics, including leadership patterns, participation equity, and argument quality, would illuminate the specific mechanisms linking collective reasoning to tactical performance. Neuroimaging investigations could differentiate the neural substrates underlying collaborative versus algorithmically guided tactical processing. Extension to other invasion team sports would establish generalizability.

Conclusion

This investigation establishes that structured, human-led, reflective dialogue is a more effective catalyst for improving tactical behavior and problem-solving skills in soccer training than AI-generated action plans or conventional methods. The HID approach consistently generated the largest improvements across tactical indicators and cognitive variables, underscoring the importance of socially mediated reasoning, shared mental models, and collaborative interpretation in shaping adaptive decision-making. Although AI support contributed positively through structured and analytically coherent recommendations, it did not replicate the depth of cognitive engagement elicited by human debate. Optimal training frameworks should embed reflective communication as the primary pedagogical vehicle while incorporating AI systems as complementary analytical resources. Practitioners should exercise caution when deploying AI-generated tactical guidance as a standalone intervention, given its dependence on players’ tactical literacy, its context-insensitivity relative to real-time game dynamics, and its inability to replicate the metacognitive and relational processes essential to collective tactical development. AI tools should be validated against player comprehension and on-field applicability before instructional integration.

Abbreviations

AI

Artificial Intelligence

AIID

AI-Ideation Delivery

AAS

Approach-Avoidance Style

ANOVA

Analysis of Variance

BMI

Body Mass Index

CB

Conquered Ball

CG

Control Group

COD

Change of Direction

EI

Efficiency Index

HID

Human Idea-Debate

HR

Heart Rate

ICC

Intraclass Correlation Coefficient

ISAK

International Society for the Advancement of Kinanthropometry

ISSEP-KEF

Higher Institute of Sport and Physical Education of El Kef

LB

Lost Ball

NB

Neutral Ball

OB

Offensive Ball

PC

Personal Control

PSC

Problem-Solving Confidence

PSI

Problem-Solving Inventory

PS

Performance Score

RB

Received Ball

RPE

Rating of Perceived Exertion

SS

Successful Shot

SSG

Small-Sided Games

TSAP

Team Sport Assessment Procedure

VP

Volume of Play

Authors’ contributions

All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by HS, MMB, and WD. The first draft of the manuscript was written by HS, and all authors commented on subsequent versions. All authors read and approved the final manuscript.

Funding

No sources of funding were used to assist in the preparation of this article.

Data availability

The datasets analyzed during the current study are available from the corresponding author upon reasonable request. Personal identifiers were removed from all datasets, and encrypted digital storage protocols were used to ensure data security throughout the study.

Declarations

Ethics approval and consent to participate

The research protocol received ethical approval from the Institutional Review Board of the Higher Institute of Sport and Physical Education of El Kef (ISSEP-KEF), University of Jendouba, Tunisia (approval code: C-0035/2023), ensuring compliance with the principles of the Declaration of Helsinki for human research. All participants and their legal guardians provided written informed consent following a comprehensive briefing on study objectives, procedures, potential risks, and data confidentiality measures. Participation was voluntary, with the explicit right to withdraw at any time without penalty.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.