AI Chatbots and Cognitive Control: Enhancing Executive Functions Through Chatbot Interactions: A Systematic Review

Synthesis: Pergantis, Bamicha, Skianis and Drigas (2025) review how AI chatbots (conversational agents) have been used to train executive functions — working memory, inhibition, cognitive flexibility, planning and attention — and what the evidence says about whether those gains last. Following PRISMA 2020, they screened 115 records from Scopus, Web of Science, PubMed and Google Scholar and retained 10 studies covering 1606 participants aged 6–75+. The review reports positive effects across most of the included studies, but almost all were immediate or task-specific; only two studies reported one-month follow-up, and the authors conclude that chatbot-based executive-function training is still at a primary stage and difficult to generalize or reproduce.

What the Review Asked and How It Searched

The review asks which conversational chatbots were used for executive-function training, what outcomes they produced, and how long improvements lasted. The protocol was registered on the Open Science Framework (osf.io/jxmg4) and follows PRISMA 2020. Scopus, Web of Science, PubMed and complementary Google Scholar were searched for peer-reviewed, English-language, open-access studies published from 2021 onward; the authors describe the coverage as January 2021 to 2024. Keywords combined "AI chatbots", "chatbots" and "conversational agents" with "executive functions", "attention", "memory", "cognitive load" and "emotional regulation", and a snowball search supplemented the database results.

Seven inclusion criteria define the corpus. Studies had to be experimental, observational or mixed methods; use an AI-based chatbot or conversational agent to support executive functions or conditions that affect them, named as stress, anxiety, depression, memory, attention, cognitive load and behavioral change; involve general or neurodevelopmental and neurological populations; be peer-reviewed, in English, and available in full text. Six exclusion criteria removed pre-2021 work, reviews, non-AI or non-interactive tools, studies not targeting executive skills, non-English studies and paywalled studies. The flow reports 115 records screened, 8 duplicates removed in Mendeley Desktop 1.19.8, 73 excluded at title and abstract, 42 texts sought, 4 not retrievable, 38 assessed in full text and 28 excluded, leaving 10. Two reviewers judged eligibility independently, with two further reviewers resolving disagreements.

The Ten Included Studies

The 10 studies are organized into four age bands — children and adolescents, university students and young adults, adults, and older adults — spanning 6–75+. Three came from the Republic of Korea, two from the USA, one from Sweden, one from the UK and one was multinational (USA, UK and Canada); two did not specify a country. Designs vary as much as settings: a cross-sectional survey, a pilot randomized controlled trial, a single-subject AB design, a feasibility study, a randomized controlled trial, a mixed-methods exploratory study, a longitudinal intervention study, an experimental comparative study, a structural equation modeling analysis and an exploratory study.

Sample sizes range from 3 students aged 11–13 to 744 adolescents aged 12–19 (385 in Study 1 and 359 in Study 2). The remaining studies tested 46 adults aged 19–60, 27 children aged 6–12, 132 children aged 6–7, 47 adults aged 18–74, 32 older adults aged 60+, 256 adults aged 19–40, 193 chatbot users and 126 university students aged 18–32. Three studies focused on ADHD, one on age-related memory impairment, and six on general populations.

Chatbot technologies differ in kind, not branding: a decision-tree stress-management bot (Popbots), a psychoeducational cognitive-behavioral bot (Tobaki), a gamified cognitive-training bot (NUROW), a goal-directed bot for children with ADHD (ForME), general-purpose generative AI including ChatGPT, ChatGPT-4 and Copy.ai, a cognitive-training bot for older adults, and a service-oriented bot for cognitive-load management. Targets follow the same spread: working memory, attention, inhibition, planning, cognitive flexibility, emotional regulation, task efficiency and perceived cognitive load. Drawing on developmental psychology and neuropsychology, the review treats executive function as a multidimensional construct with "hot" emotional and "cool" cognitive dimensions, reporting gains in both. It also ties executive skills to metacognition and self-regulated learning, consistent with a cognitive psychology framing in which the skills improve through repeated practice.

How Long Did the Gains Last?

Durability is where the evidence thins out. Klarin et al. (2024) found that task-efficiency and working-memory improvements lasted only during chatbot-assisted assignments. Fabio et al. (2024) found cognitive-flexibility and reasoning gains were task-specific and limited to the duration of chatbot interaction, and Koivisto and Grassini (2023) observed creativity and divergent-thinking gains only during task performance with the chatbot. Park et al. (2024) sustained behavioral-regulation and task-adherence improvements across a six-week intervention but reported no post-intervention follow-up. Sun et al. (2023) found attention and behavioral-control gains during a four-week intervention and did not measure long-term impact. Rostami and Abadi (2023) found working-memory effects limited to the immediate task, with no evidence of retention or transfer, and Chou and Hsu (2021) found cognitive-load reduction only during active use.

Two studies reach further. Jang et al. (2021) reported reduced impulsivity, improved attention and improved emotional regulation maintained for one month post-intervention, and Kim et al. (2024) reported episodic-memory improvement and reduced anxiety sustained for up to one month. No included study followed participants beyond one month, making one month the ceiling of the evidence rather than a plateau.

Key Findings

1. Ten studies survived screening from 115 records. The flow records 115 records, 8 duplicates, 73 title-and-abstract exclusions, 42 texts sought, 38 full texts assessed and 28 full-text exclusions, ending at 10 studies covering 1606 participants aged 6–75+.
2. The evidence base is geographically narrow. Three studies came from the Republic of Korea, two from the USA, one each from Sweden and the UK, one was multinational across the USA, UK and Canada, and two did not specify a country; 744 of the 1606 participants come from a single Swedish survey of adolescents.
3. Chatbot type varies more than the generic label suggests. The corpus includes decision-tree, psychoeducational CBT, gamified, goal-directed, cognitive-training, service-oriented and general-purpose generative AI bots, and the largest test of a general LLM chatbot is a cross-sectional survey of 744 Swedish adolescents.
4. The most consistent gains concern attention, behavioral regulation and planning in ADHD samples. Park et al. (2024) tested 132 children aged 6–7 over six weeks, Sun et al. (2023) tested 27 children aged 6–12, and Jang et al. (2021) tested 46 adults aged 19–60 with one-month retention.
5. Working-memory evidence is thinner and mixed. Klarin et al. (2024) reported improved task efficiency and working memory, while Rostami and Abadi (2023), with 3 students aged 11–13, found mixed working-memory results and a risk of over-reliance on the chatbot.
6. Durability, not direction, is the weak point. Only Jang et al. (2021) and Kim et al. (2024) reported one-month post-intervention effects; the other eight reported immediate, task-specific or during-intervention effects, and no study reported a longer follow-up.
7. The authors' own verdict is cautious. They describe chatbot use for executive-skills training as "at a primary stage" that is still difficult to generalize and reproduce, and they call for a unified framework, better study designs, more diverse populations, larger samples and longitudinal follow-up.

What this means for practice

• Treat chatbot effects as task-level support, not developmental change. The review's own duration findings show gains that track the chatbot interaction; plan for fading support and transfer activities rather than assuming durable executive-function growth.
• Start with structured, goal-directed bots where the evidence is strongest. The ADHD cluster — ForME, NUROW and Tobaki — targets behavioral regulation, attention and planning with predefined dialogs, the most consistent body of evidence here.
• Do not assume open-ended LLM chat builds working memory. The largest LLM study is a cross-sectional survey, and the only experimental LLM working-memory study had 3 participants and mixed results plus an over-reliance risk.
• Measure executive function directly and follow up past one month. The corpus relies on BRIEF-2, BRIEF, CANTAB, N-Back, CAARS and self-report stress scales; programs should pair a direct EF measure with a delayed follow-up instead of stopping at post-test.
• Build privacy, transparency and bias review into procurement. The authors flag privacy violation, algorithmic bias, hallucinations and reduced human interaction as design risks, and no safeguard is empirically evaluated.

Limitations

• No formal risk-of-bias or methodological-quality instrument is reported; the synthesis is a narrative "critical overview" rather than a scored appraisal, so confidence in individual studies rests on the authors' prose judgment.
• The 10 studies use incompatible designs, comparators and outcome measures, which precludes meta-analysis; "most studies were positive" is effectively a vote count across heterogeneous evidence.
• Several samples are small (3, 27, 32, 46 and 47 participants) and only two studies reported one-month follow-up, so the durability question is answered by a maximum of one month of data.
• The inclusion criteria fold stress, anxiety, depression and cognitive load into the executive-function construct, and some outcomes (PHQ-4 stress, GDS and GAI scales, semantic-distance creativity ratings) are proxies rather than direct executive-function tests.
• ADHD and age-related memory impairment are overrepresented relative to general education, two studies did not specify a country, and the authors acknowledge that home, clinic and gamified settings were controlled and do not generalize to real-world conditions.
• The search was restricted to English, peer-reviewed, open-access studies from 2021 to 2024 across four databases plus Google Scholar, with no gray literature or meta-analysis, so publication bias and language bias are unaddressed.

Citation

Pergantis, P., Bamicha, V., Skianis, C., & Drigas, A. (2025). AI chatbots and cognitive control: Enhancing executive functions through chatbot interactions: A systematic review. Brain Sciences, 15(1), 47. https://doi.org/10.3390/brainsci15010047