AI in Gifted Education: Smarter Screening, AI Tutors and Twice-Exceptional Learners, and Why the Research Urges Caution

0:00 / 30:29Voice: Guy
Chapters (10)
Primary school children in matching blue tracksuits sit on the wooden floor of a Japanese classroom, most of them seen from behind, listening to a man in a grey work jacket at the front, next to a green chalkboard with Japanese writing
Photo: Nhrdsk, “金屋石を使っての授業の様子(2014.12.3)” (“Lesson using Kanaya stone”, 2014), CC BY-SA 4.0, via Wikimedia Commons – https://commons.wikimedia.org/wiki/File:%E9%87%91%E5%B1%8B%E7%9F%B3%E3%82%92%E4%BD%BF%E3%81%A3%E3%81%A6%E3%81%AE%E6%8E%88%E6%A5%AD%E3%81%AE%E6%A7%98%E5%AD%90%EF%BC%882014.12.3%EF%BC%89.JPG (resized) · License: CC BY-SA 4.0 (Creative Commons Attribution-ShareAlike 4.0 International)

Gifted education has always had a resource problem. A child who finishes the maths unit in a week needs harder material, faster pacing and, ideally, a mentor who knows the field. Most teachers do not have time to build that for one or two students. Generative artificial intelligence (AI) looks, at first sight, like the perfect answer: a tireless tutor that can explain calculus to a ten-year-old at midnight.

Since 2023, researchers in gifted education have started to test that promise. The result, so far, is a mixed picture. AI can save teachers time, help some students write and think, and may help find talented children who are now missed. But the research base is small, mostly from 2025 and 2026, and several larger studies from general education show that AI can raise homework scores while lowering real learning. This post explains what is known, what is only hoped for, and what experts warn about.

How we checked: we read original papers, abstracts and publisher pages where possible: Gifted Child Quarterly, Gifted Child Today, Journal of Advanced Academics, Gifted Education International, Frontiers in Psychology and Frontiers in Education, Education Sciences, PNAS, Scientific Reports, and working papers from NBER, CEPR and the World Bank. We also used pages from NAGC, the World Council for Gifted and Talented Children (WCGTC), ECHA, the OECD, UNESCO and the EU AI Act Service Desk. Some items come only from search-result summaries, conference pages or community posts. We say so where that is the case. We did not test any AI tools with students.

Key terms:

  • Gifted and talented: students with high ability or high potential in one or more areas, such as maths, language, science, arts or leadership. Definitions differ a lot between countries and school systems.
  • Twice-exceptional (2e): students who are gifted and have a disability or learning difference, such as ADHD, autism, dyslexia or dysgraphia. The two can hide each other (“masking”).
  • Identification / screening: the process of deciding which students get gifted services, usually through tests, teacher ratings, grades or portfolios.
  • Universal screening: testing all students instead of only those referred by teachers or parents.
  • Acceleration and enrichment: moving faster through content (acceleration), or going deeper and wider (enrichment).
  • Machine learning / generative AI: software that learns patterns from data; generative AI (such as ChatGPT or Gemini) uses this to produce text, images or code.
  • Cognitive offloading: letting a tool do mental work for you. Useful in small doses, harmful when it replaces the thinking you are meant to learn.

1. The state of research: early, small and growing fast

The clearest overview so far is a scoping review published in Frontiers in Psychology on 29 July 2026 by Omar Alsamani and Yasir Alsamiri. They searched four databases up to May 2026 and found just 26 eligible studies on AI in gifted and talented education. Nineteen of them were published in 2025 and four in 2026. Nine came from the United States and seven from Türkiye; the review found none from Sub-Saharan Africa, Latin America or South Asia (Frontiers in Psychology).

The authors see two main strands:

  1. Teaching and development: generative AI, AI tutors and AI-made materials for differentiation, writing, questioning, creativity and mentoring.
  2. Assessment: machine learning used to identify gifted and twice-exceptional students.

Their overall verdict is cautious. The field, they write, “is still describing possibilities more than demonstrating outcomes.” Most studies are conceptual or small, and longitudinal and comparative studies are almost missing. They propose a principle they call “supervised augmentation”: AI may extend what teachers can do, but a qualified human keeps the final say on any important decision.

Much early work appeared in a few journal issues: Gifted Child Today (April 2025), whose editorial cites Del Siegle’s idea that AI can support the “three legs” of talent development (acceleration, depth and complexity, interest-based learning) (editorial), and a 2025 themed issue of the Journal of Advanced Academics (volume 36, issue 4). Gifted Education International has an open special issue, edited by Robert J. Sternberg, on “The Future of Gifted Identification and Education in the Age of AI”; the submission deadline was 1 October 2026 (GEI call for papers, PDF).

2. Identification and screening: can algorithms find hidden talent?

Why identification is a problem in the first place

Who ends up in a gifted programme depends a lot on who gets noticed. Research before the AI boom showed this clearly. In a large US school district, introducing universal screening raised the share of low-income, Black and Hispanic students identified as gifted, compared with the old system based on teacher and parent referrals (Card & Giuliano, PNAS 2016). The 2026 scoping review lists biased teacher referral and narrow test-based criteria as main reasons why some groups are underidentified (Frontiers in Psychology).

This is the gap AI promises to fill: more data, fewer subjective judgements, and patterns that humans overlook.

What the recent studies tried

  • Teacher rating scales in South Korea. Lee, Cho and Walsh (2025, Journal of Advanced Academics) applied machine-learning models (Gaussian mixture models, support vector machines and a neural network) to HOPE teacher rating scale responses. Fifty-five teachers rated third- to sixth-grade students in Seoul. The models reported very high accuracy (above 99% in cross-validation). But gifted students were rare in the data, and the authors also looked at signs of teacher rating bias. The scoping review notes that the study found challenges linked to underrepresented students and rating bias across socioeconomic and ethnic groups (JAA, Lee et al. 2025).
  • Twice-exceptional identification in Türkiye. Atmaca and Baloğlu developed the Cognitive Assessment Battery for Twice Exceptionality (2eCAB), published in Gifted Child Quarterly. It measures nonverbal ability, memory, rapid naming and pseudoword reading. The authors used a decision-tree method (CART) to sort 565 Turkish-speaking primary school students into typically developing, gifted, twice-exceptional and specific-learning-disability groups (Gifted Child Quarterly; Hacettepe University record).
  • Beyond school. A 2025 study at Troy University (US) used academic records of 829 computer science students and identified “talented” students with about 80% accuracy (JAA 2025).

AI is also entering creativity assessment, which many identification systems use. The Open Creativity Scoring project at the University of Denver offers “Ocsai”, a fine-tuned language model that scores the originality of answers to divergent-thinking tasks such as “list unusual uses for a brick”. The developers report correlations up to about r = .81 with human raters (Organisciak et al., ERIC; Open Creativity Scoring).

The catch: algorithms learn yesterday’s bias

Every review we read warns about the same thing. A model trained on past referrals, ratings and placements learns the patterns in that data, including its unfairness. The 2026 scoping review puts it bluntly: “algorithmic identification is not automatically equitable identification.” It recommends that any pilot should validate models on local, diverse samples, report results by subgroup, document which data are used, keep human review for every consequential decision and monitor for disparate impact (Frontiers in Psychology). Very high overall accuracy can also hide poor results for the small group that matters most.

In Europe, there is also a legal angle. Under the EU AI Act, AI systems that decide on access or admission to education, assign people to institutions, or evaluate learning outcomes are listed as high-risk in Annex III, with duties such as human oversight; some narrow, preparatory uses may be exempt (AI Act Service Desk, Annex III; Education use case). Whether a given gifted-screening tool falls under these rules depends on how it is used. We are not giving legal advice here.

Does IQ still mean anything?

A more radical debate is about what should be identified. Sternberg argues that AI already beats humans on many tasks used in IQ tests. His GEI call for papers asks: “Should we be identifying children as gifted on the basis of tests that measure skills in which AI performance already beats human performance?” He suggests looking instead at critical thinking, moral intelligence, “paradigm-defying creativity”, resilience, epistemic humility and AI literacy (GEI call, PDF; GEI article). This is a position paper, not an empirical finding, and many researchers in the field still use cognitive tests as one part of a broader identification process.

3. Personalised acceleration and enrichment

The most common use of AI in gifted education so far is not student-facing at all. It is teachers using AI to prepare differentiated material faster.

  • Personalised playscripts. In Project Brilliance, an after-school theatre programme supported by the US Javits gifted education grant programme, teachers used ChatGPT to tailor playscripts to each third- and fourth-grader’s interests, reading level and creative ideas. In this action research study, students were more engaged and showed growth in fluency, comprehension and creativity (Gifted Child Today editorial).
  • Mentoring and independent research. Guilbault, Wang and McCormick (2025) describe case studies in which ChatGPT supported differentiation, independent research, feedback and “virtual mentorship” for secondary gifted students, aligned with NAGC standards. They stress prompt-writing skills and critical thinking (Gifted Child Today).
  • Better questions. Akdeniz, Clark and Roberts (2025) asked ChatGPT-4, Google Bard (now Gemini) and an experienced teacher to write questions for each level of Bloom’s taxonomy. Fifteen gifted-education experts rated them. AI questions were well aligned at lower levels, but human questions were better at the higher, more complex levels (Journal of Advanced Academics). In other words: AI can save time on routine questions, but the challenging ones still need an expert.

NAGC, the US National Association for Gifted Children, has published blog posts and a professional-learning module on AI. The module says generative AI “can be a helpful tool” for differentiated instruction when paired with professional expertise, that all output “must be carefully examined for cultural bias and accuracy,” and that AI “cannot replace the personal and professional expertise of a human teacher” (NAGC learning module; NAGC blog). We did not find a formal NAGC position statement dedicated to AI on its position papers page; this may change.

4. AI tutors: the evidence from gifted programmes and beyond

A tutor built for advanced learners

The best-documented gifted-specific trial comes from the Johns Hopkins Center for Talented Youth (CTY). In an online course called “Diagnosis: Be the Doctor”, one class of 22 middle and high school students had access to a GPT-4o chatbot (“Dr. Smith”) designed to ask Socratic questions instead of giving answers. A second class with the same teacher had no chatbot. There was no significant difference in final assessment scores. Many students used the bot to look up facts or ask for the answer, not for the back-and-forth reasoning it was built for. The team concludes that students and teachers need structured training in how to use such tools (Johns Hopkins Hub, 22 Oct 2025; Journal of Advanced Academics).

In a small interview study, seven gifted students at a rural Texas high school valued their AI tutor’s fast, in-depth answers, but mentioned inaccuracies and worries about over-reliance (Sam Houston State University dissertation, 2025). A 2025 framework by Ecker, Eckert and Cummings maps AI tutoring onto affect, content, process and product (Journal of Advanced Academics), but it is conceptual and untested (Tirri & Tirri, GEI 2026).

What large studies in general education show

Because gifted-specific studies are so small, researchers often look at general education. Here the evidence is larger and more sobering:

  • Turkish high-school maths (PNAS, 2025). In a field experiment with nearly 1,000 students, access to standard GPT-4 improved practice scores by 48%, but students then scored 17% worse on an exam without AI than the control group. A “GPT Tutor” version with guardrails (hints instead of answers) improved practice scores by 127% and largely avoided the harm (Bastani et al., PNAS).
  • Chinese secondary schools (CEPR, 2026). Using 30 months of data on 26,811 students in grades 7–12, economists found that AI adoption raised homework scores by 18% and cut homework time by 30%, but lowered monthly closed-book exam scores by 20% within six months. Entrance-exam scores fell by 18% and 24%. The losses were “especially large for junior students, high-achieving students, and boys,” and concentrated among students whose behaviour suggests outsourcing homework (Strömberg, Lei & Wu, CEPR DP21577 via RePEc). This is a discussion paper, not yet peer reviewed.
  • Khanmigo in Tennessee (NBER, 2026). In a two-year randomised trial in 18 middle schools, Khan Academy with its AI tutor Khanmigo (set to coach, not give answers) raised maths achievement by about 0.06 to 0.08 standard deviations per year, similar to Khan Academy without AI. Students rarely engaged: the median student messaged the tutor in only 17% of exercise sessions where they made a mistake (Oreopoulos & Low, NBER w35620).
  • Positive results do exist. A Harvard randomised trial with 194 undergraduate physics students found that a carefully designed AI tutor produced larger learning gains than an in-class active-learning lesson (Kestin et al., Scientific Reports 2025). A World Bank trial in Nigeria found gains of about 0.31 standard deviations after six weeks of teacher-guided, after-school English tutoring with GPT-4 (World Bank working paper).

The OECD Digital Education Outlook 2026 sums up the pattern: generative AI can scale personalised tutoring, but “overreliance on generative AI tools that provide direct answers can reduce students’ active engagement, improving task performance without corresponding learning gains.” The report calls for educational AI designed with teachers, so that teachers can monitor and shape how students use it (OECD Digital Education Outlook 2026, PDF).

For gifted students, the lesson is uncomfortable. Being fast and capable does not protect against offloading. In the Chinese data, high achievers were among those who lost most.

Do gifted students use AI differently?

Some studies suggest they do, at least in self-reports. In Taiwan, a survey of 484 junior high students (197 gifted) found that gifted students reported lower use of AI for scientific inquiry tasks, but higher AI literacy and more confidence in using AI safely. Almost all students had used ChatGPT (Li, Kuo & Wu, Education Sciences 2025). In Türkiye, a study of nine gifted university students found both benefits (efficiency, incidental learning) and risks (integrity, originality, dependence) in their ChatGPT-assisted essay writing (Hasanbaşoğlu & Baloglu, 2026).

5. Mentoring: can a chatbot replace a mentor?

Mentoring is central to talent development models. At ECHA 2026 in Dublin, one attendee’s report notes that mentors have a central place in the revised Talent Development Megamodel, where they spark interest, open networks and act as intellectual sparring partners (Hoogbloeier conference report).

Some authors see AI as a “virtual mentor” that gives gifted students feedback and career guidance when no human expert is nearby (Guilbault et al., Gifted Child Today). Others are sceptical. In a 2026 article in Gifted Education International, Henry Tirri (Aalto University, a computer scientist and former Nokia CTO) and Kirsi Tirri (University of Helsinki, former ECHA president) argue that the key question “is not whether AI helps. It is which cognitive work stays with the learner.” They warn that when AI writes the essay, the product no longer shows how a student thinks. That removes the main tool a mentor uses to judge a young person’s development. Their recommendation is to put AI “in the teacher’s hands rather than the learner’s”, for example as an assistant for tutors (Tirri & Tirri, GEI 2026). Their list of capacities worth teaching includes problem-finding, cognitive autonomy and self-regulation.

6. Twice-exceptional learners: high hopes, very thin evidence

Twice-exceptional students might benefit most from adaptive technology, because they need challenge and support at the same time. A text-to-speech tool or a writing scaffold can remove a barrier, while advanced content keeps them engaged.

The research, however, is tiny. A scoping review in Frontiers in Psychology (2026) searched for empirical studies on AI and 2e learners from 2020 to August 2026 and found only two that met its criteria, both from the same 2025 journal issue (Frontiers in Psychology, 2e scoping review):

  • Human–AI hybrid writing tutoring in Iran. Twelve twice-exceptional bilingual adolescents with motor dysgraphia and ADHD took part in 32 sessions combining an intelligent tutoring system and a human instructor. The authors reported gains in handwriting fluency, legibility, orthographic coding and composition (Mahmoudi-Dehaki & Nasr-Esfahani, 2025, summarised in the review).
  • Rehearsing pedagogy with “synthetic” students. Ronksley-Pavia, Ronksley-Pavia and Bigum (Australia) used invented but evidence-informed learner profiles to test how generative AI could help teachers plan for twice- and multi-exceptional students. They found some promise, but also inconsistent, sometimes biased outputs, and stress the need for human oversight (Journal of Advanced Academics).

The 2e review points out an important open question: no study has yet checked whether AI systems that adapt to a student’s visible performance might reinforce masking. If a disability hides a student’s ability, an adaptive system may simply give easier material and never discover the strength. Teacher-education programmes are beginning to train future teachers to use AI to design both scaffolds and enrichment for 2e learners, while checking outputs critically (UCF teaching repository, Hughes 2025).

7. Competitions, olympiads and the new cheating problem

Competitions have long been a pathway for gifted students. AI changes them in two ways.

First, AI itself now competes at the top level. In July 2025, an advanced version of Google DeepMind’s Gemini Deep Think was officially graded at gold-medal standard at the International Mathematical Olympiad, scoring 35 of 42 points; OpenAI said its experimental model reached the same score, graded by former medallists rather than through official entry (Google DeepMind; Reuters). Some mathematicians questioned how comparable these results are to human performance (Scientific American). For gifted education, this raises Sternberg’s question again: if machines solve olympiad problems, what exactly are we training and selecting for?

Second, online competitions face AI cheating. The competitive programming platform Codeforces introduced an official rule in September 2024 that restricts AI use in rated rounds: AI may translate problem statements or give minor syntax help, but not generate solution logic (Codeforces rule). Community members have since claimed that a noticeable share of top-ranked accounts use AI; one widely discussed post claimed “more than 4%” of top ranks were AI cheaters (Codeforces community post). We could not verify this figure; it comes from a user post, not from the platform. In-person olympiads like the IMO ban computers and communication devices for contestants (IMO General Regulations, PDF), which is one reason supervised, offline formats remain important.

There are also new competitions about AI. In the United States, the government-backed Presidential AI Challenge invites K–12 students to use AI to solve community problems (participation guidebook, PDF; timeline). In these contests, AI is not the threat but the tool.

8. Risks: over-reliance, cheating, bias and the “sameness” of ideas

Across the literature, the same risks come up again and again:

  • Over-reliance and “cognitive debt”. Sternberg’s 2026 essay in Frontiers in Education argues that whether AI increases or decreases giftedness “is not preordained, but rather reflects a choice.” Citing studies on cognitive offloading, he warns that “the greatest risk, quite simply, is that AI makes us ‘stupid’” (Sternberg, Frontiers in Education 2026). Some studies he cites are small preprints.
  • Less diverse ideas. A Science Advances study cited by Sternberg found AI-assisted stories were rated more creative but were more similar to each other (Doshi & Hauser, 2024).
  • Cheating and the “broken artifact”. When AI can produce a polished essay, the essay no longer proves what the student can do. Tirri and Tirri propose assessing the process (drafts, predictions before using AI, final unaided tasks) instead of only the product (GEI 2026).
  • Bias and privacy. Biased data can distort both identification and AI-made content, and these tools may collect sensitive data on children. The scoping review calls for consent, data governance and appeal procedures for families (Frontiers in Psychology).
  • A new equity gap. Tirri and Tirri suggest, as an untested hypothesis, that able students may build much better “AI partners” than their peers, so gaps could widen in ways that look like differences in ability (GEI 2026).

9. What the international organisations say

  • WCGTC. The 2025 World Conference in Braga, Portugal (29 July–2 August) included a keynote on “Gifted Education in the Age of AI” by Dr. Dai of the University at Albany, asking whether gifted education could become “obsolete and irrelevant with the rise of so-called super-intelligence” or gain added significance (WCGTC keynotes).
  • ECHA. The European Council for High Ability held its 2026 conference in Dublin, with AI and technology among the conference themes (ECHA 2026). One attendee found it “refreshing” that ECHA 2026 “was not an AI conference disguised as a conference on giftedness,” and noted a broader shift from labels toward talent development across the lifespan. The next ECHA conference is planned for Leuven, Belgium, on 9–12 August 2028 (Hoogbloeier report).
  • NAGC. Besides the module above, NAGC’s 2025 convention included a session on a “ThinkAI Framework” for AI literacy among gifted learners (NAGC25 session).
  • OECD and UNESCO. Neither organisation has, as far as we found, specific guidance on AI for gifted learners. Their general guidance still applies: the OECD warns against answer-giving tools that replace effort (OECD 2026), and UNESCO’s AI Competency Framework for Students sets out 12 competencies, from a human-centred mindset and ethics to AI techniques and system design (UNESCO).

10. A brief look at AI in education in general

Gifted education is a small part of a much bigger shift. According to the OECD’s 2024 Teaching and Learning International Survey (TALIS), cited in the Digital Education Outlook 2026, 37% of teachers already use generative AI for work tasks such as summarising topics or planning lessons. The same report cites a 31% reduction in planning time for secondary science teachers in England using AI support, and a 9-percentage-point rise in pass rates when less experienced tutors used an AI assistant (OECD Digital Education Outlook 2026, PDF).

The general trend matches gifted research: AI works best when it supports teachers and learning goals, and worst as a shortcut for students’ thinking.

What we could not verify

  • The High Ability Studies article (2026) reporting that AI-assisted learning with environmental scenarios improved cognitive and problem-solving skills of 41 gifted middle-school students appeared in search results, but the publisher page timed out. We do not report its details as confirmed.
  • The claim that “more than 4%” of top Codeforces ranks are AI cheaters comes from a community blog post, not from Codeforces itself.
  • Some studies (for example Mahmoudi-Dehaki & Nasr-Esfahani 2025) are reported only through the two Frontiers scoping reviews. We did not read the full papers.
  • The World Bank Nigeria and Harvard physics effect sizes come from abstracts and search summaries; we did not read the full papers.
  • For Roeper Review and the Journal for the Education of the Gifted, we did not find a 2024–2026 empirical article focused on AI in our searches. This does not mean none exists.
  • The ECHA 2026 themes page did not load for us; its AI theme is confirmed only via search summaries and attendee reports.

Key takeaways

  • Early days: a July 2026 scoping review found only 26 studies on AI in gifted education, mostly small or conceptual.
  • Identification: machine learning may help spot hidden or twice-exceptional talent, but can repeat old biases. It should be decision support, not decision replacement.
  • Teaching: the strongest current use is helping teachers prepare differentiated material faster, with expert review.
  • AI tutors: no clear gains yet in gifted-specific trials; large general studies show higher homework scores but lower exam results, including among high achievers.
  • Big question: researchers ask whether gifted education should shift from skills AI can copy toward problem-finding, creativity, ethics and self-regulation.

Sources

← Back to all posts