Interactive AI tools, personalized feedback, and EFL learners’ intention to use AI

Synthesis: Alshammari and Abdelaty (2026) extend the Technology Acceptance Model with two pedagogical affordances — interactive AI features and AI-generated personalized feedback — and test the extension on 596 Egyptian university students learning English as a foreign language, then interview eight of them. Perceived usefulness and perceived ease of use both predicted the intention to use AI, while the two new affordances reached intention only indirectly, and by different routes: interactivity through usefulness, feedback through ease of use. The model explained 72% of the variance in intention, with adequate fit indices and a measurement model that held, though two construct pairs sat exactly at the discriminant-validity threshold. The interviews supply the mechanism — learners valued dialogue they could repeat without embarrassment, and Feedback that named their errors instead of rewriting their work — and a condition the survey could not show: continued use was contingent on verifying output and on teacher judgment. The outcome throughout is behavioral intention measured by self-report, not use and not learning.

Key Findings

1. Perceived usefulness and perceived ease of use both predicted behavioral intention, with the ease-of-use path larger (0.710 against 0.346 in unstandardized estimates); the model explained 72% of the variance in intention.
2. Interactive AI features raised perceived usefulness (1.031) and reached intention only indirectly, through usefulness (indirect 0.357, 95% BC CI (.079,.643), p = .013).
3. Personalized learning feedback raised perceived ease of use (0.980) and reached intention only indirectly, through ease of use (indirect 0.696, 95% BC CI (.426,.996), p < .001).
4. Fit was adequate — χ²/df = 2.802, CFI = 0.964, TLI = 0.958, RMSEA = 0.055 — on a five-point instrument back-translated into Arabic and piloted with 30 students.
5. Eight interviews produced four themes: usefulness through interactive engagement, feedback as a route to manageability, AI as a psychologically safe environment, and learner agency with a continuing need for human guidance.
6. Interviewees described questioning the tool repeatedly without embarrassment, yet they conditioned continued use on checking output against course material, on teacher explanation and on their own judgment about accuracy.

Design and what the model was built to test

The study is an explanatory sequential mixed-methods design. The survey phase sampled 596 students by convenience at one public and one private Egyptian university, measuring personalized learning feedback, interactive AI tools, perceived usefulness, perceived ease of use and behavioral intention, with the feedback and interactivity items adapted from Yaseen et al. (2025) and the rest from Davis (1989). Only 16.3% had formal training; 84.6% were undergraduates. The interview phase then selected eight participants — seven undergraduates and one postgraduate trainee teacher — interviewed in Arabic over Zoom for roughly 30–50 minutes and analyzed thematically.

The extension is deliberately narrow, and the authors say so: learner agency and human–AI partnership frame the paper but are not measured, and AI literacy, proficiency, self-regulation and prior experience are excluded as controls. Generative AI acceptance here is a function of perceived affordances and little else.

Two affordances, two routes into intention

The direction of each path is the paper's main theoretical claim. Interactive features fed perceived usefulness; personalized feedback fed perceived ease of use; neither reached intention directly. Read together, the two affordances do different work: dialogue, clarification and follow-up practice made the tool seem worth using, whereas feedback that identified errors and simplified explanations made the task feel manageable. Ease of use here means reduced uncertainty about the work, not interface simplicity — which is why the authors insist the two features are not interchangeable design characteristics.

The paper reports unstandardized regression weights (IAT → PU = 1.031, PLF → PEU = 0.980, PU → BI = 0.346, PEU → BI = 0.710), so magnitudes sit on each instrument's own scale. Measurement was adequate but not comfortable: composite reliability ran 0.851–0.927 and AVE 0.588–0.720, yet HTMT sat exactly at the 0.90 cut-off for the feedback–interactivity and usefulness–ease-of-use pairs, which the authors defend as theoretically plausible rather than resolve empirically.

What the eight interviews add

The interviews supply the experience behind the coefficients. Ahmad, an English major, asked the tool to explain step by step "until everything becomes clear" — "it's like having a private teacher" — the usefulness path in a learner's own words. The same participant used feedback selectively: "I want it to show my mistakes, not change everything." That is a claim about control over revision, closer to the ease-of-use path than to anything the survey measured.

A second contribution is affective. Participants contrasted AI with a classroom where asking risks judgment; Hala valued repeating a question "without feeling embarrassed," and Noha described learning that "depends on you," more flexible than the classroom because depth and pace were hers to set. This is student–AI interaction as a low-stakes rehearsal space, and it explains why repeated questioning — not one good answer — was what learners called useful. The anxiety relieved is social, not technical.

The third contribution qualifies the model. Participants did not treat AI as a replacement for instruction: Eman preferred face-to-face study for "direct interaction," and Ahmad worried about inaccurate responses accepted uncritically. Intention to continue was conditional on verification and teacher guidance.

What the design can and cannot tell us

The interviews explain the survey's paths only in the weak sense of illustrating them. Eight people were drawn from a 596-person sample, the paper does not report whether their survey responses were linked to their accounts, and coding began deductively from the model — a design that could hardly contradict the constructs it was built to illuminate.

What they do establish independently is the conditionality. Acceptance coexists with critical awareness: the same learners who found the tool useful described checking it and keeping the teacher in the loop. That is what the instrument could not detect, since no item asked about verification or trust.

What this means for practice

• Instructors. Design AI tasks around follow-up dialogue and error-specific feedback rather than one-shot answers. Interactivity and feedback moved different beliefs, so optimizing for one does not substitute for the other.
• Instructors. Require students to compare AI output against course materials and to say which suggestions they rejected and why. The interviewees who used the tool well described exactly that selective practice.
• Educational technology developers. Prioritize adjustable explanations, requests for simpler or fuller responses, and guidance that names the error instead of rewriting the text.
• Administrators. Fund training in prompt formulation, verification, privacy and integrity, and publish rules separating acceptable assistance from misconduct; formal training reached a small minority here while most learned alone.
• Researchers. Pair intention instruments with usage logs or performance measures before claiming that acceptance predicts learning.

Limitations

• The dependent variable is behavioral intention, not use and not learning: 596 questionnaires at one time point, with no usage records, observation or language-performance data behind the 72% of explained variance.
• The setting is one national higher-education system: two Egyptian universities (one public, one private), convenience sampling, and university affiliation not recorded as a questionnaire item, so the public–private split could not be examined.
• The qualitative phase rests on eight interviews — seven undergraduates and one postgraduate trainee teacher — with no reported link between their survey responses and their accounts, so the themes illustrate the model rather than validate it.
• Feedback and interactivity were not empirically clean constructs: HTMT was 0.90 for the PLF–IAT and PU–PEU pairs, exactly at the accepted threshold, which weakens the claim that the two affordances route through distinct beliefs.
• Prior AI experience was neither controlled nor analyzed: 81.0% had used ChatGPT for English learning and 48.0% were self-taught, so the model describes an already-adopting population without testing whether experience changes the paths.

Citation

Alshammari, S. H., & Abdelaty, Y. (2026). Interactive AI tools, personalized feedback, and EFL learners’ intention to use AI: an extended TAM study in Egyptian higher education. Frontiers in Education, 11, 1923377.