AI Is Doing Children’s Homework: Learning Aid or Shortcut That Weakens Thinking?

AI Is Doing Children’s Homework: Learning Aid or Shortcut That Weakens Thinking?

Rating
—/5 0 ratings
Follow us on Google News
Education & Careers22 min read

Artificial intelligence is doing children’s homework in seconds: solving equations, writing essays, summarising novels and explaining lessons. But there is a difference technology cannot erase between a correct answer and a child who has understood the lesson. A well-written paper alone does not show that a pupil could defend the ideas in it. On the other hand, an adapted explanation, a well-chosen question or a hint offered at the right moment can help a pupil overcome a real stumbling block.

This is no longer a debate about the future. In the OECD’s PISA 2025 results note for Romania, published on 8 September 2026, 40% of Romanian 15-year-olds said they use AI chatbots for learning support at least weekly. One-third use them to research or write texts for homework. These are pupils’ survey responses, not a direct measure of how well they have learned.

The useful question for parents and teachers is not simply, “Did the child use AI?”, but “Which part of the intellectual work remained with the child?” Research published in 2025 and 2026 points to both benefits and risks, depending on the tool’s design, the activity and the extent of the pupil’s involvement. This InfoBuz guide sets out the evidence, ten homework activities and a practical rule for putting understanding before the finished product.

For more on digital education, see also Children’s online safety: a guide for parents, from TikTok to online games, which provides complementary information on the subject.

AI is doing children’s homework: how widespread is it already in Romania?

PISA 2025 asked 15-year-olds in Romania how they use AI chatbots in schoolwork. The publication of the results in September 2026 provides a more useful national reference point than assumptions that “all pupils” use them or that “nobody learns any more”.

Use reported by Romanian pupils Percentage What it does and does not mean
Using AI for learning at least weekly 40% Self-reported frequency; not an indicator of benefit or harm
Initial research on a new topic 33% Describes a purpose of use; does not show the quality of sources consulted
Writing texts for homework or school assignments 33% Does not in itself distinguish between brainstorming and submitting an entirely generated text
Summarising a text the pupil was meant to read 31% May support reading or replace it, depending on the process
Never or almost never using chatbots for the tasks surveyed 9% Does not mean everyone else uses AI every day or for every homework assignment

Source: OECD, PISA 2025 Results (Volume I), Romania. The percentages concern different questions and purposes, which may overlap; they should not be added together.

These figures describe the reported behaviour of pupils who took part in PISA, not all children in primary, lower-secondary and upper-secondary education. Nor do they show whether every pupil had a teacher’s permission or whether the work produced was better. For context on school performance, see InfoBuz’s analysis of Romania’s PISA 2025 results.

Another sign comes from the Ministry of Education and Research. More than 17,600 pupils and over 930 teachers from 295 schools took part in the “Explorers in the World of Artificial Intelligence” programme, presented on 27 July 2026. Activities included algorithmic bias, ethics and responsible use. Participant numbers show interest and educational activity, but not an experiment demonstrating that AI automatically improved school performance.

Correct homework is not necessarily a lesson learned

A pupil can receive an impeccable answer from a chatbot and still not know why it is correct. The distinction is between performance on the immediate task and acquired capability: can the child apply the method to another exercise, explain the reasoning or identify an error?

In its OECD Digital Education Outlook 2026, the organisation stresses that generative AI can improve the quality of a final product without necessarily producing learning gains. Tools explicitly designed for education, with questions, hints and feedback, have different potential from general chatbots used only for instant solutions.

A simple comparison illustrates the point. If a child asks, “Write my essay about autumn,” and copies the response, they obtain a text. If they ask, “Ask me three questions that will help me find my own ideas,” then write their own essay, they practise imagination, organisation and expression. AI was used in both cases, but the intellectual work left to the pupil was different.

Not every quick use is a problem. Checking a definition after understanding a lesson can save useful time. The concern arises when a pupil systematically delegates the very steps that should build a new skill.

What research from 2025–2026 shows: AI can help or reduce learning

The scientific evidence is mixed. It supports neither the slogan that “AI makes children smarter” nor the verdict that “AI is destroying a generation’s thinking”. Studies vary by participants’ age, subject, model used, task, quality of assistance and the assessment carried out without AI.

The maths experiment: better exercise results, but a risk in the test without AI

A 2025 study in Proceedings of the National Academy of Sciences, by Hamsa Bastani and colleagues, tested access to AI assistants among nearly 1,000 secondary-school pupils studying mathematics. During exercises, access to AI improved marks on practice tasks. The authors reported practice-grade improvements of 48% for a system resembling a general chatbot and 127% for a pedagogically configured tutor, compared with the control group.

The important difference emerged after the assistance was withdrawn: in the subsequent assessment without AI, pupils who had used the chatbot without pedagogical safeguards scored 17% lower than pupils without AI access. The negative effects were largely mitigated when the tutor was designed to give hints and not reveal the solution immediately. Source: PNAS, 2025.

These percentages are results from one specific experiment, not the rates by which the marks of every Romanian child who opens a chatbot would fall. But the study offers a practical lesson: how AI responds can matter as much as access to AI itself.

A July 2026 study also found learning gains

In a randomised experiment involving university students, published as an IZA Discussion Paper in July 2026, Zara Contractor and Germán Reyes found that AI access increased immediate knowledge-check scores by 0.27 standard deviations, with benefits persisting one week later. Writing results varied: participants who used the tool to explain ideas, rather than automatically produce text, had better delayed gains on certain criteria of independent writing.

A standard deviation is a statistical unit; it does not mean a 27% increase in a grade. The participants were university students, not primary-school pupils, so the findings cannot be directly transferred to children of every age.

Other experiments show that access alone does not guarantee engagement

An IZA study from October 2026, involving 2,379 students and 30 teachers, reported lower final grades and lower participation in the learning platform in the university sections given access to an integrated AI tutor. The paper is a research working paper, and generalising its findings to all education systems would be unwarranted.

In another direction, a two-year experiment in 18 middle schools in Tennessee, presented in August 2026, observed small increases in maths performance for pupils offered Khanmigo. The gains were comparable with those from Khan Academy activities without that assistant, and actual conversational use of the tutor was limited.

Taken together, the research does not show that AI has a single effect on learning. It suggests that outcomes depend on what AI replaces, how the tutor is designed, how active the pupil remains and how competence is assessed without support.

Ten homework examples: when AI can assist teaching and when it becomes a shortcut

The following ten scenarios are original editorial examples, not the results of an experiment InfoBuz conducted with children. Teachers and parents can adapt them to age, curriculum and school rules. For children below a service’s permitted age, interaction with it should be managed by an adult, not left to the child.

1. Mathematics: ask for a hint, not the full solution

Homework: the pupil must solve a first-degree equation and gets stuck after the first step.

Useful use: “Do not solve the exercise for me. Ask me what operation I could do first, and explain only where I make a mistake.” The pupil suggests the steps, checks the transformations and writes the result independently.

Avoid: “Solve all exercises 1–20 and write only the answers.” If the system performs the operations the child is meant to learn, the homework may be finished without the skill being developed.

Check: after the explanation, the pupil solves a similar exercise with different numbers unaided. A correct answer to a different exercise is a better sign of understanding than flawlessly copied homework.

2. Romanian language: an essay structure, not a finished essay

Homework: a 150–200-word description of an autumn day, at an age-appropriate level.

Useful use: “Ask me five questions about what I saw, heard and felt. Do not write the essay for me.” The child can organise ideas into an opening, body and conclusion, then rewrite sentences they dislike independently.

Avoid: submitting a fully generated text containing wording the pupil cannot explain. Authorship is not the only issue: the pupil loses an opportunity to practise putting ideas into words.

Check: the child shows an outline, a first version and two changes they can justify.

3. Reading: questions about the book before a summary

Homework: reading a literary passage and identifying a character’s main motivation.

Useful use: after the text has actually been read, an adult can ask AI to suggest comprehension questions, including one whose answer must be supported by a passage from the book. The pupil finds the page and frames the explanation.

Avoid: systematically replacing the assigned text with an automated summary. A summary may omit essential details, introduce unsupported interpretations and does not practise reading.

Check: the pupil accurately quotes the real passage and explains why it supports the interpretation.

4. History: verify the source and chronology

Homework: comparing two causes of a historical event.

Useful use: “Suggest questions I can use to assess a historical document: who wrote it, when, for whom and for what purpose?” The pupil then uses the textbook or a verifiable source to answer them.

Avoid: accepting AI-invented quotations and bibliographic references. A system can sometimes produce names, pages or sources that look real but do not exist.

Check: every quotation is found in the original source, not just in the chatbot’s explanation.

5. Science: form the hypothesis before the explanation

Homework: observing how light affects a plant’s development, in an age-appropriate activity.

Useful use: “Help me formulate a testable hypothesis and identify which variable needs to stay constant.” The child makes observations, records results and discusses the experiment’s limits.

Avoid: generating a table of fictional observations and presenting it as real measurement. AI can help explain data, but it should not invent the results of an experiment that was not conducted.

Safety: for experiments involving substances, high temperatures or dangerous instruments, the teacher’s rules and adult supervision take priority over any automated recommendation.

6. Biology: use AI to identify misconceptions

Homework: explaining photosynthesis and how it differs from cellular respiration.

Useful use: “Give me three statements about photosynthesis, one of them wrong. Let me identify it and explain why.” The pupil then compares the answers with the textbook.

Avoid: memorising an elegantly worded answer that mixes up biological mechanisms or contains errors. Diagnostic questions and explaining a misconception are more valuable than repeating an uncomprehended definition.

Check: the child can explain the mechanism in their own words and correct a false statement.

7. Geography: the map before the generated conclusion

Homework: comparing two regions by climate, relief and economic activity.

Useful use: AI can suggest a table with headings or comparison questions. The pupil finds the data in an atlas, textbook and official sources, and records what has been checked.

Avoid: putting rainfall or population figures into a project without a year or source. A precise figure may be outdated, incompatible with the region or simply fabricated.

Check: the pupil locates the regions on a map and can explain how relief or geographical position contributes to the differences.

8. Foreign languages: a conversation partner, not a translator that does everything

Homework: an English dialogue about organising a trip.

Useful use: “Have a simple conversation with me at my level. Ask one question at a time and briefly correct important mistakes after I answer.” The pupil produces sentences and observes their progress.

Avoid: submitting a perfect translation of a text the child could not say aloud. An improvised dialogue with new questions can be a better test of competence.

Check: the pupil uses five new expressions in their own conversation without the model open.

9. Computing: explain the error, rather than provide code to copy

Homework: a simple program does not produce the expected result.

Useful use: “Show me where the problem might be and ask me to check the hypothesis before offering corrected code.” The pupil understands how variables, conditions or loops work.

Avoid: replacing the entire file with a generated solution the pupil cannot alter or explain. In assessed projects, the teacher’s rules on outside help also matter.

Check: the pupil can reproduce the correction and explain why the earlier test failed.

10. Projects and debates: an argument opponent, not the author of the conclusion

Homework: a short speech about phone use at school.

Useful use: “I support this argument. Give me three serious objections so I can respond better.” The pupil must bring sources, identify weaknesses and clarify their position.

Avoid: requesting a complete, emotive speech full of unverified “quotations”, then memorising it. The educational purpose is to practise thinking and expression, not repeat an opinion the pupil did not formulate.

Check: the pupil independently defends the argument, acknowledges a legitimate objection and can revise a conclusion in light of new evidence.

The AI homework traffic light: useful uses and uses that need discussion

Situation Recommendation Reason
AI asks questions, explains an error or suggests hints Generally appropriate with clear rules The pupil continues to think and formulate the answer
AI offers structural ideas, practice tests or feedback Use with checking and transparency The subject, teacher’s requirement and who wrote the final version matter
AI completes the homework, invents sources or generates experimental data Do not use as one’s own work It conceals the learning process and may breach assessment rules
A pupil enters names, photographs, grades, medical records or classmates’ data Avoid it and check the service’s policy Privacy and children’s data protection

This table is an editorial recommendation, not a national rule overriding school policy. A teacher may ban, conditionally allow or require a particular use of technology in an assignment.

The 20-minute rule: a simple at-home protocol

Parents can turn an AI-assisted session into an exercise in independence. The protocol below is a practical InfoBuz proposal, not a method experimentally validated for every age group and subject. Timings are indicative and can be adapted.

  1. First five minutes – independent attempt: the child reads the task, notes what they know and attempts a first step. The adult does not immediately open the chatbot.
  2. Next five minutes – one question for AI: ask for a short explanation or hint, not the full answer.
  3. Next five minutes – independent work: set the screen aside and have the pupil apply the explanation in their notebook.
  4. Final five minutes – checking and reflection: the child explains what they got wrong, what they understood and what they would do differently in a similar exercise.

For a particularly difficult task or a pupil with particular educational needs, the timing and type of support should be adapted. A parent need not turn every homework assignment into an examination, and a child should not be humiliated for asking for help. The aim is to reduce dependence on an immediate answer.

Five AI prompts that can preserve the pupil’s thinking

For teachers and adults using services permitted by age rules and school policy, the wording of a request changes how the tool helps. These examples can be adapted without entering children’s personal data:

  1. “Do not give me the solution. Ask me a question that helps me choose the first step.”
  2. “Explain this idea with one simple example, then let me create a second example myself.”
  3. “Check the reasoning I wrote and point out the first error, without rewriting all my homework.”
  4. “Give me three questions of increasing difficulty to test my understanding.”
  5. “Compare two interpretations and tell me which sources I should check myself.”

A chatbot may not follow a request perfectly or may provide incorrect explanations. The teacher, textbook and credible sources remain the reference points for accuracy, and children should be encouraged to say when an explanation makes no sense.

How teachers can assess learning without turning every pupil into a suspect

Banning every AI tool may be difficult to enforce, while accepting any generated text as proof of competence can undermine assessment. An alternative is to make the process more visible, without claiming that every teacher can determine precise authorship simply by reading.

An assignment can require an initial plan, a paragraph written in class, two verified sources and a short explanation of decisions. Teachers can assess ideas and justifications, not only the polish of the final prose. In relevant subjects, a new exercise completed individually can complement a project prepared at home.

An indicative marking rubric might allocate: 30% for conceptual accuracy, 25% for explaining the reasoning, 20% for source verification, 15% for organisation and clarity, and 10% for reflection on errors. This marking example is proposed by InfoBuz; it is not an official rubric. Each teacher should adapt it to the objectives, age group and curriculum.

In an earlier article, InfoBuz separately explained which skills a pupil should be able to demonstrate without AI. This guide complements that analysis with experimental studies, subject-specific activities and recommendations for families.

AI text detectors are not sufficient evidence to accuse a pupil

Some tools attempt to estimate whether a text was generated by a model. Their results can be wrong: human-written text may be labelled synthetic, while generated text may go undetected after editing. A numerical “AI probability” score is not a reliable test of authorship and does not prove whether a pupil has learned.

A fairer educational procedure begins by clarifying the rules of use, gathering evidence about the process and allowing the pupil to explain their work. A teacher can discuss how the child completed an assignment without turning stylistic observations or a detector into a final accusation.

It is not justified automatically to penalise a pupil because they use more mature vocabulary, write more accurately than usual or receive a high score in a detection application. Any disciplinary procedure must respect school rules and the pupil’s rights.

Are there particular risks for younger children? Age, data and supervised use

AI tools have different terms and age limits. OpenAI states that ChatGPT is not intended for children under 13, and that users aged 13 to 18 need a parent’s or guardian’s consent. For educational activities involving children under 13, interaction with ChatGPT must be carried out by an adult. These are the rules of one particular service, not a universal legal age threshold for all AI applications.

UNESCO recommends a human-centred approach, protection of privacy and age-based rules for independent use of generative platforms. UNICEF, in guidance updated in 2025, stresses the need for AI systems that respect children’s rights, safety and data protection.

In practice, parents and teachers should avoid uploading minors’ full names, addresses, photographs, individual grades, medical records, school documents or private conversations to tools where there is no justification and no adequate safeguards. An exercise example can use invented data without identifying a pupil.

A parent can use AI themselves to prepare educational questions or adapted worksheets without allowing a child to create an account that is incompatible with the service’s terms. For logic exercises independent of chatbots, InfoBuz also offers 40 challenges for Years 1–4.

What the European Commission recommends for schools integrating AI

In June 2026, the European Commission published an updated version of its ethical guidelines on the use of artificial intelligence and data in education, aimed particularly at primary and secondary-school teachers. The material includes scenarios and assessment questions, transparency, data protection and the implications of the AI Act.

The guidelines are not an instruction to use chatbots in every lesson, nor an argument for abandoning screen-free exercises. They offer a reference point for pedagogical choices: the tool should serve the learning objective, and the teacher should be able to explain why it is being used.

In Romania, the Ministry of Education announced a pilot project on AI use in education in March 2025. Local experiments, pilots and training programmes need to be assessed through real outcomes, not only through the number of accounts created or sessions held.

AI homework and educational equity: who receives support and who is left behind?

Not every family has the same internet connection, comparable devices, time for supervision or digital skills. One child may have access to a specialised tutor, while another has only a shared phone. Even with free tools, language barriers, accessibility and the quality of explanations may differ.

Conversely, a well-configured assistant could, in some circumstances, offer repeated explanations without the cost of one-to-one private tutoring. That opportunity does not mean a chatbot can replace a teacher, psycho-pedagogical support or an individual plan for pupils with special educational needs.

An equitable education policy must provide alternatives for children without technology access, teacher training, clear rules on data and independent evaluations of benefits. Otherwise, AI use may deepen existing inequalities, even where the original intention is modernisation.

What parents can do without becoming the “homework police”

Parents do not need to read every AI conversation or suspect every well-phrased sentence. A calmer discussion about purpose is more useful: “What did it help you with?”, “Which part did you solve yourself?”, “What did you check?”, “What do you understand better?”

A simple family agreement can be established: AI is allowed for explanations and extra practice, not for submitting a complete piece of work as one’s own; important information is checked; personal data are not entered into conversations; and the teacher’s rules take priority. For younger children, use should be supervised and adapted to the service’s age requirements.

A warning sign is not simply opening a chatbot, but a repeated situation in which the child no longer tries to solve anything, cannot explain what they submitted or becomes anxious without access to the tool. In such cases, it is worth speaking with the teacher and reconsidering the working method, rather than immediately attaching a label to the child’s intelligence or motivation.

Checklist for parents and teachers: before and after AI-assisted homework

  • Is the tool permitted for this age group and activity?
  • Do we know exactly which skill the homework is intended to develop?
  • Has the pupil first attempted an independent step?
  • Is AI offering hints and explanations, or producing the entire work?
  • Can the child check the information in the textbook or credible sources?
  • Have the personal data of pupils and classmates been avoided?
  • Has the school explained which uses are permitted and how they should be declared?
  • Is there a record of the process: an outline, versions, corrections or explanations?
  • Can the pupil apply the idea to a new task without help?
  • Have the activity and timing been adapted to the child’s age and needs?
  • Does technology use leave room for reading, handwriting, discussion and practice?
  • If AI makes a mistake, does the child know how to question and correct it?

Frequently asked questions about AI and children’s homework

Is it good for a child to use AI for homework?

It can be useful when the system provides hints, feedback or explanations that the pupil actively uses, in line with school rules and age limits. It becomes problematic when it automatically completes a task the pupil is meant to learn to do.

Does artificial intelligence make children think less?

There is no universal conclusion. Research shows different results according to the type of assistance and the pupil’s involvement. A chatbot that provides the solution can reduce the cognitive effort required, while questions and hints can support learning.

How many Romanian pupils use AI for learning?

According to PISA 2025, 40% of participating Romanian 15-year-olds said they use AI chatbots for learning at least weekly. The data do not represent all children and do not directly measure performance attributable to AI.

Can AI explain maths without giving the answer?

Yes. It can be asked for a hint, an intermediate question or feedback on a step taken by the pupil. AI responses must still be checked, and the tool does not always follow instructions perfectly.

Is it right for a pupil to submit an essay written by ChatGPT?

If an assignment requires independent writing and full use is not allowed, presenting generated text as personal work may breach assessment rules. Brainstorming and feedback may be treated differently, depending on the teacher’s instructions.

Can teachers reliably detect whether a text was written with AI?

No automated detector can guarantee authorship. Assessment should rest on transparent rules, the work process and demonstration of competence, not only on a percentage supplied by software.

From what age can a child use ChatGPT?

Under OpenAI’s rules, the service is not intended for children under 13, while those aged 13 to 18 require a parent’s or guardian’s consent. In educational activities involving children under 13, interaction with ChatGPT must be led by an adult. Other services may have different terms.

Should AI be banned in schools?

A general decision either to ban AI or to allow unlimited use does not by itself solve the problem. OECD and European Commission guidance recommends use adapted to pedagogical objectives, while protecting independent thinking and children’s data.

What can a parent do when a child asks only for the final answer?

They can begin with a brief attempt without AI, then ask for hints and return to a similar problem solved independently. Adjusting the level of difficulty matters; the entire activity should not become a punishment.

Is it enough for a pupil to declare that they used AI?

Transparency is important, but it does not replace checking understanding. It should be clear what the system did, what the pupil did and whether the assignment rules were followed.

Conclusion: the question is not who wrote the homework, but who learned from it

Artificial intelligence is doing children’s homework faster than teachers or parents can answer every question. But speed is not the same as learning, and an impeccable answer does not show that a pupil can think independently.

Studies published in 2025 and 2026 point to both benefits and risks: pedagogical support through questions and feedback can help, while automatic problem-solving can replace the effort that builds skills. PISA figures show that the technology is already part of schoolwork for many Romanian teenagers; the issue is how it is used, not whether we can pretend it does not exist.

A practical rule is for the pupil to begin with their own reasoning, receive the minimum help needed, check the information and be able to explain the result independently. AI can be a learning tool; it should not become the pupil who does the child’s homework.

Illustrative featured image: photograph by Katerina Holmes / Pexels, used under the platform’s licence. The image does not depict participants in the cited studies.

Sources and methodology

Etichete: AI in schoolsAI LiteracyChatGPT HomeworkChildren and Technologydigital educationAI Ethicsschool assessmentcritical thinking
Share this article
COMUNITATEA INFOBUZ

Discuția cititorilor

Ai o opinie sau o întrebare despre acest articol? Spune-ne ce crezi.

0 comentarii

Fii primul care deschide discuția.

Lasă un comentariu

Email address este vizibilă doar administratorilor și nu va fi publicată. Câmpurile marcate cu * sunt obligatorii.

Te rugăm să păstrezi un limbaj civilizat și să discuți la subiect. Comentariile spam sunt filtrate automat.