The cons of AI in education most often named are weaker thinking skills, answers students do not check, less connection to teachers, cheating and the burden of policing it, data breaches, biased or unfair treatment, and cost. They are not equal. A 2025 national survey by the Center for Democracy and Technology of 806 teachers, 1,018 parents and 1,030 students measured the first five: 71 percent of teachers worry AI weakens academic skills, half of students feel less connected to their teacher, 71 percent of teachers report the added work of checking whether work is a student's own, and 23 percent of schools had a large-scale data breach, rising to 28 percent where AI use was heaviest. Bias is a strong worry with no outcome measure yet. Cost and displacement are arguments, and the most-quoted dependence statistic was retracted in February 2026. The measured cons share one root: the answer arrives before the student has formed a view.
On this page+
The cons of AI in education most often named are weaker thinking skills, answers students never check, less connection to teachers, cheating and the work of policing it, data breaches, biased or unfair treatment, and cost. They are not equal. Some were measured in 2025 in national samples of teachers, parents and students. Some are still only worries. One was retracted.
You are probably reading this because a decision is close: a school is rolling out a tool, a child has started using one for homework, or you have noticed your own hand reaching for the chat window before your own head. Every list you have found gives the same cons at the same volume. This piece sorts the cons of AI in education by what sits under them, so you can spend your worry where the evidence is.
What are the cons of AI in education?
The cons of AI in education, in the order they appear across the top results, are these:
Weaker thinking skills. Students hand the reasoning to the tool and stop practising it.
Unchecked answers. Generative tools state wrong things fluently, and students accept them.
Less connection to teachers. The tool answers faster than a person, so the person is asked less.
Cheating, and the burden of policing it. Work arrives that may not be the student's own.
Student data privacy. More systems hold more student data, and breaches follow.
Bias and unfair treatment. Detectors and graders misjudge some groups more than others.
Cost and displacement. Licences, training, and the fear that teachers become optional.
Written that way, the list looks like seven facts. It is closer to three kinds of claim. The first five were measured in 2025 by the Center for Democracy and Technology in a national survey of 806 teachers, 1,018 parents and 1,030 students.1 The sixth is a concern with a worry measure and no outcome measure yet. The seventh is an argument, and the most-quoted number behind it was retracted in February 2026.2
Which cons have actually been measured?
Five of the seven cons of AI in education have numbers behind them from the same 2025 dataset, and the numbers move together. The CDT survey, fielded from June to August 2025, found that 86 percent of students had used AI in the 2024-25 school year. Of teachers, 71 percent worried that AI weakens the academic skills students need, including writing, reading comprehension, critical thinking and research; the same share worried students would not question whether the tool's answers were accurate. Half of students said using AI in class made them feel less connected to their teacher. Seventy-one percent of teachers said student AI use had added the burden of working out whether a piece of work was the student's own. And 23 percent of teachers said their school had suffered a large-scale data breach that year, 28 percent where AI was used for many purposes and 18 percent where it was barely used.1
Two things about that block matter more than any single figure. First, every risk rose with use: the report's own framing is that the emerging risks "all increase the more that a school uses AI".1 Second, who worries runs opposite for adults and children. Teachers and parents who used AI more were less concerned. Students who used it more were more concerned. The people closest to the effect were the ones reporting it.
Does AI weaken students' thinking?
Students and teachers both believe it does, and the best study of the mechanism says the belief is about right, with one twist. In the CDT sample, 64 percent of students agreed that using AI weakens important skills they need to learn; among students who had held back-and-forth conversations with a tool, 66 percent.1 These are self-reports, not test scores, and the survey does not claim otherwise.
The twist comes from a different population. Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real uses of generative AI and published the results at CHI in 2025. Their central finding: "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking."3 The tool was present in both cases. What changed the amount of thinking was where the person placed their trust. That is the whole con in one sentence, and also the way out of it.
We made the longer case for that mechanism in why AI should not be used in education. The short version is that AI and critical thinking are not enemies by nature. They become enemies the moment the answer arrives before the student has formed a guess. In sessions we watch the same thing in adults: the person who writes their own view down first argues with the tool; the person who asks first agrees with it.
That exercise is the shape of a session with Inwitt AI, an AI thinking partner app that never hands over the answer. When you ask it for facts, the research reveal makes you guess before the lookup, so your own view exists before the tool's does. Confidence stays with you.
Why do students feel less connected to their teachers?
Because the tool answers faster than a person can, and asking a person is the thing that built the relationship. In the CDT data, 50 percent of all students agreed that using AI in class made them feel less connected to their teacher. In schools that used AI for many purposes the figure was 56 percent; in schools that used it for few or none, 46 percent. Thirty-eight percent of students said that when they did not understand something they would rather work with AI than with a teacher.1
The tool answers faster than a person can, and half of students say the relationship was the price.
None of the five competitor pages we read carried this number. "Loss of human connection" usually gets written as a soft, unmeasurable worry, conceded in one sentence before the pros resume. It turns out to be one of the most measurable cons on the list. A similar share of students said a teacher who uses AI in class "is really not doing their job as a teacher".1 That is a comment on what a student thinks a teacher is for.
The mechanism is the same one from the thinking section. A question asked of a person costs something: a pause, a bit of exposure, a wait. A question asked of a tool costs nothing, so it gets asked first, and the person is asked last or never. The relationship is not destroyed by AI in schools. It is starved by the order in which questions get asked.
What do the cheating and privacy numbers say?
They say the cost of AI cheating lands mostly on teachers, and the cost of student data privacy failures rises with every extra use. Seventy-one percent of teachers agreed that student AI use had created an additional burden on them to work out whether a student's work was their own. Thirty-nine percent said AI tools were more of a distraction from learning than a way to improve it; among teachers who barely used AI themselves, 50 percent.1 The academic integrity con is less about the number of cheats and more about the hours of verification added to a job that had no spare hours.
The privacy con has the cleanest dose-response curve in the report. Just under a quarter of teachers said their school had experienced a large-scale data breach, ransomware attack, or accidental sharing of student data in 2024-25. The share was 28 percent where AI was used for many school purposes and 18 percent where it was used for few or none.1 The report does not claim AI caused the breaches. It shows that the schools reaching hardest for the tools are the schools whose data is leaking most, which makes privacy a con with evidence rather than a con with a paragraph.
One more number belongs beside these: only about one in ten teachers had received training on how to respond when a student's AI use turns harmful.1 The tools arrived; the training did not. Most of the practical risks of AI in schools sit in that gap.
Which cons are argued but not yet shown?
Two cons on every list have no outcome measure behind them yet, and one carries a number that should no longer be used. The whole list, with its evidence tier:
Con
Evidence tier
What the evidence is
Who pays first
Weaker thinking skills
Measured (self-report)
71% of teachers, 64% of students agree; CHI 2025 confidence finding
Students
Unchecked answers
Measured (self-report)
71% of teachers, 60% of parents worried
Students
Less connection to teachers
Measured
50% of students agree; 56% in high-use schools
Students, then teachers
Cheating and policing it
Measured
71% of teachers report added burden
Teachers
Data breaches
Measured (incidence)
23% of schools; 28% high-use vs 18% low-use
Families
Bias and unfair treatment
Reported concern only
31% of students worry; 39% with an IEP or 504 plan
Students already at a disadvantage
Cost and teacher displacement
Argued
No outcome data in the sources we read
Unknown
Bias is the con we would most like to see measured, because the worry is not evenly spread. In the CDT sample, 31 percent of students worried an AI tool would treat them unfairly; among students with an individualised education programme or a 504 plan, 39 percent.1 A worry measure is not a finding of unfair outcomes. The pages that state detectors flag non-native English writers more often may well be right; none gave a source we could open, so we are not repeating it as fact.
Cost and teacher displacement are arguments, and reasonable ones, but arguments. And the number most often quoted under the "dependence" con, that AI accounts for 68.9 percent of human laziness and 27.7 percent of lost decision-making, comes from a 2023 paper of 285 university students in Pakistan and China that its journal retracted on 3 February 2026 over insufficient ethical approval; the authors disagree with the retraction.2 Whatever the merits of the dispute, the figure is now unsupported. If you find it in a list of the disadvantages of AI in education, treat the rest of that list with care.
Common mistakes when weighing the cons of AI in education
The first mistake is treating the cons of AI in education as seven equal items. A parent who spends the same worry on teacher displacement as on their child never checking an answer has spent most of it on the con with the least evidence. Rank first, then worry.
The second is quoting the retracted number. It travels because it is precise, and precision feels like proof. A number with a retraction notice attached is worse than no number, because it tells an informed reader that the writer did not check.
The third is banning the tool instead of moving the confidence. Lee's finding was not that the tool reduced thinking; it was that trusting the tool more than yourself did.3 A ban removes the tool and leaves the trust where it was, ready for the next tool. The habit that survives a ban is the one where the student writes a view before asking. We described what actually improves critical thinking elsewhere, and it is not abstinence.
The fourth is reading adult concern as student experience. The CDT report shows the two moving in opposite directions: adults who use AI more relax about it, students who use it more do not.1 If you want to know what over-reliance on AI feels like from the inside, ask the students, using the question in the exercise above.
Questions people ask about AI in education
Is AI in education good or bad?
Both, and the split is between uses, not tools. The same CDT teachers reported more personalised learning and more time with students alongside more verification work and more breaches.1 The negative effects of AI in education cluster around one use, letting the tool think first, and that use can change without removing the tool.
Can AI replace teachers?
Not on the evidence, and students are the ones saying so. Half of students in the 2025 sample said a teacher who uses AI in class is not really doing their job.1 What students describe wanting is a person who asks them what they think. No tool currently meets that job description.
What are the negative effects of AI on students specifically?
Weaker skills by their own account, less trust in their teachers, and more worry, in that order. Sixty-four percent of students said AI use weakens skills they need, 50 percent felt less connected to a teacher, and students who used AI more reported more concern, not less.1
How can students use AI without becoming dependent on it?
By putting their own answer on the page before the tool's. Write the guess, then look. Check one claim in every answer rather than none. That is reasoning critically in its smallest form, and it is the difference Lee's study found between people whose thinking went up with the tool and people whose thinking went down.3 A tool that draws the student's reasoning onto a thinking board before it says anything makes the order hard to skip.
The cons of AI in education that survive the evidence share one root: the answer arrives before the question has done its work. Every session at Inwitt runs the other way, one question at a time, your reasoning on the board, a receipt at the end in your own words. If there is a decision about AI in front of you, at school or at home, join the Inwitt waitlist and bring it.
Retraction Note (3 February 2026) to Ahmad, S. F. et al. (2023), "Impact of artificial intelligence on human loss in decision making, laziness and safety in education," Humanities and Social Sciences Communications. The journal cites insufficient ethical approval for the research involving human participants; the authors disagree with the retraction. https://www.nature.com/articles/s41599-026-06602-8↩↩2
