Dymocks Tutoring and Talent 100 closed five Sydney centers on Sept. 27 after telling parents to spend their money on ChatGPT or Gemini instead, even as it arranged continued human tutoring for students approaching their final exams.
The advice, first reported by Maani Truu in The Australian Financial Review, came from a business that had bought a competitor just two years earlier. Its chief executive, Mark Buckland, was making a judgment about price as much as teaching: human tutors were better, he said, but an A.I. subscription cost a fraction of their fees.
Research supports part of his argument. In controlled experiments, some A.I. tutors have helped students learn as much as human instructors at much lower operating cost. Other experiments have found that students who did better with a chatbot performed worse once it was taken away. The distinction is between getting a problem right with assistance and being able to solve the next one alone.
The Sydney exit
About 38 students preparing for their school-leaving examinations were to keep individual online sessions with their assigned tutors until the exams, according to GIGAZINE’s account. The arrangements made the shutdown less absolute than the recommendation to use A.I. suggested.
Uplift Education was to take over the Parramatta classes and online classes from the following teaching period. Students at the other four centers were offered discounted tuition through Art of Smart. Those were routes to another human provider, rather than instructions to replace every lesson with a chatbot.
Mr. Buckland’s argument concerned families paying for an academic advantage. Dymocks charged as much as A$900 per subject. A subscription could provide explanations, practice and feedback across subjects for much less. In remarks to the Financial Review reproduced by Streamline Feed, he acknowledged the difference in quality.
Sure AI won’t perform as well as our human tutors with all the support behind them, but it’s also $30 a month.
The A$30 figure was his comparison at the time, not a fixed price for every product or family. Nor was it evidence that a subscription reproduced everything included in tutoring fees. It priced access to software against a service involving staff, premises and scheduled instruction.
Mohan Dhall, chief executive of the Australian Tutoring Association, said the withdrawal surprised him. David Ramagge, chief executive of Matrix Education, offered a different business judgment: about 8,000 students attended its Sydney-area classes each teaching period, and enrollment was growing, according to the account of his Financial Review interview. Matrix also offered A.I. tools. Mr. Ramagge expected high achievers to continue seeking in-person instruction.
Those statements came from people selling or representing tutoring. They do not settle which service teaches better. They do establish that Dymocks’s conclusion wasn’t shared across its industry. It was not clear how much of the decision reflected students switching to A.I., rather than the economics of this particular operation.
The earlier expansion
In March 2019, Mr. Buckland was describing enrollment growth of 600 percent and plans for 30 corporate hubs across Australia, with franchised centers to follow, Inside Franchise Business reported. The business then used the name Potentia. Dymocks, the bookseller, had announced its first hub the previous September.
The plan depended on a trusted brand making tutoring more consistent. Corporate hubs would establish a presence in each state. Franchisees would extend it, including into regional areas. Mr. Buckland also discussed educational technology and publishing the company’s own learning materials. Software was already part of the proposed business, though not a reason to abandon it.
Dymocks Education’s FY24 annual update recorded the acquisition of Talent 100 in July 2024 and described it as the beginning of a consolidation effort. The sequence matters: Dymocks was still buying capacity less than two years before deciding that cheaper alternatives made its tutoring difficult to justify.
Australia’s tutoring market had grown without a dependable national account of who used it or what families received. In a September 2025 article, education researchers Ben Zunica, Bronwyn Reid O’Connor and Katherin Cartwright cited estimates that one in six Australian pupils received tutoring during their schooling, rising to one in four in places including Sydney. Those are estimates of participation over schooling, not the share attending lessons in any given week.
Their concern was that families already struggled to judge quality in a largely unregulated industry. Cheap software doesn’t resolve that problem. It changes the product parents must assess.
The promise of individualized instruction also predates chatbots. In 1984, the educational psychologist Benjamin Bloom popularized a finding that tutoring could produce gains of two standard deviations, roughly enough to move an average student to the 98th percentile. The figure became a recurring justification for efforts to make personal tutoring affordable.
Paul T. von Hippel, a University of Texas education researcher, revisited that history in Education Next in 2024. He traced the celebrated result to short experiments combining tutoring with mastery learning, including repeated assessment and correction. A later review of 96 studies found an average tutoring effect of 0.37 standard deviations. None reached two.
Sal Khan, the founder of Khan Academy, presented the case for an A.I. tutor for every student in a 2023 TED talk. His demonstration showed a program asking questions and guiding reasoning, rather than simply supplying an answer.
The strongest replacement evidence
A study released on Sept. 23, the day the Sydney closure became news, tested 2,383 adults on Graduate Record Examination material. Its conclusion was unusually relevant to Mr. Buckland’s argument: pooled A.I. tutoring produced immediate learning gains statistically equivalent to expert human tutoring.
Curtis Northcutt and five co-authors published StudentBench as an arXiv preprint, rather than a peer-reviewed journal article. Participants received A.I. instruction, an individual human tutor or unrelated videos. They took assessments before and after an hour in their assigned condition. The study paid participants $50 per completed session, regardless of scores.
The Sept. 30 revision supplies important qualifications. Participants were 18 to 67, with a median age of 21. The pooled A.I.-minus-human difference was minus 0.58 percentage points, with a 90 percent interval from minus 2.18 to 1.03. Equivalence held for the combined quantitative and verbal result and for quantitative questions separately, but not for verbal questions separately.
Mr. Northcutt promoted the result on X, including the comparison between an expert’s hourly charge and the model’s computing cost. That is the economic case for substitution in its most direct form.
Today we establish that AI is as good as expert human tutors for immediate GRE learning gains (p=.015, n=2,383 students, study conducted July-Sep 2026).
One hour with an expert GRE tutor: $75.
One equivalent hour with an AI tutor: 7 cents .
918x cheaper, equivalent learning…
The paper’s cheapest model that met an individual equivalence test cost about 918 times less per percentage point of learning gain than the human comparison. The calculation used model inference spending on one side and a reference tutoring price of $75 an hour on the other. It was not a comparison of complete businesses: software development, support and other deployment expenses weren’t included in the inference figure.
Even with that limitation, the gap was large. Estimated inference costs ranged from $0.067 to $21.24 a session across models. A provider doesn’t need to eliminate every other expense for that difference to affect what it can charge.
The experiment nevertheless tested adults who completed a short, paid session. It did not establish lasting retention, success on an actual GRE or whether a teenager would return to study without someone expecting them. Its strongest finding concerns immediate learning under those conditions.
Another positive result came from Greg Kestin and colleagues’ 194-student Harvard physics experiment, published in Scientific Reports on June 3, 2025. Students using a purpose-built tutor achieved more than twice the median learning gains of students in an active-learning class. The comparison involved specific physics material over two consecutive weeks, not an entire degree or a human tutor working one-to-one.
When help prevents learning
Nearly 1,000 high school mathematics students in Turkey supplied a less reassuring result. Hamsa Bastani and her co-authors compared ordinary practice with two GPT-4-based systems: a general chat interface and a tutor designed with teachers to give incremental hints.
Their paper in Proceedings of the National Academy of Sciences, published in June 2025, found that practice scores rose 48 percent with the general interface and 127 percent with the protected tutor, relative to the control group. But students who had used the general interface scored 17 percent lower than the control group on an exam without A.I. The hint-based tutor largely avoided that loss.
The practice results and the exam results measured different things. One showed what students could accomplish while receiving help. The other tested what they had learned. Knowledge at Wharton reported that students using A.I. also overestimated how much they had learned.
On Hacker News, a reader using the name saadn92 raised that objection to the Sydney firm’s recommendation.
I’m all for saving money where you can, but these models really aren’t great for learning because they love giving you answers and it’s so easy to get an answer, but the process to get to the answer is how you actually learn. Unless you put guardrails in place to actually learn, I don’t think this advice is good.
The Turkey experiment supports the warning about unrestricted answers. It does not support rejecting every A.I. tutor: the protected version behaved differently. Nor do the positive Harvard and GRE results erase the negative finding. They involved different students, different software and different comparisons.
The harder objection concerns the work a tutor does before a student asks a question. Another Hacker News reader, jvvw, described hiring a tutor for one child and valuing the relationship as well as subject knowledge.
Having got a tutor for one subject for one of my children, I’d say that tutors are about motivation/structure/instilling confidence not just subject matter. Our tutor has been great as a general mentor during the period leading up to exams, not just as a subject expert.
That is a parent’s account, not a measured comparison. But it identifies an outcome the short experiments don’t establish: sustained participation through an examination period. A system can explain algebra accurately and still fail a student who doesn’t open it.
Mary Burns, writing in a January review for Brookings, found encouraging evidence for purpose-built generative A.I. tutors, particularly systems that guided students through problems. She also identified continuing problems with accuracy, dependence and personalization. The evidence favors particular teaching designs, not the assumption that any conversation with a capable model is a lesson.
Three business responses
Chegg announced plans to dismiss about 388 employees, or 45 percent of its work force, in October 2025. The homework-help company attributed falling traffic and revenue to A.I. and reduced referrals from Google, Fox Business reported. Its experience shows that students can abandon a paid service without anyone first proving that its replacement teaches better.
A second response keeps the human tutor and supplies software advice. Rose E. Wang and colleagues’ 2025 Tutor CoPilot study, listed by Stanford’s National Student Support Accelerator, involved more than 700 tutors and 1,000 students from underserved communities. Students whose tutors had access to the tool were four percentage points more likely to master mathematics topics. Students of lower-rated tutors gained nine points.
The system cost about $20 per tutor per year. Tutors using it asked more probing questions and offered less generic praise. Here the low-cost product supported a person conducting the lesson, rather than replacing that person.
A third response puts software at the center of the school day. Alpha School’s model attracted attention in the Sydney discussion because it assigns academic work to adaptive learning programs and uses adults called guides. But A.I.-led schooling is a different proposition from a subscription supplementing an ordinary school education.
In March, WBUR’s Here & Now discussed an investigation by Emanuel Maiberg of 404 Media into poorly constructed A.I.-generated lessons and multiple-choice questions that didn’t make sense. Alpha disputed the account, including its descriptions of the technology, privacy practices and learning tools, and said it remained committed to research-backed education.
Neither those complaints nor the school’s response establish a general result for A.I. instruction. They show why a successful laboratory or course-specific tutor cannot validate every commercial product sold under the same description.
The next exam
On Oct. 16, a separate Central News article reported Mr. Dhall’s warning that tutoring providers were exposing children’s identities, schools and whereabouts in promotional material. The concern was about human-run businesses, not simply chatbots. Parents weighing the two face questions about safeguarding on both sides.
Dymocks’s recommendation therefore leaves a more specific decision than whether machines can teach. A family buying explanations and additional practice has reason to consider a cheaper tool. A family buying scheduled attention, diagnosis of persistent mistakes or help keeping a child engaged is purchasing something the replacement studies haven’t fully tested.
The Sydney firm’s own transition arrangements recognized that difference in practice. It recommended inexpensive software while directing families toward other tutoring businesses and preserving individual instruction for about 38 exam candidates.
Those students are now in the examination period. According to the New South Wales Education Standards Authority’s timetable announcement, this year’s 123 written Higher School Certificate exams began on Oct. 13. The final examination, physics, is scheduled for the morning of Nov. 5.