How Quantum Academy uses AI, and why we put live teaching first

A participant in a post-quantum migration course might ask what to do when a critical supplier can’t meet the organization’s timetable. Say the supplier makes embedded controllers for equipment that stays in service for decades, and its quantum-safe firmware won’t ship until a year after the organization’s own deadline.

An AI tool can produce a tidy answer in seconds. Renegotiate the contract, add compensating controls, escalate the risk. Each of those could be right, depending on details the participant hasn’t mentioned. A useful discussion starts with questions of its own. Which system is involved? Is the gap in communications security, firmware signing or the hardware root of trust? What creates the deadline – a regulator, a customer contract or an internal policy? What can the organization actually change? An answer that sounds sensible in the abstract often needs rethinking once those details come out.

That exchange is what we built Quantum Academy to make possible. That explains two decisions about how we work. We use AI to help prepare our material, and we put live teaching by experienced practitioners first. We launched with live courses only, online and in person, because we think discussion with people who do the work is the best foundation for professional education. AI has a useful role in that model, under the direction of people who decide what we teach and answer for it.

What AI does at Quantum Academy

Yes, we use AI, and we’d be foolish not to. Our team uses commercial AI tools in four kinds of work, and our AI Editorial Policy describes them.

• Research. AI search tools help us find standards, regulatory texts, papers and announcements faster. Someone on the team reads each source before we cite it.
• Drafting. A subject-matter expert decides the scope and substance of each piece. We then use AI tools to produce first drafts of course material, web pages, articles and emails.
• Editing. We run drafts through AI checks for clarity, consistent terminology and house style.
• Building. We build and maintain the software behind our website with AI coding tools.

Everything an AI tool produces is a draft until a person has reviewed it. Our instructors and course authors are practitioners with client work of their own. We use AI for the mechanical parts of course production, such as first drafts, formatting and consistency checks, so that our experts can spend their hours on judgment calls, worked examples and the questions learners will ask.

When the rules change

Post-quantum standards and regulations change often, and our courses have to keep up. Between March 2025 and June 2026, the UK, the EU and the US each published timelines for the move to post-quantum cryptography. The UK’s National Cyber Security Centre published migration milestones for 2028, 2031 and 2035. The EU’s coordinated implementation roadmap said member states should start the transition by the end of 2026 and move high-risk use cases as soon as possible, and no later than the end of 2030. In the US, Executive Order 14412, signed on June 22, 2026, directs the Office of Management and Budget to require federal agencies to move their high-value assets and high-impact systems to post-quantum cryptography for key establishment by December 31, 2030. The deadline for digital signatures is December 31, 2031.

When an order like that is published, updating a course takes two kinds of work. Finding every lesson, exercise and page that mentions US federal timelines is search work, and AI tools do it faster than a person rereading every module. Deciding what the order means for learners is judgment work. A security architect at a federal contractor and a CISO at a European bank need different answers from the same document. An instructor who has run migrations decides what changes in the course, and an expert reviews the update before we publish it.

What our reviewers check

Before we publish a course, page or article, a reviewer checks each algorithm name against NIST’s published standards. Each statistic, date, regulatory deadline and quotation is checked against its original source. The reviewer also checks whether each qubit count means physical or logical qubits. For each capability claim, they determine whether it describes a demonstrated result or an announcement.

A first draft might say that a product “supports Kyber.” The reviewer has to establish which Kyber. NIST standardized ML-KEM in FIPS 203 in August 2024, and the standard differs from the round-three CRYSTALS-Kyber submission in ways that break compatibility. In ML-KEM, the shared secret is no longer derived from a hash of the ciphertext, the encapsulation key has to pass input checks, and the final standard added domain separation to key generation. Software written to the round-three submission won’t interoperate with ML-KEM, so “supports Kyber” can describe a product that implements a pre-standard algorithm.

Status needs the same care. NIST’s draft transition guidance, NIST IR 8547, proposes deprecating quantum-vulnerable public-key algorithms that provide 112 bits of security, such as RSA-2048, after 2030, and disallowing quantum-vulnerable public-key algorithms at every security level after 2035. A draft that calls those dates NIST requirements overstates them. As of October 2026, IR 8547 is still the initial public draft that NIST published in November 2024.

Checking a source also means checking that it hasn’t been withdrawn. A 2025 meta-analysis that reported large learning gains from ChatGPT was retracted in April 2026, after other researchers raised concerns about discrepancies in the analysis. An AI tool working from training data that ends before April 2026 can repeat the paper’s findings without mentioning the retraction.

Why experts check what we publish

Fluent text can still be wrong. In a 2024 study, researchers asked general-purpose chatbots verifiable questions about US federal court cases, and the models gave wrong answers between 58% and 88% of the time, depending on the model. Those were 2023 models. In a follow-up study, AI legal research tools built for lawyers still gave an incorrect answer, or cited a source that didn’t support the answer, on between 17% and 33% of queries.

Law isn’t cryptography, and neither study measured post-quantum topics or the tools available today. In both fields, a confident wrong answer costs the person who acts on it, so nothing an AI tool drafts is published until an expert has checked it.

Checking also has to be somebody’s explicit job. In a 2025 survey of 319 knowledge workers, Hao-Ping Lee and colleagues found that people with more confidence in generative AI reported less critical thinking when they used it, while people more confident in their own ability to do the task reported more. Our reviewers are subject-matter experts, and checking AI output is part of the role we give them.

In 1966, the chemist and philosopher Michael Polanyi wrote that “we can know more than we can tell.” Language models are trained on what people have written down, and much of what a migration lead knows was never written down. They know which supplier always misses its dates, how long a change freeze really lasts and which legacy system nobody wants to touch, even if there is no policy against it. A writer without that knowledge can describe the documented process and miss what really happens. So a person who has done the work sets the scope of every course before any draft exists.

People have the first and final say

Human judgment starts before the first draft. Someone with relevant expertise has to decide which problems deserve attention, what the audience already knows, and how much detail a lesson needs to be useful.

In a course on evaluating a quantum computing pilot, a clear explanation of the hardware helps. The learning objectives also have to cover how to choose a workload, how to set up a fair classical comparison and what evidence would justify further investment. The person who sets those objectives should have run or reviewed a pilot.

The same judgment is needed in review. A paragraph can be readable and factually correct and still leave out the qualification that makes it useful. A writer can state, correctly, that hybrid key establishment in TLS combines a classical algorithm with a post-quantum one. A practitioner also needs to know that the post-quantum key share makes the TLS ClientHello larger, and that when Chrome 124 turned hybrid key exchange on by default in April 2024, some servers and network devices failed to handle the bigger message. An expert reviewer asks whether an example exposes the important assumptions and whether a conclusion goes further than the evidence. They also ask where readers should be told what is still uncertain.

Quantum Academy takes editorial responsibility for everything published under its name, including anything drafted with AI tools, and we never list an AI tool as an author. Errors are still possible. If you find one in a course, a page or an article, email support@quantumacademy.com. We check every report and correct confirmed errors.

Exams, results and your data

Multiple-choice and scenario items are scored automatically by our exam software, as our Candidate Guide explains. No AI tool scores your exam or decides your result. Subject-matter experts approve every exam question and set the passing standard for each exam with the modified Angoff method, in which a panel of experts estimates, for each question, the likelihood that a minimally competent candidate would answer correctly. We review every exam booking by hand, along with every request for an exception to the training prerequisite. Two assessors who didn’t teach the cohort mark each capstone in the Certified Chief Quantum Officer (CCQO) program. A person reviews every appeal.

We don’t put learners’ or candidates’ personal information, exam answers or results into AI tools. If a client shares material with us for private team training, we use AI tools on it only with the client’s written agreement. We use AI services under business terms that prohibit the provider from training its models on what we enter.

Why we put live teaching first

Go back to the supplier question. In a live session, an instructor can work through the assumptions with the group. One participant is thinking about equipment that’s expensive to replace and can’t be patched remotely. Another points out that the obvious workaround depends on a maintenance window their organization can’t get. A third has negotiated a similar contract and knows which clauses suppliers accept. By the end of the exercise, the group has identified the conditions that make the recommendation workable.

That exchange needs people who have done the work, in the room or on the call. You can ask an instructor how they reached a conclusion, what would change it and when a different approach would be reasonable.

In 1991, the learning researchers Allan Collins, John Seely Brown and Ann Holum wrote that “the processes of thinking are often invisible to both the students and the teacher.” Their remedy, which they called cognitive apprenticeship, is to make an expert’s thinking visible. The expert works through a problem aloud, coaches learners through their own attempts, and asks them to explain their reasoning and compare it with the expert’s. In a live session, an instructor can do all three and respond to your reasoning as it develops.

In a 2020 meta-analysis of 435 studies, Benedikt Wisniewski, Klaus Zierer and John Hattie found an average effect size for feedback of 0.48. Feedback with more information about the task, the process and how learners manage their own work had roughly twice that effect, at 0.99. Reinforcement or punishment came in at 0.24. In a live session, you explain your migration sequence to someone who has run a migration, and they tell you what’s missing and why. That’s high-information feedback.

In a 2009 study in an undergraduate genetics course, Michelle Smith and colleagues found that discussing a concept question with classmates improved understanding, “even when none of the students in a discussion group originally knows the correct answer.” In a cohort of security architects, risk managers and engineers, the person next to you has often met the constraint you haven’t hit yet.

Who teaches

Each of our instructors is a senior practitioner who still works in the field. Between them, they have led cryptographic migrations and deployed quantum key distribution systems and quantum networks. Some have assembled quantum computers from modular components, and others have briefed boards.

Expertise alone doesn’t qualify anyone to teach. In a 2003 study, Mitchell Nathan and Anthony Petrosino found that trainee teachers with more advanced mathematics backgrounds were more likely to assume students must master symbolic equations before they can solve word problems. Students’ actual results don’t follow that order, and the researchers called the pattern expert blind spot. Every Quantum Academy instructor passes a technical interview with faculty in their field and a recorded teaching audition, then co-teaches or shadows a live session with an existing faculty member before teaching alone. Our selection process is public.

A scheduled session alone doesn’t guarantee that kind of exchange. Our instructors plan live time around exercises, group problem-solving and realistic scenarios that the cohort works through together, and the written materials are yours to keep afterward.

Teaching format and learning: the research

When researchers have compared formats directly, teaching methods have counted for more than the format. In a 2006 meta-analysis, Traci Sitzmann and colleagues found that web-based instruction was 6% more effective than classroom instruction for teaching declarative knowledge, and that the two were equally effective for procedural knowledge. When both formats used the same instructional methods, the declarative-knowledge difference disappeared, “suggesting media effects are spurious,” in the authors’ words.

In a 2014 analysis of 225 studies in undergraduate science, engineering and mathematics, Scott Freeman and colleagues found that exam scores rose by about 6% under active learning, and that students in traditional lectures were 1.5 times more likely to fail. In active learning, the learners do the work in class, solving problems and discussing them, while the instructor guides.

Active learning also feels harder while it’s happening. In a 2019 study in Harvard’s introductory physics course, Louis Deslauriers and colleagues randomly assigned students to active-learning sessions or to polished lectures, then switched the groups. Students learned more from the active sessions and felt they had learned more from the lectures. Deslauriers told the Harvard Gazette that “the effort involved in active learning can be misinterpreted as a sign of poor learning.” A smooth lecture, a slick video and a fluent AI explanation can all feel like learning.

Researchers at the US Military Academy at West Point ran a direct comparison of online and in-person teaching. In fall 2020, Michael Kofoed and colleagues randomly assigned 551 cadets in a required economics course to online or in-person sections, and each of the twelve instructors taught half their sections online and half in person. Final grades in the online sections were 0.215 standard deviations lower. In a post-course survey, online students said they had struggled to concentrate in class and felt less connected to their instructors and classmates. The cadets took a required course during a pandemic semester. Senior professionals who enroll by choice may respond differently. Within those limits, cadets taught in person did better.

In a 2022 study in Nature, Melanie Brucks and Jonathan Levav ran two experiments, one in a lab and one in the field with 1,490 people at five country sites of a telecommunications infrastructure company. Pairs working face to face produced more ideas than pairs working over video, and more creative ones. The researchers attributed the gap to focusing on a screen, which narrowed participants’ cognitive focus. Video pairs were no worse at choosing which idea to pursue. Brucks and Levav tested creative idea generation in pairs, and their results don’t extend directly to professional training or migration decisions. With that caution, we read them as a reason to favor in-person time for exploratory work, such as listing ways around a supplier who can’t deliver.

University students took part in most of these studies, and none of them establishes a single best format for every learner or subject. They do show the value of practice, explanation and useful feedback. We prioritize live courses because they let participants question experienced practitioners, compare constraints and work through decisions together. In Kofoed’s West Point study, cadets taught in person earned higher final grades than cadets taught online. Our on-demand courses will apply the same principles through structured lessons, worked examples and exercises with explanatory feedback.

AI as a tutor

In a 2025 randomized study of 194 students in Harvard’s introductory physics course, Gregory Kestin and colleagues compared an AI tutor used at home with an active-learning class on the same material. The study covered two lessons, with the groups switching conditions, and measured learning with a test after each one. Median learning gains in the AI-tutored group were more than double those in the classroom group. The median student spent 49 minutes with the tutor, against about 60 minutes of learning time in a 75-minute class. The instructors built the tutor themselves and wrote step-by-step solutions into its instructions. It took students through each problem one part at a time.

The authors were careful about where it fits. They wrote that an AI tutor “should not replace in-person teaching” and should be used “to bring all students up to a level where they can achieve the maximum benefit from their time in class.” They also wrote that they did not “presume that structured AI tutoring will always outperform in-class active learning in all contexts.”

A 2025 paper in PNAS is titled “Generative AI without guardrails can harm learning.” Hamsa Bastani and colleagues ran a field experiment with nearly a thousand high-school students in Turkey. Some classes practiced with a plain GPT-4 chat interface, and others with a version that included teacher-written solutions and instructions to give hints one step at a time. With the plain chatbot, practice grades rose 48%. On the exam, taken without AI, those students scored 17% lower than the control group, which never had AI. With the guided version, practice grades rose 127%. Exam results in those classes were statistically no different from the control group’s. The researchers concluded that students had used the unrestricted chatbot as a crutch.

Kestin’s and Bastani’s results show why help during practice and independent learning have to be measured separately. Kestin’s instructor-built tutor improved learning on the lessons tested. In Bastani’s experiment, the guardrails raised practice grades and removed the harm on the unaided exam, but they didn’t produce an advantage over the control group there. Our AI Editorial Policy covers four kinds of work, and tutoring isn’t one of them. We expect teaching tools to keep improving, and we’re open to learning from that work.

On-demand courses

We will also offer on-demand courses for people who can’t attend a scheduled session, or who want to work at their own pace. Structured written lessons are the core of each one. Each diagram appears next to the text it illustrates, and we write explanatory feedback for every exercise, so if you get one wrong, you see why. Short videos, typically three to seven minutes long, explain a concept or demonstrate a tool, and instructors and other senior practitioners contribute short pieces on decisions they’ve faced. How We Teach describes the design in full.

Everyone who appears or speaks as an instructor, expert or practitioner in our course videos is a real person. We don’t use AI-generated avatars to impersonate instructors or experts, and we don’t use synthetic voices to present personal experience or professional judgment as if it came from a real practitioner. If we use synthetic media for another purpose, such as accessibility, localization, animation or demonstration, we won’t present it as a real person speaking from experience.

Avatars can convey facts. In a 2025 study, Torbjørn Netland and colleagues showed 447 online participants short operations-management videos made either by teachers or with AI tools. Test results showed no statistically significant difference, although participants enjoyed the human-made videos more. We keep our rule for a different reason. When someone presents a decision as their own professional experience, you should be able to trust that it was.

The format of a course and the use of AI are separate choices. A recorded course can be taught by a human expert, and a live one can use material prepared with AI assistance. We tell you how each course is delivered and who is responsible for its substance.

Live courses will remain our priority. We’re building Quantum Academy around access to experienced instructors and the chance to learn through discussion, and we’ll add other formats only where they serve learners without weakening that.

Questions to ask any training provider about AI

Whichever provider you choose, including us, you can ask six questions before you enroll.

1. Who decided what this course covers, and what have they done in the field?
2. Who checked the material, and against which sources?
3. Who will answer my questions during the course, and do they still work in the field?
4. Is everyone presented as an instructor or expert a real person?
5. Does an AI tool score my exam or decide my result?
6. What happens to my personal data, or my company’s material, if I share it?

Our answers are in our AI Editorial Policy, and we update it when our tools or practices change.

Back to the supplier question

AI helps us arrive better prepared. When a participant asks what changes in their own situation, we want an experienced instructor there to work through that question with them, in person or live online. You can see upcoming courses on our live training schedule and meet the people who teach them on our instructors page.