We read a hundred reviews of the top AI courses — here's what people actually hate

In short: on 17 July 2026 we pulled numbers straight from Udemy's internal API and from Coursera's review pages — 14 courses on AI and prompting, 6,208,053 enrolments, 557,032 reviews on the Udemy side alone. On the two most popular Udemy courses, the share of ratings at 3.5★ or below is 9.61% and 10.36%. On Coursera it's 1.5–3.5%. The obvious conclusion is that Coursera is three to seven times better. The conclusion we actually reached, after looking at the denominators, is the opposite one — and it's far more useful to you. Meanwhile the verbatim complaints don't add up to "bad lecturer". They add up to six repeating structural holes that picking a different course on the same platform will not fix.

What we measured, and what we didn't

No magic involved: we took the eight biggest non-technical Udemy courses on ChatGPT and generative AI and the six Coursera flagships, pulled the course records out of their own API, and read the reviews. This is our primary source; these figures don't appear in any public write-up.

The limits, stated up front, because you'll need them shortly. Coursera's star filter runs in the browser — their server HTML always returns page one. So the Coursera quotes you're about to read come from the default listing: they're exactly what any human sees on opening the course page, not something dredged from an archive. We could not confirm the exact price of ChatGPT Plus: every OpenAI domain returns 403 to us. Reddit is closed to our crawler. And we did not measure what people learned — we measured what they complain about. Those are different objects and must not be confused.

The scale: this isn't a niche, it's a stadium

Before criticising anything, take in the size. Coursera:

• Google AI Essentials — 1,876,929 enrolments, 22,557 reviews, 4.8 rating, roughly 8 hours.
• Generative AI for Everyone, DeepLearning.AI (Andrew Ng) — 814,083, 5,060 reviews, 4.8, 6 hours.
• Prompt Engineering for ChatGPT, Vanderbilt (Jules White) — 698,444, 7,915 reviews, 4.8, 19 hours.
• Generative AI: Prompt Engineering Basics, IBM — 654,256, 2,313 reviews, 4.7, about 10 hours.
• Google Prompting Essentials — 365,425, 7,219 reviews, 4.8, under 10 hours.
• Prompt Engineering Specialization, Vanderbilt — 138,967, 9,151 reviews, 4.8, 39 hours.

And Udemy:

• Generative AI for Beginners (Aakriti E-Learning) — 409,492 students, 121,079 reviews, 4.53, 4.5 hours, 29 lectures, updated 18.04.2026.
• The Complete AI Guide (Julian Melanson and others) — 376,845, 62,104 reviews, 4.52, 42 hours across 545 lectures, updated 23.06.2026.
• ChatGPT: Complete Course For Work (Steve Ballinger) — 281,823, 130,274 reviews, 4.46, 16.5 hours, 158 lectures, updated 01.07.2026.
• How to use ChatGPT and GenAI for content (Justin Barnett) — 152,821, 76,216 reviews, 4.48, 7 hours, 75 lectures.
• Intro to ChatGPT and Generative AI (365 Careers) — 125,040, 55,238 reviews, 4.51, 2.5 hours, 50 lectures.
• Prompt Engineering with ChatGPT Masterclass (RPATech) — 118,803, 57,836 reviews, 4.48, 4 hours, 41 lectures.
• ChatGPT & GenAI — The Complete Guide (Academind) — 110,183, 29,884 reviews, 4.53, 21.5 hours, 382 lectures.
• Prompt and Context Engineering 101 (Mike Wheeler) — 84,942, 24,401 reviews, 4.31 — the worst rating in our sample — 2.5 hours, 34 lectures, updated 15.07.2026.

Add the enrolments up: 4,548,104 on Coursera and 1,659,949 on Udemy — 6,208,053 across fourteen courses. The caveat matters: an enrolment isn't a person, and certainly isn't a completion. The same human can enrol on three of them, and Coursera enrolment is free, so it includes people who never open lesson one. This is not an audience figure. It's an upper bound.

One more thing worth lifting straight out of that table: no course requires code anywhere. Verbatim from the prerequisites: "No prerequisites as ChatGPT is a tool anyone can access and use immediately" (Ballinger); "No coding skills needed" (Wheeler); "A desire to learn" (Aakriti); "No prior experience with AI or programming is needed" (Complete AI Guide). The "beginners, no code" niche is not an opening. Count exactly what we just quoted: those four courses hold 1,153,102 students between them (Aakriti 409,492 + The Complete AI Guide 376,845 + Ballinger 281,823 + Wheeler 84,942). More than a million people have already bought the promise that no code is needed — from four instructors out of eight.

9.61% versus 1.5% — and why that isn't what it looks like

Now the headline number. Share of ratings at 3.5★ or below:

• Generative AI for Beginners — 9.61% of 121,079 reviews. That's roughly 11,600 unhappy people.
• The Complete AI Guide — 10.36% of 62,104 reviews. About 6,400 unhappy people.
• Prompt and Context Engineering 101 — 4.00% or more (on the worst rating in our set, 4.31).
• Coursera across the board — 1.5–3.5%.

A three-to-sevenfold gap. The temptation to read it as "Coursera is three times better" is enormous, and nearly every article on this subject would stop right there. We looked at the denominator instead, and the conclusion fell apart.

Do the arithmetic yourself, on our own numbers. Ballinger has 130,274 reviews from 281,823 students — every second student left one (46%). Aakriti: 121,079 from 409,492, nearly one in three (29.6%). Now Coursera. Google AI Essentials: 22,557 reviews on 1,876,929 enrolments — 1.2%. Andrew Ng: 5,060 on 814,083 — 0.62%. IBM: 2,313 on 654,256 — 0.35%.

The review density differs by a factor of thirty to eighty. That is not "the same people, only happier". That is a different opinion-collection machine. On Udemy you paid €19.99, the course sits in your account and asks you to rate it; a disappointed person who paid is annoyed, and annoyance walks all the way to the form. On Coursera enrolment is free: the person who gave up two minutes in simply closes the tab and never returns. He isn't a negative review — he isn't a review at all. What reaches the form is mostly the people who finished, and people who finish are, by construction, the satisfied ones.

So here's the honest conclusion, which is inconvenient for us as much as anyone: 1.5% negativity on Coursera and 9.61% on Udemy is not a measurement of quality. It's a measurement of who got a microphone. We are not claiming Coursera is worse. We're claiming that no quality verdict follows from those two numbers, and anyone who draws one — including us, had we not checked — is drawing it out of thin air. The useful signal lies elsewhere: not in the share of complaints but in their content. And the content matches across both platforms. That coincidence is the real finding.

Complaint #1: "he just reads off the slide"

The most frequent and the angriest. Verbatim:

"why read straight from the slide? I can do that. This was not a helpful course at all"
— Janie I., 02.07.2026, 1★ (Justin Barnett)

"50% of this course is reading script like a robot from the slides. …they are just reading text from the slides which you can also you from any good website"
— Shashank T., 13.04.2026, 1★ (The Complete AI Guide)

"They are just reading the prompter sometimes without even knowing the point. Hating myself after purchasing it."
— Sachin S., 13.07.2026, 1★ (The Complete AI Guide)

"Written by AI, delivered by AI. The slides are way too crowded to be useful and it really doesn't help to have the bot read them out word-for-word"
— Bradley M., 16.04.2026, 1★ (RPATech, 118,803 students)

"AI voice, not recommended. didn't learned anything"
— Ashvin D., 22.06.2026, 1★ (RPATech)

Look closely at Janie's four words: "I can do that." That isn't a complaint about diction. It's a precise economic diagnosis: what exactly did the money buy, if the only difference from the text on the screen is that someone says it out loud? Video earns its cost where it shows what text cannot carry — hands moving through an interface, a train of thought, a mistake and its repair. A narrated slide is text with the reading speed, the page search and the ability to re-read a paragraph surgically removed. A reader controls the pace. A viewer doesn't.

Note also what Bradley M. is describing: an AI-written script read out by an AI voice, sold as a course, to 118,803 people. His complaint isn't about the technology. It's that the format claims a human's presence and doesn't deliver one.

Complaint #2: "I submitted my work and instantly got 100% — and learned nothing"

This one isn't about a lecturer. It's about how the platform is built, and it's the most interesting finding in the set.

"you send your assignments and immediatly you got your results: 100% correct. I am still speechless. I put all that effort in and have no idea whether I my answer was correct or not."
— Shinysheep, 21.09.2023, 1★ (Vanderbilt Prompt Engineering)

"Videos are too short and superficial so you end up memorizing sentence by sentence to pass quizzes. Not a learning experience."
— Laurie J Phillips, 12.04.2024, 3★ (IBM)

"quizzes give unhelpful feedback for incorrect answers and just say 'watch the video again'"
— Cory Covino, 04.05.2024, 2★ (IBM)

"All the coding is done in the labs for you. You won't have to debug anything or figure anything out, just press shift-enter."
— Cornelius Griggs, 1★ (Generative AI with LLMs)

Walk through the mechanics. The course teaches prompt writing. The student writes a prompt and submits it. The platform replies: 100% correct. How does it know? It doesn't: there is no model inside the Coursera lesson. The platform physically cannot run your prompt and look at what came out. It can grade a multiple-choice quiz, and there its powers end. "100% correct" here does not mean "your prompt is good". It means "you pressed the button".

This isn't one author cutting corners; it's a constraint of the construction. And it explains the second quote: when there's no way to test skill, you test memory instead — hence "memorizing sentence by sentence". It explains the third one too. To tell a student why an answer is wrong you have to understand the answer, and a quiz understands only its number. "Watch the video again" is the only sentence the machine is honestly capable of.

Here we're obliged to say the uncomfortable thing about ourselves, or none of the above is worth reading. On our platform prompts run inside the lesson — every prompt block in this article has a Run button, and you're about to confirm that yourself. That closes the exact hole Shinysheep fell into. But it does not mean we grade your skill: a live model response tells you what your prompt produced, not whether your prompt was good. You still judge. The difference between us and an automatic 100% isn't that we have a verdict — it's that we have a fact to judge from. The automatic grader issues a verdict with no fact behind it. That's worse than a fact with no verdict, but a fact is not a verdict either, and we won't pretend otherwise.

Test that claim now. The prompt below makes a model grade someone else's answer against explicit criteria — and, more importantly, name what it cannot grade:

You are a strict grader. Below is a student's prompt and the answer a model gave.

STUDENT'S PROMPT:
"Summarise this quarterly report for my boss."

MODEL'S ANSWER:
"The report shows strong growth across key metrics, with revenue trending
upward and several areas of improvement identified. Overall the quarter was
positive, though some challenges remain that management should address
going forward."

Do five things:
1. Score the answer 0-2 on each criterion, one sentence of justification each:
   specificity, verifiability of claims, usefulness to a busy executive,
   absence of filler.
2. Extract every phrase in the answer that would stay true for ANY quarterly
   report whatsoever. Report them as a percentage of the text.
3. Name what you fundamentally CANNOT assess here, and explain why.
4. State what the student had to give you for your grade to mean anything.
5. Rewrite the student's prompt so that its answer becomes checkable, and name
   the exact check I could then run against the new answer.

Watch point 3. The model will most likely say: "I never saw the report." That is precisely the sentence the platform should have returned instead of "100% correct".

Complaint #3: "you promised prompting and gave me an overview of prompt types"

We deliberately pulled these from courses with prompt engineering in the title. This isn't nitpicking the marketing — it's a gap between the promise and the contents:

"There is nothing teached about creating a good prompt. It is just an overview of types of prompts."
— Geralt O., 01.07.2026, 2★ (Mike Wheeler)

"No specific guidance on prompt engineering… what to avoid while asking, how to organize your thoughts, how to give feedback to AI based on its answers etc."
— Bharat Ram A., 05.06.2026, 1.5★ (Mike Wheeler)

"i thought it would go deeper in prompts and have more examples and sessions to master or enhance our current prompts."
— Manuel L., 15.06.2026, 2★ (Prompt Engineering for Everyone)

"Too much background on ChatGPT. Get me to how to prompt GPT"
— Mark R., 16.09.2025, 1.5★ (Academind)

Read what Bharat Ram is actually asking for: "how to give feedback to AI based on its answers". That is not "give me the prompt formula" — everybody gives the formula. It's a request to be taught the loop: look at the output, work out what's wrong, change precisely that. The prompt didn't work — what do you change first? The role? The format? The example? The length constraint? Nobody teaches this, and the complaints in our sample ask for exactly it.

Note that Wheeler's course — the 4.31, the worst we found — was updated on 15.07.2026, two days before our snapshot. Freshness didn't save it. The complaint is structural, not about staleness.

Here is debugging in its pure form: the thing that was paid for and not delivered. Run it:

Below is a weak prompt. Do NOT execute it — diagnose it.

WEAK PROMPT:
"Write a professional LinkedIn post about our new product. Make it engaging."

Do six things:
1. Number every decision this prompt silently dumps on you (audience, length,
   tone, what the product even is, the goal of the post, the call to action...).
2. For each one, state the default you would have picked, and why that default
   is probably the wrong one.
3. Rank the missing pieces: which SINGLE addition would change the output most?
   Second? Third? Rank by size of effect, not by order of appearance.
4. Rewrite the prompt three times: (a) minimal fix — one added sentence;
   (b) a full brief; (c) deliberately over-specified.
5. For each of the three versions, name an OBJECTIVE check I could run on the
   output — something countable or verifiable, no opinions.
6. Say at which version the extra detail started hurting rather than helping,
   and name one thing I will have to judge for myself regardless.

Point 3 is the heart of the whole subject. The order of the fixes, not the list of them, is what separates someone who can repair a prompt from someone who has read about types of prompts.

Complaint #4: filler and repetition

"I finished week 2. And all the lessons could be in 1 lesson. I jumped to week 3, 4, 5 and 6, and I see some concept reapeted. Very bad time spending for me."
— Shai Mizrachi, 22.07.2023, 1★ (Vanderbilt)

"The teacher is saying a whole lot of nothing. going to see if I can get a refund."
— David Prince, 03.07.2023, 1★ (Vanderbilt)

"The majority of the class he just rambles around unimportant subject… There's really no "meat" in this course. A major waste of time and money!"
— Kevin K., 20.02.2026, 1★ (ChatGPT for Work)

"The course is far too drawn out… padding the course with filler and making it longer than it needs to be."
— Victoria X., 13.02.2026, 2★ (The Complete AI Guide)

Now hold that against the table. The Complete AI Guide: 42 hours, 545 lectures. Academind: 21.5 hours, 382 lectures. And Aakriti — the single most popular course in the entire sample — 4.5 hours, 29 lectures. A tenfold spread on the same subject, with all of them demanding the same prerequisites: none.

Which means duration does not describe the amount of knowledge. It describes an author's decision about slicing. "545 lectures" on the sales page reads as "more for the same money"; in the reviews the same fact reads as "padding the course with filler". Both readings come from one number, and the number tells you nothing on its own.

Complaint #5: "only theory, which I already knew"

"The content of this course is incredibly simple to the point of uselessness, it is essentially stating that generative AI exists and listing a bunch of example models."
— Nicholas Munford, 1★ (Generative AI: Introduction and Applications)

"Didnt meet expectations, only theory is discussed which we already know."
— Aditya Nagavolu, 19.11.2023, 1★ (Generative AI for Everyone)

"you know quite little, just a thin, very thin layer of knowledge, not even enough for friends discussion. To pay 45$ fo this is realovercharge"
— Jean-Philippe Bernard, 27.08.2019, 1★ (AI For Everyone)

"This is NOT a beginner course. After listening to the 20-minute lecture and reading some of the material, I was more confused than before I started."
— Zlatica S Hoke, 02.07.2023, 1★ (Introduction to Generative AI)

Those last two contradict each other — and that isn't a sampling error, it's the problem itself. A "course for everyone" serves both the person hearing the word "prompt" for the first time and the person who has used ChatGPT daily for six months. The first finds it hard; the second finds it empty. A mass course aims at the middle and misses both edges, while on the sales page the two of them look identical: "for beginners, no experience needed".

Which gives one cheap action before buying: measure where you actually stand. Not by feel — by questions.

Build me a 10-question diagnostic on practical work with generative AI.

Rules:
— No trivia: no dates, company names, history, or model names.
— Only things that change the result in practice: how to repair a prompt that
  returned the wrong thing; when the model's confident tone means nothing;
  how to verify an output; how context and limits work; when NOT to use AI
  at all.
— Rising difficulty: questions 1-3 — anyone who has opened a chat knows them;
  4-7 — the level of a typical beginner course; 8-10 — what beginner courses
  usually skip.
— Multiple choice, exactly one correct option.

Output format:
First questions 1-10 with NO answers.
Then a line exactly like this: === KEY ===
After it, for each question: the correct option, one sentence of why, and one
sentence on the practical mistake made by someone who answered wrong.
Finish with: what a score of 8+ out of 10 implies about buying a beginner
course, and what to look for instead.

Answer all ten honestly before you scroll to the key. If you score 8 or more, a "for beginners" course will sell you what you already own — and, per the reviews, take between 2.5 and 42 hours to hand it over.

Complaint #6: content rots, and the "updated" badge doesn't show it

"Most content is from 2024. This course is not bad for its time, but just too dated now."
— Martin F., 27.05.2026, 2★ (Generative AI for Beginners)

"The content is mostly from 2023. …I invested my 41 hours and Im learning content which is from 2023. Very disappointed."
— Harsh A., 09.04.2026, 1.5★ (The Complete AI Guide)

"Stable diffusion, Dalle and Midjourney are not GAN architecture powered. They're diffusion models."
— Jose C., 12.05.2026, 1★ (Justin Barnett)

The first quote is a direct hit on the sales page. Martin F. writes "most content is from 2024" on 27 May 2026, about a course whose record says "updated 18.04.2026". Both can be true at once: editing one lecture refreshes the date on the whole course. Which means the "updated" badge does not measure the freshness of the material — it records that some edit happened. As a selection criterion it is worthless, and it is one of the two things buyers look at hardest.

The third quote is its own genre. A student is correcting the course on a matter of fact: Stable Diffusion, DALL·E and Midjourney aren't GANs, they're diffusion models. The error sits in the lectures and in the quizzes. Which means a student who knows the right answer is required to select the wrong one in order to pass. That is where the value of the certificate ends — not as a rhetorical flourish, but literally: the credential now certifies compliance with an error.

And here's the consequence that explains the whole picture. Text is fixed in ten minutes. A 42-hour video course is not reshot — that's lighting, editing, voice, 545 files. The economics of repairing video make it rational to touch nothing and refresh the date instead. Not because authors are crooks, but because the alternative doesn't add up. Harsh A. invested 41 hours to find that out.

Complaint #7: pace and transcripts

"It is too fast for a beginner. …the screens they are displaying while teaching is so fast that, until you try to see and understand it the next screen pops up"
— Amit A., 29.04.2026, 2★ (The Complete AI Guide)

"I am very frustrated that I can not easily access a transcript to reference later. The instructors are talking so quickly - I can not follow."
— Vince D., 16.03.2026, 1★

This is complaint #1 seen from the other end. With text, the reader sets the pace, and the transcript exists by definition — it is the text. Video doesn't always lose: you cannot show hands moving through an interface in prose. It loses precisely where the video is a narrated document.

The money: there is no Udemy discount

Now the reason we went into the API at all. The received wisdom is that Udemy courses cost €199 and you, by happy accident, always arrive during a 90%-off sale. Here is the raw response from their own API:

"price": {"amount": 24.99, "currency": "EUR"},
"list_price": {"amount": 24.99, "currency": "EUR"},
"saving_price": {"amount": 0.0},
"has_discount_saving": false,
"discount_percent": 0

It reads unambiguously: saving_price: 0.0, has_discount_saving: false, discount_percent: 0. Price equals list price. There is no discount. Not "the discount is small" — it does not exist as an object. Eleven of the twelve top courses we priced cost €19.99; one (The Complete AI Guide) costs €24.99.

The practical conclusion deserves its own line: the countdown timer on the page has nothing to do with you. The price will be the same tomorrow. The only thing the timer accomplishes is stopping you from reading one page of reviews filtered to one star — which is the single most valuable thing you could do before buying.

The rest of the economics:

• Udemy Personal Plan — €20.00/month, promo price €10.00/month.
• Coursera Plus — €50/month or €343/year, with a 14-day refund window.
• Standalone Google programmes on Coursera — $49/month after a 7-day trial.
• The tool itself: Google AI Pro — €21.99/month (Spanish pricing), Google AI Plus — €4.99/month, Google AI Ultra — from €99.99/month.

And one fact to know before paying: Vanderbilt's Prompt Engineering Specialization requires a paid ChatGPT+ subscription to complete the assignments. Add it up: Coursera Plus at €50/month plus a model subscription (we could not confirm the exact ChatGPT Plus price — OpenAI returns 403; secondary sources put it around $20/month). That lands at roughly €70 a month to watch videos and receive an automatic 100%. The student pays for the platform and for the model, and there is still no check — because the platform cannot see what your subscription produced. You bought the grader's missing organ and it still can't use it.

What to do with all this if you're choosing a course

We're not going to finish with "so buy ours". Here is what follows from the data, most of which applies to any course, ours included.

• Ignore the rating; read the composition of the complaints. 4.53 and 4.31 differ in nothing useful. But 9.61% of people all writing the word "slide" is a completely different object from 9.61% writing "too fast".
• Split complaints into taste and structure. Pace, voice, manner — taste; those may suit you fine. "Nobody checked my prompt" and "it never taught me to fix a prompt" — structure; a different course on the same platform is built the same way.
• Don't trust the "updated" badge. Judge freshness from reviews in the last two months, not from the date on the card. Martin F. is the ready-made proof of why.
• Hours and lecture counts don't measure knowledge. 42 hours and 4.5 hours promise the same outcome and demand the same prerequisites. The difference lives in the slicing.
• Check whether a paid subscription is required on top. Vanderbilt needs ChatGPT+; that isn't in the price on the page.
• Don't hurry because of the timer. There's no discount — it's visible in their own API. Spend the freed-up minutes on the one-star page.
• Ask the course one question: what will check that I learned this? If the answer is "a multiple-choice quiz", the course will test your memory, not your ability. That's not automatically bad — just don't mistake it for a skill check.
• Measure yourself before, not after. The diagnostic prompt above costs five minutes and can save you €19.99 and forty-one hours.

And here's a tool that does that job for you, on a live specimen of sales-page promises:

Below is a typical "what you'll learn" list from an online course sales page.

THE LIST:
— Understand what generative AI is and how it works
— Explore the different types of prompts
— Master ChatGPT for work and productivity
— Learn prompt engineering best practices
— Discover 50+ AI tools to boost your workflow
— Lifetime access and a certificate of completion

For EACH bullet, do three things:
1. Rewrite it as a task I could be given, with an objective pass/fail criterion.
   If it cannot be turned into one, write "unfalsifiable claim" and say why.
2. State what a person who has "got" this bullet can do on Monday that they
   could not do on Friday. One sentence, with an action verb.
3. Estimate honestly, in minutes, how long transferring this skill takes —
   if it is a skill at all.

Then answer:
— which bullets are the same bullet in different words?
— which bullet is not a learning outcome at all, but a purchase term?
— which ONE bullet is worth money, and why?
Finish with a single sentence: what is missing from this list that had to be
on it.

That last question is the valuable one. As a rule, the model's answer lands on exactly what Geralt and Bharat Ram were asking for: repairing the thing that didn't work.

What our data doesn't say — and where we're weak ourselves

An article that ends only with conclusions flattering to its author shouldn't be believed. So: the boundaries.

We don't know what those six million people learned. We read complaints, and complaints are systematically skewed — satisfied people stay silent. The share of negativity, as we showed above, doesn't measure quality in the first place. Coursera's star filter isn't available to us server-side, so their quotes come from the default listing. We could not confirm the exact ChatGPT Plus price. Our sample of Udemy courses is eight, deliberately chosen for popularity, which is a bias in itself: these are the courses that succeeded at selling. And we measured one snapshot, on 17 July 2026 — courses get edited, numbers will drift.

Now about us. We do not have 121,079 reviews, and we cannot show you our negative share — not because it's flattering, but because at our size the number would mean nothing. Ballinger collected 130,274 ratings; that kind of evidence is bought with years and hundreds of thousands of students, and we haven't bought it. What we can show is the thing they don't have physically: a prompt that runs inside the lesson, where you see the output yourself rather than reading a description of it. Our prompt engineering course is built around complaint #3 — around the loop "it didn't work → what do I change first". Whether that works is something to judge from the Run buttons in this article, not from our claims about ourselves: the four prompts above are made of the same material as the course. Run them and decide.

Next in this topic: the prompt engineer profession, free AI tools and their limits, what a prompt is and AI agents in plain words.