When I was sixteen I worked at Asda, and before they let me near a cage of stock I had to sit a health and safety test. Manual handling, hazard spotting, what to do about a spillage, why you do not run a pallet truck over your own foot. You could fail it. People did fail it. There was a right answer and a wrong answer and the wrong answer had consequences that a supervisor could point at.
I hold several AI platform certifications now. I have done the OpenAI ones and I am partway through the Anthropic set. And I want to be careful here, because this is not the usual sneering: the Asda test was harder. Not more important. Harder. It asked more of me, it was invigilated by someone standing there, and failing it meant something.
The thing nobody says out loud
Most of what I sat asked me to recall things a model will tell you instantly. Which service does which job. What the tiers are called. Which feature belongs to which product name. This is exactly the category of knowledge that has become least valuable, because it is the category the tools now hold for you.
And then the recursion, which is genuinely funny once you see it. Many of these are online and unproctored. Nothing stops a candidate having the vendor's own model open in a second tab, answering questions about itself, in its own words, correctly. I am not going to pretend I do not know how that would go.
An AI certification you can pass using AI is not a test. It is a form. That is not a moral complaint about people who take them, it is a design observation. If the assessment cannot distinguish between someone who knows the material and someone with a browser tab open, then whatever it is measuring, it is not competence.
Contrast that with the shop floor. You cannot ask a language model to carry a cage safely on your behalf. The test worked because the skill and the assessment were the same activity.
So what can a certification still measure?
Only things that survive the model being available. There are three, and the industry is thin on all of them.
Recall under invigilation. A proctored exam still proves you carried it in your head. Whether that matters is arguable, but it is at least a real filter, and it is the only thing on this list most vendors currently attempt.
Judgement under constraint. Given a budget, a deadline and a nervous stakeholder, what do you cut? Models are good at options and bad at consequences, because they have never had to look someone in the eye and explain the overspend. Almost nothing tests this.
Work you can inspect. Assessment against something you actually built. Expensive to run, which is why it barely exists, and worth more than the rest combined.
What the market actually offers
We track this daily on the Cert Ledger, so rather than assert, here is the live picture at the time of writing.
16 of Microsoft's 18 were added or amended in the last 45 days
4 proctored Anthropic exams, all gated behind partner membership
0 certifications from Cursor, or from xAI
5-10 hours per OpenAI partner programme, self-paced and unproctored Live at /certifications/. Counts move; the shape does not.
Take the first number seriously. Microsoft rewrote sixteen of its eighteen AI credentials inside six weeks. That is not a scandal, it is a vendor keeping pace with its own products. But think about what it does to the value of the badge. A qualification whose syllabus turns over faster than you can put it on a CV is not a qualification, it is a subscription. You are not certified. You are current, briefly.
And it is not one vendor's quirk. Inside the OpenAI partner portal, where I did mine, superseded programmes sit in the catalogue alongside their replacements, tagged as the old version. You can watch a curriculum turn over in place. I have completed things there that have since been relabelled as the previous generation, which is a strange feeling for something described as a qualification. Nobody did anything wrong; the products moved. But a credential that ages the way a phone contract ages is not the thing the word implies.
The second number is the interesting one. Anthropic's four exams are properly proctored through Pearson VUE, which by any honest standard puts them above every completion badge in this market. They can fail you. But all four require a company email tied to a partner organisation, so however technical the content, the credential is structurally a channel qualification. Both things are true and neither cancels the other.
The third number is the one I would think hardest about. The tools I actually work in all day certify nobody. There is no Cursor certification. There is no Grok credential. The agentic workbench where the job now happens has no exam, no badge and no framework, while the platforms have a growing catalogue. Certification is tracking where the commercial relationships are, not where the work is.
Fine. So is it bragging rights?
Let us be adults about it. A large share of certification has always been signalling, and signalling is not fraud. You are telling a stranger something about yourself cheaply, because the alternative is asking them to take your word for it. That is a legitimate thing to want, particularly early on.
So if what you need is a signal, buy a good one. The test is not how impressive the badge looks on a banner, it is whether it survives three questions from someone who knows the field.
- Could it have failed you? Proctored, timed, a real pass mark. If completing the material guarantees the certificate, it is a receipt for attendance.
- Can a stranger verify it without asking you? A registry entry beats a PDF and a LinkedIn banner.
- Does the market price it? If it appears in job specs, it does work. If it appears only on the issuer's marketing, it does work for them.
Applied honestly, that gets you to a short answer: one proctored exam on the platform you actually use. One. The AWS, Azure and Google AI exams clear the bar. Anthropic's clear it too if you can register. After the first, the returns collapse, because the second badge tells a hiring manager nothing the first did not, and the sixth tells them something you did not intend.
The tell. Six entry-level badges across six vendors does not read as breadth. It reads as someone who found the certifying easier than the doing, and every engineering lead I know reads it exactly that way. One hard credential and one thing you built beats a wall of them, every time.
The part where I argue against myself
Everything above is a comfortable position for someone with a track record and people who will vouch for him. That is worth saying plainly, because the cynicism has a cost and it does not land on me.
If you are switching careers at forty, or you are in a country where nobody has heard of your university, or you simply do not know anyone in this industry, a certification may be the only lever you can pull on your own. It is cheap next to a degree, available anywhere, and does not require knowing the right people. Plenty of very good practitioners got their first interview off a badge that someone like me would call thin. That interview is the whole point.
So the honest position is narrower than "certs are rubbish". It is that a certification is a door, not a destination. Use one to get read. Use the work to get hired. What I object to is not people taking them, it is an industry that has quietly changed what the word certified means while keeping the price.
What I would tell someone starting now
Take one proctored exam in the platform your clients actually use. Put the rest of that time into something a stranger can inspect without your help: a repository, a write-up of a migration that went sideways and what you did about it, a small tool that solves a real irritation. Our own wreck-watch exists because a problem annoyed us enough to fix it, and it has started more conversations than any badge ever has.
And if you are choosing between two candidates and one has a stack of certificates while the other has a thing they built and can explain, you already know. You have always known. The certificates never told you what you needed, they just made the decision feel more procedural.
The health and safety test worked because you could fail it, someone watched you take it, and getting it wrong meant something. Three conditions. Most of this market meets none of them, and I say that as somebody holding the badges.
Check the claims
→ The Cert Ledger, every AI certification we can verify, refreshed daily, with the grading framework
→ All seven datasets, free to cite, each stating its own limits
→ Disagree? If you hire for AI roles and see it differently, that is more useful than agreement and we will publish the correction