What did 424 AI-obtained credentials prove?
What my completed experiment reveals about course records, personal ability and the decisions we make from them.
AI agents working through my account collected 424 credentials that could be verified on the providers' public pages. That establishes what the providers issued to the account; it leaves my ability to do the work untested.
From 25 to 28 October 2026, I authorised agents to take free online courses and assessments on my behalf. I tracked the process, including courses that ended without a verified credential. The complete experiment record remains public.
A hiring manager could check the 424 issuances and still have no answer to a fairly ordinary question: which of these subjects can Jacob use competently in a job?
That is the question I want to keep attached to the result. A real credential can be asked to carry a stronger claim than the work behind it supports. When someone uses it to choose an employee, approve a training programme or assign responsibility, that gap has consequences.
What the agents collected
Most of the 424 credentials were learning achievements. The providers' categories break down into 315 learning achievements, 90 course certificates, 15 skill badges and four awards the providers call certifications.
Those labels describe different kinds of evidence. A short module-completion achievement and a course certificate do not become equivalent because both appear in a profile. Nor does the word certification, by itself, establish accreditation, professional licensing or supervision of an assessment.
The count also follows the providers' award structures. It includes 23 aggregate awards, such as learning-path trophies, issued on top of module awards that also count. They are separate credentials under the experiment's counting rule. They are not 23 additional, independent demonstrations of skill.
There were 470 course records in total. The other 46 remained in the register when I stopped the experiment. Some assessments failed. Other courses were blocked by access or technical problems, needed further work, or had not been attempted. Calling all 46 failures would tell the wrong story.
The recording period was just over three days. Agents could work in parallel, and the timestamps include coordination and verification. Dividing an advertised course duration by that period would not produce an honest measure of learning speed or time saved. My own setup and support time was never measured.
What a verified credential tells you
Verification established that the provider had issued the credential to my account. It did not independently establish who supplied the assessment answers or what I had learned.
That distinction is easy to lose when a check returns a reassuring result. The holder matches. The title matches. The provider confirms the issuance. Those are useful checks, especially when the alternative is accepting a fabricated certificate. They answer the authenticity question.
The question about capability needs more. What work was required? Under what conditions? Which parts were assessed, and whose contribution was observed? Some credentials include criteria or supporting evidence that help answer those questions. They deserve to be read, rather than treating the verification link as the whole assessment.
In this experiment, I supplied authorisation and helped with account setup, sign-ins and acceptance of terms. The register records the assessment answering as AI-performed. That is the experiment's own record of the work, not an independent observation proving that no human was involved at any point.
My permission to use my account did not settle the providers' assessment rules. The Saylor records explicitly disclose that AI-performed finals departed from the provider's expectation of individual work. An issuance should not be read as the provider endorsing how the assessment was completed.
I cannot take the issuances and turn them into a claim that I personally mastered the subjects. Equally, someone reading a credential without the experiment's disclosure would not necessarily know how the work had been divided. The account name alone cannot settle that.
What the unsuccessful courses add
The courses without a verified award help define what this experiment tested. They stop the headline from becoming a claim that agents can complete any assessment placed in front of them.
Four records ended in failed exams or units without a later successful retake before closure. Twenty were recorded as access blocks. Those blocks included content that did not render, unavailable practical environments and four finals that required webcam, face registration or screen sharing.
Those four finals were never started. The experiment excluded that kind of camera and proctoring access, so reaching the requirement triggered its own stop rule. It would be misleading to describe the result as proof that proctoring defeated an agent, or that an agent could bypass it. Neither was tested.
The course selection matters too. This was a deliberately chosen set of free online offerings, not a representative sample of education or professional examinations. No fee was recorded for the 424 issued credentials. Compute, subscriptions and human time were not costed, so the experiment has no total-cost result.
These limits leave a substantial observation intact: agents completed work through an authorised account, and providers issued credentials that another reader could verify. There is enough there to question how those credentials are used without making a claim about every course, candidate or qualification.
Which decision are you using the credential to make?
The evidence you need depends on the responsibility you are about to give someone. A course record may help you choose what to discuss with a candidate, while being insufficient to decide that they can handle a customer problem independently.
Suppose a candidate has a relevant course certificate. I would start with the work the job requires and ask them to explain how they would approach it with the tools they would actually have. Where would AI help? What information would they need before accepting its answer? What would make them stop and ask someone else?
Their explanation gives the credential some context. A short piece of work can give more: a recommendation, the assumptions behind it and the checks they made. Changing a condition then helps reveal whether the person can work through the consequences or only present the original output. This is a proposed way to assess the contribution, not a hiring method validated by my experiment.
The same question applies inside a company. If a completion report is used to decide that a team is ready for a new responsibility, ask what observation supports that step. The report may accurately record completed activities while saying little about the decisions employees can now make at work. The experiment involved no corporate learning system or employees, so it does not establish how often that happens in a business.
I had already raised the question of who actually passed an AI training assessment. The completed experiment gives it a larger, inspectable basis. It still does not let me infer a person's ability from the number of entries beside their name.
What remains worth learning
A course can help someone learn even when its completion record gives limited evidence about that learning. The agents obtaining a credential does not tell us what a person would gain from studying the material.
It also leaves open a useful kind of competence: working well with AI. I delegate a lot of work to agents myself. If a job allows those tools, an assessment that excludes them may miss part of how the person would perform. But access to an agent does not answer whether the person can recognise a bad recommendation, check a claim or decide when a task exceeds their authority.
The next step is to define the work for which the person will be responsible and observe enough of their contribution to make that decision. Where the stakes are low, a discussion may be sufficient. Where mistakes would affect customers or other people, a completion count needs more support.
The 424 credentials remain real issuances. When I decide what responsibility to give someone, I still need to understand what they can do with the knowledge and tools in front of them.