Endorsed uses a machine learning model to create a prediction score for every candidate’s fraud risk. The system looks at hundreds of data points about a candidate to produce this single risk level.

The four risk levels are:

This is the most important thing to understand: the risk score is produced by the AI model independently. The signals you see in the dashboard are an explanation layer. They show you summarized pieces of what the model is considering, but they don’t show you everything.

An analogy: Imagine a doctor gives you a health score. They then show you a few of the test results that went into that score: your blood pressure, your cholesterol, your weight. Those test results help you understand the score, but the doctor also considered dozens of other lab results, your family history, and patterns across thousands of other patients. The test results you see illustrate the score. They didn’t create it.

You might see two candidates with a similar number of green and yellow dots, but one is “Clear” and the other is “High Risk.” This can feel confusing. Here’s why it happens:

The dashboard shows you signals across several categories:

You may see yellow or red resume signals like “Resume: Word Choice” or “Resume: Writing Style.” These mean that Endorsed’s AI has detected statistical patterns in the resume text that are correlated with fraudulent applications.

We intentionally don’t go into detail about what these patterns are. There are two reasons for this. First, if we explained exactly which word patterns or formatting details we look for, fraudsters could simply adjust their resumes to avoid detection.

Second, some of these signals are based on statistical correlations that are genuinely useful for detection but would raise more questions than they answer if shown in detail. They’re the kind of thing that sounds strange in isolation but our data has shown they are highly predictive when the AI evaluates it alongside everything else.

The important things to know: firstly, we are not penalizing candidates for using AI to write their resumes. Secondly, these signals have been validated against thousands of known fraud cases. If a resume signal fires, it means the text matches patterns that fraudsters disproportionately exhibit.

Endorsed compares the education and work experience on a candidate’s resume to what’s on their LinkedIn profile. When no matches are found between the two, it’s a warning sign. Some fraudulent applicants use completely different histories on their resume vs. their LinkedIn.

Fraud is an adversarial problem. The people committing fraud are actively trying to figure out how detection systems work so they can beat them. Every signal we expose publicly is a signal that fraudsters can learn to game.

Our data science team has identified patterns that fraudsters don’t know we’re looking for, and keeping them confidential is what makes them effective. If we revealed how these signals work, bad actors would reverse-engineer the system and adjust their behavior to avoid detection.

At the top of every candidate’s predicted risk report, you’ll see “Top Signals Relevant to Risk Score.” These are the most important signals for this candidate, the ones that had the biggest influence on the score. Start here.

Below that, you can expand “View Comprehensive Fraud Evaluation” to see every signal we can show you, organized into the categories described above.

Remember: a few yellow signals don’t automatically mean fraud. The model weighs each signal differently. What matters is the overall risk level at the top.

Endorsed’s fraud model is approximately 94% accurate, but no system is perfect. Here’s what you should know about how it handles mistakes:

The system is designed to err on the side of caution for real candidates. For both legal and compliance reasons, our highest priority is making sure real candidates are not wrongly flagged as fraudulent. This means the system is more likely to miss a fraudster (giving them a lower score than they deserve) than it is to wrongly flag a real person as high risk.

In plain terms: if the system predicts someone is shown as a predicted High Risk, you should take that very seriously. If it says someone is Clear, we predict that they are highly likely to be real, but occasionally a sophisticated fraudster might slip through with a lower score.

If you believe the system got it wrong, there are two ways to report it:

Either way, every correction helps. The more examples our team can analyze, the more accurately the system performs for everyone.

“Why is this candidate High Risk when all their signals look green?”

Because the model evaluates signals you can’t see (the confidential signals), and some signals carry dramatically more weight than others. A single heavy signal can outweigh many light ones. We are working on making this more transparent over time.

“Why is this candidate Clear when they have yellow warning signals?”

Because those yellow signals were low-weight. The model considered them but determined they weren’t strong enough, individually or together, to meaningfully increase the fraud risk.

“What are the resume signals checking?”

Statistical patterns in how the resume is written: word choice, writing style, and formatting. These patterns are correlated with fraudulent applications based on thousands of confirmed cases. Details are kept confidential to prevent fraudsters from adjusting their resumes to avoid detection.

“What are the secret signals?”

Additional fraud indicators that our data science team has identified through extensive research. They’re kept confidential to prevent bad actors from learning how to circumvent them. If “Secret Signals: Suspicious activity detected” appears, it means multiple confidential checks have flagged this candidate.

“Should I override the system’s recommendation?”

If your strong gut sense is that the risk score is wrong, trust your gut, but do more work to verify the candidate is real during the interview process. Keep in mind that sophisticated fraudsters can steal real people’s identities, and some go as far as having the real person show up to interviews while a different person does the actual work. Even when overriding, stay vigilant.

“How does Endorsed get better over time?”

We upgrade the system via two methods: 1) adding additional signals, and 2) improved AI training. In any cases where we discover candidates improperly flagged, those cases help us ensure that those candidates and similar candidates to them are accurately scored in the future.

Questions? Reach out to your Endorsed contact or email [email protected]