False Positives and False Negatives in AI Underwriting Flags: How to Log and Review Them
Updated November 5, 2026: corrected for accuracy.
A false positive is an AI underwriting flag that marks a sound file as risky. A false negative is a risky file the model lets pass. Lenders should send scores near the approval cutoff to manual review and record every flag outcome in a simple log, so repeated errors become visible and fixable.
Key Takeaways
- A false positive flags a sound file. A false negative passes a file that has a real problem.
- False positives cost time and borrower goodwill. False negatives cost credit quality, and they are harder to see.
- Scores just above or just below the approval cutoff deserve a human review, because the model is least reliable there.
- A flag-review log should record the flag, the reviewer’s finding, the error type and the root cause.
- Track the two error types separately so one does not hide the other.
Related Topics
- 5 Red Flags in AI in Underwriting: Opportunities and Limits
- Key Terms in AI in Underwriting: Opportunities and Limits
- Manual vs. Automated: AI in Underwriting for Private Mortgage Lenders
- How to Get Started With AI in Underwriting
- Late Fee and Grace Period Management in Private Mortgage Lending
- Seven Servicing Habits That Earn Repeat Borrower Business
What Is a False Positive in an AI Underwriting Flag?
A false positive happens when the model raises a risk flag on a file that a careful human reviewer would clear. The flag looks serious, but the facts behind it do not support it.
Payment timing is a common source in private lending. A model can treat any payment received after the due date as a late payment. That reading ignores the grace period written into the note. A payment that arrives inside the grace period often does not trigger a late fee under the note’s own terms. The model still scores it as a delinquency signal. How a grace period works depends on the note terms and on state law. The guide to late fee and grace period management covers how lenders set and apply those terms.
A second source is a repeat borrower whose record the model cannot weigh. The model sees one pattern in isolation. It does not see a long history of closed and repaid loans with the same lender. A human sees that context at once. The model does not, unless someone feeds it in.
False positives have real costs. They slow decisions. They pull reviewers away from files that do need attention. They can also strain a relationship with a strong borrower. The post on servicing habits that earn repeat borrower business explains why those relationships matter.
What Is a False Negative, and Why Is It Harder to Catch?
A false negative is the reverse error. The model passes a file that has a real problem. No flag appears, so no one looks closer.
False negatives tend to come from gaps in what the model can see. Missing documents, thin data on unusual collateral and borrower structures the model was not built for are typical causes. The output still looks confident. That confidence is the danger.
A false positive announces itself, because a reviewer has to clear it. A false negative stays quiet. Sometimes a human reviewer catches it during a full file review. Sometimes it surfaces only after funding. Either way, record it in the log the moment someone finds it. The overview of underwriting red flags every lender should know lists the issues a model should be catching.
Why Do Scores Near the Approval Cutoff Need Manual Review?
A model turns many inputs into one score. A cutoff then splits that score into pass and flag. Far from the cutoff, small data errors rarely change the outcome. Near the cutoff, one misread field can flip a file from approve to decline.
That makes the margin the least reliable zone. Set a band on both sides of the cutoff. Send every file in that band to a human reviewer, whichever way the model leaned. A file just above the line can hide a false negative. A file just below it can be a false positive.
The reviewer should read the model’s stated reason for the score, then check it against the full file. The comparison of manual and automated underwriting explains where each approach fits.
What Should a Flag-Review Log Record?
A flag-review log does not need special software. A shared spreadsheet works if every entry follows the same format. Record these fields for each reviewed file:
- Internal file reference. Use an ID, not borrower names or personal details.
- Review date and reviewer. Note who made the call and when.
- Model version and cutoff in use. Results change when the model or cutoff changes.
- Model output. Record the score, the decision and the reason the model gave.
- Reviewer finding. State whether the reviewer agreed, and why.
- Error type. Mark it as a correct flag, a false positive, a false negative or unresolved.
- Root cause. Pick from a short, fixed list, such as grace period misread, missing history, missing document or data feed error.
- Action taken. Note the final decision and any fix requested.
A fixed list of root causes matters most. Free-text notes are hard to sort. Fixed categories let patterns show up.
How Should a Lender Review the Log?
Review the log on a regular schedule, not only when a problem appears. Sort the entries by error type, then by root cause. Look at false positives and false negatives as separate groups. A drop in one can hide a rise in the other.
Recurring causes point to fixes. If grace period misreads keep appearing, the payment data may lack a clear grace period field. If repeat borrower flags keep appearing, the model may not be receiving relationship history. Clean servicing records help here. A payment history that shows the due date, the received date and grace period status gives a model less room to misread timing.
Send each fix to whoever configures the model. Then log the change, so later entries can show whether it worked. The post on red flags in AI underwriting describes warning signs worth adding to the review. Depending on loan type, state and loan purpose, a decline can carry notice or record-keeping duties. Consult qualified legal counsel about what your log and your decline files need to contain.
Frequently Asked Questions
What is the difference between a false positive and a false negative?
A false positive flags a file that turns out to be sound. A false negative passes a file that turns out to have a real problem. The first costs time and goodwill. The second costs credit quality.
Why would a model read an on-time payment as late?
Some models compare the payment date only to the due date. They do not account for a grace period in the note. Grace period terms depend on the note and on state law, so the data has to show them clearly.
How wide should the manual review band around the cutoff be?
No single width fits every lender. Start with a band you can staff. Then use the log to see whether errors cluster outside it, and adjust.
Should the log contain borrower personal information?
Keep it to an internal file reference where you can. The full file holds the details. Privacy and retention duties vary by loan type, state and loan purpose, so consult qualified legal counsel.
Who should review flagged files?
Someone with private lending underwriting experience who can read the full file. That person should also have the authority to override the model and record why.
Does a review log replace human underwriting?
No. The log supports human review. It shows where the model needs help and where reviewers should focus.
Sources
- Underwriting, decline-notice, privacy and record-keeping requirements vary by loan type, state and loan purpose (business or consumer). Consult qualified legal counsel about your own obligations.
Next Steps
If you want to talk about how clean servicing records, such as clear payment dates and grace period status, support better underwriting review, we can help. Contact Note Servicing Center to start a conversation about your servicing needs.
This content is provided for general informational purposes only and does not constitute legal, financial, or compliance advice. Always consult a qualified attorney or advisor regarding your specific situation.
Share This Story, Choose Your Platform!
Disclaimer
The information provided in this article is for general educational and informational purposes only and does not constitute legal, financial, investment, tax, or professional advice. Note Servicing Center, Inc. is a licensed loan servicer and does not provide legal counsel, investment recommendations, or financial planning services. Reading this content does not create an attorney-client, fiduciary, or advisory relationship of any kind. Nothing in this article constitutes an offer to sell, a solicitation of an offer to buy, or a recommendation regarding any security, promissory note, mortgage note, fractional interest, or other investment product. Any references to notes, yields, returns, or investment structures are illustrative and educational only. Past performance is not indicative of future results, and all investments involve risk, including the potential loss of principal. Note investing, real estate transactions, and lending activities are subject to federal, state, and local laws that vary by jurisdiction and change over time. Before making any decision based on the information in this article, you should consult with a qualified attorney, licensed financial advisor, certified public accountant, or other appropriate professional who can evaluate your specific circumstances. Some articles on this site include hypothetical stories, examples, and scenarios created to illustrate concepts and demonstrate the types of situations Note Servicing Center, Inc. handles. Any names, companies, properties, and circumstances in these examples are fictitious or have been anonymized to protect confidentiality, and any resemblance to actual persons or entities is coincidental. These examples do not describe specific clients and do not guarantee any particular outcome. Some content may be created with the assistance of generative AI tools and may contain errors or omissions. While we make reasonable efforts to ensure the accuracy of the information presented, Note Servicing Center, Inc. makes no warranties or representations regarding the completeness, accuracy, or current applicability of any content. We disclaim all liability for actions taken or not taken in reliance on this article.
