A prediction error where a model classifies a genuinely positive case as negative, such as a lead scoring model that marks a qualified buyer as low-priority or a brand safety tool that fails to flag harmful content. For agencies, false negatives are the errors of omission: the opportunities missed, the violations that went uncaught, and the costs that accumulate invisibly until the pattern surfaces.
Also known as miss, type II error, false omission
In a binary classification system, every prediction falls into one of four categories. True positives are cases the model correctly identifies as positive. True negatives are cases the model correctly identifies as negative. False positives are cases the model incorrectly predicts as positive. False negatives are cases the model incorrectly predicts as negative when the true label is positive. The false negative rate is the proportion of all actual positives that the model misclassifies as negative: a model with a 15% false negative rate misses 15% of the cases it should be catching.
Recall, also called sensitivity or the true positive rate, measures the complement of the false negative rate: a model with 85% recall has a 15% false negative rate. Improving recall typically requires accepting more false positives, because the threshold adjustment that catches more true positives also catches more negatives that happen to resemble positives. This tradeoff is fundamental and unavoidable in most classification problems: the right operating point depends on the relative costs of missing a positive versus incorrectly flagging a negative.
The cost asymmetry between false negatives and false positives varies dramatically by application. In medical screening, a false negative, failing to detect a disease, is far more costly than a false positive, flagging a healthy patient for follow-up testing. In spam filtering, the relative costs may reverse: a false positive that blocks a legitimate email may be more disruptive than a false negative that lets a spam email through. Agencies deploying AI classifiers need to be explicit about this asymmetry in their system design and communicate it clearly to clients who may not recognize that a model’s threshold is a value choice, not a technical constant.
Most agency AI systems involve classification decisions where one error type is substantially more costly than the other. A working ad agency that evaluates model performance only on aggregate accuracy is ignoring this asymmetry and may be optimizing for the wrong thing. Understanding which error type matters more in each deployment context, and configuring models accordingly, is a practical competency that directly affects client outcomes.
Lead scoring false negatives are the invisible revenue leak. A lead scoring model that classifies genuinely qualified buyers as low-priority means those buyers receive less attention, slower follow-up, or no outreach at all. The cost is a missed sale, which never appears as a model error because the negative outcome, no sale, confirms the model’s prediction. High false negative rates in lead scoring can persist undetected for months because the missed opportunities are invisible by definition. Auditing lead scoring models by following up on a sample of low-scored leads is the only way to measure how many real buyers the model is discarding.
Brand safety false negatives are the headline risk. A brand safety tool that fails to flag genuinely unsafe content places the brand adjacent to content the client explicitly wants to avoid. A single high-profile false negative in brand safety can create more reputational damage than dozens of false positives. Agencies managing brand safety programs need to set their classifier thresholds to reflect this asymmetry rather than accepting the default configuration that optimizes for balanced accuracy.
Content moderation false negatives compound over time. In platforms that depend on AI-assisted moderation, content that escapes detection establishes norms and precedents that influence future submissions. A false negative rate that appears acceptable in isolation creates a compounding problem in production because each piece of content that slips through reinforces the community expectation that this type of content is permitted. Catching it earlier is always cheaper than catching it after the pattern has established itself.
An agency operates a content relevance classifier for a B2B publisher that screens submitted articles for quality and editorial fit. The classifier is tuned for high precision to minimize false positives, articles incorrectly approved that the editors have to reject manually. After six months, the editorial team notices that the volume of high-quality submissions from a specific industry vertical has dropped despite growing audience interest in that vertical. An audit reveals that the classifier has a 31% false negative rate on submissions from that vertical because the training data contained few examples from it. The model has been discarding roughly one-third of genuinely strong articles from a growing segment. The agency adds 200 labeled examples from the vertical to the training set and retrains. The false negative rate drops to 8% and the editorial pipeline for that vertical normalizes within two publishing cycles.
The generative AI foundations module of the workshop covers how to evaluate model performance honestly, including how to audit for the error type that costs most in each specific deployment context.