A model can compare evidence, not detect betrayal
Machine learning can help compare available dating-profile images with a reference photo. It cannot determine whether someone has broken a relationship agreement. Those are different questions: one concerns resemblance between images, while the other requires reliable facts, timing, account ownership, and the boundaries the people involved actually agreed to.
A tool described as an “infidelity detector” should therefore explain the narrower task it performs. Ask what data it examines and what its output represents. A candidate profile is material to review, not a diagnosis of someone’s intentions. A number attached to a candidate does not make the broader conclusion more reliable if the underlying information is missing.
This guide is a checklist for evaluating model claims. For an explanation of the labels shown in an OopsBusted report, see how to read a facial match.
Separate the stages of the search
Start by distinguishing input, coverage, comparison, and interpretation. The input describes whom you are looking for. Coverage is the material the service has available. Comparison measures a relationship between that input and candidate evidence. Interpretation asks what the returned material actually supports. Each stage has its own limitations.
Consider a clear reference photo and a missing profile. Better image quality cannot create evidence that the service does not hold. Conversely, a profile may be available but contain a photo that does not compare well with the reference. A completed search with no returned match leaves these possibilities open unless the report provides enough detail to distinguish them.
Amazon’s face-comparison documentation describes the method as probabilistic, acknowledges false negatives, and recommends human review before consequential decisions. That describes provider behavior; it is not an independent accuracy study of OopsBusted.
Ask what an accuracy claim measures
“Highly accurate AI” is not a useful evaluation on its own. Look for the task, the test population, the size of the evaluation, the definition of a correct answer, and the rate of errors in both directions. A model may correctly compare two faces yet provide no evidence about who uploaded the profile photograph.
A benchmark on curated images does not establish performance on every dating-app image, location, or age group. Before relying on a published figure, ask whether it measures the same conditions as your intended use. If the vendor cannot provide that information, treat the number as unverified for your situation rather than filling the gap with an optimistic assumption.
Also distinguish comparison scores from calibrated probabilities. A raw similarity percentage should not be read as a percentage chance of cheating, account ownership, or recent activity. No conversion between those claims is established by displaying a score.
Understand the inputs OopsBusted supports
Standard OopsBusted intake uses a name, adult age range, gender, and location. A reference photo is optional for a standard search and required for AI photo matching. With AI matching enabled and a reference photo supplied, the photo-only path can omit a name; the remaining required details still apply.
A subject’s phone number or email address is not a lookup input. App selections are editable clues. The initial search checks available Tinder, Bumble, and Hinge data, while a later paid unlock determines access to app reports. Paying for a wider set of reports does not expand the underlying evidence into complete platform coverage.
Review the search workflow before uploading anything. These requirements describe the current product, not a recommendation to obtain missing information through private accounts or someone else’s device.
Read outcomes without changing their meaning
“Possible face match” and “Strong face match” are comparison bands for reviewing candidates. Neither establishes ownership or behavior. “Face compared: no match” describes a performed comparison. It must remain separate from a comparison that was not requested, failed, or could not use available evidence.
Use a short review record for each candidate:
- Identify the app and the specific evidence shown, keeping different profiles separate.
- Record the capture date if supplied. Do not substitute the report-opening date for it.
- Note the comparison label and whether the comparison completed.
- List conflicts, missing context, and coverage limitations beside the supporting details.
For example, resemblance with a conflicting location does not justify silently changing the location in your notes. The conflict is part of the result. Repeating the same comparison cannot independently corroborate it, because the evidence has not changed.
Keep generated explanations separate from facts
A fluent explanation can make limited evidence sound conclusive. If a tool produces a narrative, trace each factual sentence back to a displayed source. A statement about when a person last used an app needs evidence of that event. A capture timestamp alone records collection, not the person’s activity at the moment you read the report.
Do not use a chatbot’s interpretation of someone’s tone, response time, or photograph as a substitute for that missing source. Such an answer may sound plausible while depending on assumptions you never supplied. Keep the original evidence and the model’s commentary distinct so that a later reader can assess the same limitations.
The ethical AI decision guide explains how to assess an intended use before acting on model output, including the option to stop without a conclusion.
Decide whether another search would add information
Before repeating a search, name the information you expect to change: a corrected required input, a different usable reference image, or newly available coverage. If nothing relevant has changed, another purchase is not independent evidence. Uncertainty may remain unresolved even after a technically successful comparison.
Review sample reports to understand the presentation, and review AI data-handling questions before sharing additional material. A useful stopping point is a clear account of what the available evidence supports and what remains unknown. The model cannot decide the relationship question for you.