Start with labelled examples and check their quality

At its simplest, supervised training begins with labeled examples: authentic material in one pile and synthetic or manipulated material in another. The model adjusts itself until features in those examples help separate the labels. For video it might learn spatial details inside frames, timing across frames or the relationship between lips and sound. For audio it may learn spectral and timing patterns. Text detectors work with a different set of statistical features again.

The difficult part is not making the training score rise. It is making sure the model learned forgery evidence rather than an accidental clue in the dataset. If all fake images were saved as PNG and all real images as JPEG, a wonderfully accurate model might simply become a file-format detector.

Models can learn the wrong clues

A training set contains thousands of relationships its creators never intended. The synthetic faces may involve a different group of identities, backgrounds or image sizes. One generator may leave a repeated frequency pattern. A dataset may come from a single website. The model will use any reliable shortcut it can find, even when that shortcut disappears outside the laboratory.

This is why researchers use matched real and fake examples, multiple datasets and careful hold-out tests. Work on deepfake generalization has found biases tied to identity, frequency bands and particular manipulation methods. Better training tries to stop the model from memorising those conveniences and push it toward signals that survive new subjects and tools.

The exam must contain things the model has not seen

If training and testing use near-duplicates, the final accuracy number can look magnificent and tell us very little. A serious evaluation separates the data by source, generator or time. Cross-dataset testing goes further: train on one collection and test on another with different capture conditions and manipulation methods.

Unseen generators are the hard case. Generative systems change quickly, while labeled evaluation data arrives slowly. NIST's GenAI program treats generators and discriminators as two sides of an ongoing evaluation rather than a problem solved by one frozen benchmark. That is a healthier mental model for a live product too.

Training for compression, cropping and other edits

During training, teams may resize, crop, blur or compress examples. They may vary color, noise, audio quality and other properties that should not decide the label. The goal is not to damage files for sport. It is to teach the detector that an authentic clip remains authentic after ordinary processing and that forgery evidence should survive more than one clean export.

Diversity helps, but it is not magic. Combining several datasets can introduce new conflicts; an augmentation can teach the wrong invariance; vast quantities of similar examples can still leave a blind spot. Training data needs coverage, documentation and analysis, not just volume.

Retraining can fix one gap and open another

Suppose a detector misses a new video generator. Adding those failures to the next training run sounds obvious. But a model that adapts aggressively can forget older manipulation types or shift its threshold so that more genuine media is flagged. Continual-learning research exists because retaining old knowledge while learning new attacks is a real engineering problem.

A responsible update is tested against a fixed regression set as well as new cases. Teams should measure false positives, missed detections, calibration and performance under realistic degradation. They should also keep the previous version available long enough to compare behavior. 'Newer' is not a metric.

What training means for the person reading a result

The detector has not learned a universal visual definition of truth. It has learned a decision boundary from particular examples and objectives. Ask whether the uploaded file resembles the conditions in which the system was evaluated. A short translated paragraph, an old CCTV clip and a pristine generated portrait are three very different tests.

Good products keep the model answer narrow, add provenance and source checks, and allow uncertainty. Good users do the same. The training process can produce a valuable forensic signal; it cannot give the software access to the event, the author's drafts or the uploader's intentions.

Automated results require source and context review.

Continue with independent verification.

Read our methodology
Sources

Primary reading

We use original standards, regulators, public institutions and research papers wherever possible. Sources were last checked on 14 August 2026.