Glossary of forensic models
Uthentic never leaves the verdict to a single model. Every analysis queries independent engines — on images and, separately, on texts — and fuses their readings after calibration. Here is what each one measures and how it should be read, because a number without its photographic context is not proof.
Engines for images
They read the photographic file: sensor signature, optical coherence, detail texture and the local integrity of single areas.
- Local file integrity
- Local integrity model: it looks for regions of the file whose noise and compression signature does not match the rest of the image.
- How to read itIt separates a fully generated picture from a real photograph with one area rebuilt in software. Its monochrome map shows where to look, not that the image is fake.
- Cross-generator generalisation
- Universal synthetic image detector built on general visual representations rather than on a single generator.
- How to read itIt holds up well on generators it never saw in training; on heavily compressed or resized photographs it loses sensitivity.
- Sensor and optical traces
- Family of academic detectors measuring sensor traces and colour interpolation inconsistencies left by the capture chain.
- How to read itA high signal means the camera-lens-sensor chain cannot be reconstructed: this may come from AI generation or from an aggressive re-export.
- GuastoniUnmask
- Proprietary on-device engine reading the microtexture of the epidermis and the coherence of pigmentation backlighting against the reflected light of the environment.
- How to read itIt compares skin grain with the direction of ambient light: real skin shows irregular microtexture and a coherent pigmentation-to-light ratio, while synthetic skin tends towards a uniform visual paste and a backlight that does not match the scene.
- Open-source detector (local processing)
- Detector trained on a wide set of public generators, executed locally on the device through ONNX Runtime.
- How to read itIt gives an independent reading without sending the file to external services; strong on widespread generators, weaker on very recent models.
- High-fidelity diffusion
- Family of high-fidelity text-to-image diffusion models, among the most frequent generators in the images analysed.
- How to read itIt leaves a very regular detail texture and a uniform microcontrast: photographically it reads as light that is too clean, without accidents.
- Diffusion and GAN patterns
- Statistical signatures in the frequency domain, typical of diffusion models and of generative adversarial networks respectively.
- How to read itThey are clues about the generative family, not proof: a heavily filtered real photograph can show similar periodic patterns.
Provenance and metadata
They check what the file declares about its own origin, and whether that declaration is signed and coherent.
- C2PA
- Signed provenance standard: it records inside the file who produced it and which edits it went through, with a verifiable cryptographic signature.
- How to read itA valid signature is the strongest available proof of origin. Its absence is normal in most files and does not indicate forgery.
- EXIF and capture metadata
- Technical data written by the camera or by software: body, lens, shutter speed, aperture, ISO, colour space, editing chain.
- How to read itCoherence between the declared parameters and the visible rendering of light is a strong clue; social networks and messaging apps strip this data.
Engines for texts
They establish whether a text was written by a language model and explain which sentences carry the evidence.
- Widely used open language models
- Family of open language models, widely used to generate or rewrite text; one of the references in detection benchmarks.
- How to read itThe texts produced show sentences of regular length and predictable transitions: exactly the kind of regularity burstiness statistics measure.
- Recent multilingual language models
- Family of high-performance language models, often behind recently generated Italian and English texts.
- How to read itIt produces very orderly arguments and average vocabulary: this raises the risk of a false negative on technical texts, where humans also write regularly.
- Anti-paraphrase detector
- Academic detector of generated text, trained adversarially against rewrites and paraphrases.
- How to read itIt is the most stable signal when a text has been paraphrased to mask its origin.
- Multi-generator detector
- Reference detector for generalisation: evaluated across dozens of generators and domains.
- How to read itUseful as a counterweight when a single engine drifts; its reading matters most when it agrees with the statistics.
Report metrics
The numbers that appear in the report: how they are computed and, above all, how they should not be read.
- Perplexity
- It measures how predictable a text is for a reference language model: low values mean highly expected lexical choices.
- How to read itOn its own it decides nothing: human legal, academic and regulatory texts naturally have low perplexity.
- Burstiness
- Variability of predictability between consecutive sentences: human writing alternates peaks and drops, generated writing tends to be uniform.
- How to read itIt must be read together with text length: below the minimum word count the value is not reliable and the analysis declares insufficient evidence.
- Conditional curvature
- Gap between the observed likelihood of the text and the likelihood the reference model expects across the most probable alternatives.
- How to read itIt is robust on texts sampled at high temperature, where perplexity alone comes close to human values.
- AI Probability
- The final percentage of the report, obtained from the robust fusion of independent engines after calibration.
- How to read itFrom 0 to 19% the verdict is green (real); from 20% up it is red (likely AI). The report always states how many engines answered.
10,179 runs on the fusion of engines
The fusion is never changed by feel. Every change to the weights goes through a repeatable bench of 10,179 runs, split into three categories: real photographs, real photographs with software edits (retouching, cropping, local inpainting) and AI generated images. On each category we measure the area under the ROC curve, the Brier score, the expected calibration error and risk-coverage curves.
The category of real photographs with software edits is the most important one for the credibility of the report: it makes sure an authentic photograph worked in post production is not declared AI generated. When signals disagree, or engine coverage is insufficient, the report declares insufficient evidence instead of forcing a percentage.