Adobe’s work on manipulated media includes both AI research into detecting edits and tools that record a file’s provenance. These approaches answer different questions: whether an image shows signs of a particular manipulation, and what information is available about its creation or editing history.
What are deep fakes?
Deep fakes, usually written “deepfakes”, are synthetic or manipulated media that can make someone appear to say or do something they did not. AI can generate a brand new scripted video or voice recording, or alter an existing one. Results vary, and visual realism alone is not evidence that an event happened.
Non-consensual sexual imagery was an early, harmful use of face-swapping technology. Impersonation, fraudulent messages and misleading public statements are other concerns. A clip shared on Facebook or another platform should be checked in its original context rather than trusted because many people have reposted it.
How are deep fakes made?
AI systems learn patterns from images, video or audio and use them to generate or modify media. A face-swap system and a voice-cloning system do not necessarily need the same inputs. The amount of material needed also varies, so there is no universal requirement for thousands of recordings of one person.
Not every misleading image uses AI. A genuine photograph can be given a false caption, an old video can be presented as recent, and an ordinary edit can remove crucial context. Checking only for AI manipulation misses these simpler forms of deception.
How is Adobe helping combat deep fakes?
In research published in 2019, Adobe and collaborators investigated facial edits made with Photoshop’s Face-Aware Liquify tool. Their model looked for that type of manipulation and explored how it could be reversed. Adobe Research’s subsequent image-forensics work also examined ways to identify editing traces.
These research results should not be read as a claim that Adobe can detect every deepfake video or remove false media from the internet. A detector’s usefulness depends on the manipulation it was designed for and the material it receives. Compression, unfamiliar editing methods and poor-quality copies can complicate interpretation.
What Content Credentials add
Adobe also supports Content Credentials, based on the C2PA standard, to make information about media creation and editing available for inspection. These records can include the tools or processes used and are digitally signed so a viewer can check their integrity.
A file hash helps connect a record to particular content. It is not, by itself, an AI detector. A changed hash tells you that the compared data differ; it does not establish whether a change was deceptive, a routine export or another operation.
Do not confuse provenance with truth
Provenance information can describe a file’s source or edits, but it cannot prove that the depicted claim is true. The C2PA explainer distinguishes verification of provenance records from judgments about truth. Missing metadata does not by itself prove that media are fake, either.
A real recording can still show a staged scene. Conversely, a legitimate photograph may reach you without its original metadata. Treat credentials as useful evidence alongside the source and context, not a replacement for editorial judgment.
How to check a suspicious clip
- Find the earliest available upload and identify who published it. A repost may omit the original caption or date.
- Look for the complete recording rather than a short extract. Check whether surrounding footage changes its meaning.
- Inspect available Content Credentials and read what they actually state, including any reported edits.
- Compare the claim with independent reporting or a direct statement from the person or organisation involved.
- If a clip asks you to transfer money or disclose information, verify the request through a separate, known contact channel.
Visible glitches can prompt further checking, but their absence is not proof of authenticity. If the source cannot be established, avoid presenting the clip as verified.
