AI changes what counts as evidence. And how to test it.
How we use artificial intelligence ourselves, which new sources of evidence AI chats and AI agents create, and how we forensically expose manipulated photos, videos and calls.
How we use AI ourselves
Where it helps: transcribing audio and video files, translating foreign-language documents and prioritising large data sets in review. AI pre-sorts; it does not decide.
Who decides: Every finding that goes into a report has been checked by a person. AI-assisted steps are documented and disclosed in the report.
Where your data stays: Client data does not leave our platform or Switzerland for AI functions. It is not entered into public AI services and is not used to train models.
Regulation: We align our use with the Swiss Federal Act on Data Protection and the EU AI Act, and disclose on request which models are used for what.
New sources of evidence: AI chats and AI agents
AI chats: Prompts and answers in Copilot, ChatGPT or Gemini show what someone wanted to know, what they drafted and what they passed on. In Microsoft 365, Copilot interactions are stored and can be preserved with eDiscovery tools, provided someone thinks of it in time.
AI agents: When agents send emails, change files or trigger payments, every investigation has to ask afresh whether a person did it or an agent acting in their name. The answer lies in logs, permissions and prompt histories, not in the document itself.
What this means for your case: Legal holds and preservation orders must explicitly cover AI interactions and agent logs. In review we treat them as sources in their own right, with their own chain of custody, and label AI-generated content so that authorship is not misattributed.
Real or fake: photos, videos, calls
The situation: A photo as proof of damage, a video of an alleged statement, a call in the voice of senior management ordering an urgent payment. Generative AI can fake all of this, and the other side knows it.
Check the origin, not just the image: Whether a photo is genuine rarely shows in the image alone. We preserve the device it supposedly comes from and check the original file, metadata, app data and timelines for contradictions.
Camera ballistics: Every camera sensor leaves an individual noise pattern in its images. This makes it possible to test whether a photo was really taken with a particular camera or smartphone, independently of metadata, which is easy to change.
Fake calls and voice cloning: For alleged calls, we match call logs, messenger data and device traces against the claimed sequence of events. Whether a conversation took place, and from which device, can often be answered more conclusively than how the voice sounded.
What we do not promise: No method detects every forgery, and a missing match does not prove authenticity. Our reports state what was examined, with which method and with what evidential weight.
Together with Compelson
Prague-based Compelson develops MOBILedit, one of the established platforms for forensic acquisition and analysis of smartphones, smartwatches and cloud data, as well as Camera Ballistics. We use both tools in our mandates.
MOBILedit Forensic: acquisition and analysis of mobile devices, including chat and app data from more than a thousand apps.
Camera Ballistics: attributing a photo to the camera that took it, based on the sensor fingerprint, with a report on match and evidential weight.
Is it real? Ask before you rely on it.
When a photo, a video or a call drives a decision, it pays to check it first.