# Our school wants to buy a deepfake detector for staff. What should we look for? Source: https://global100.org/forum/q/our-school-wants-to-buy-a-deepfake-detector-for-staff-what-should-we/ Site: Global 100 Forum, category Deepfakes and video Published: 2026-10-25. Updated: 2026-10-26. Replies: 9. ## Question by Dan Okafor, 2026-10-25 I teach English at a secondary school, and I have been asked to sit on a small group that will recommend whether the school buys a deepfake detector for staff. The trigger was an incident at another school in our area involving a fabricated image of a student. Our leadership team wants to be able to say we have a tool in place. I am not a technical person. The vendor demos I have seen show a page where you upload a file and get a percentage or a green and red result. They promise it works on images, video and audio. I do not know how to judge any of that. What questions should a non-specialist group ask before recommending a purchase? What would a good evaluation look like for a school, where the realistic cases are a faked image circulating among students, an impersonation of a teacher, or a scam voicemail to the office? And is a detector even the right thing to spend on first? I would rather go back to leadership with a short list of questions and some alternatives than with a yes or no I cannot defend. ## Reply 1 by Sam Whitlock, 2026-10-25 Short version, Dan: ask what happens when it is wrong. Every detector will be wrong sometimes. A vendor who answers "ours is 99% accurate" without saying on what data, generated by what tools, compressed how, is selling you confidence. Ask for their test conditions in writing. Then run your own test. Take twenty files you know the origin of, including real ones that have been resaved through a messaging app, and see what it says. ## Reply 2 by Ingrid V., 2026-10-25 From the platform side: the question that usually reorders the list is "what will staff do with the output?" If the answer is "decide whether the image is fake", the tool is being asked to do something it cannot. If the answer is "decide whether to escalate and preserve evidence", it can be a useful input. Also ask where the uploaded files go. If staff upload a photo of a pupil to a third party site, you have created a data protection question that did not exist yesterday. ## Reply 3 by Lena B., 2026-10-25 Spend on training first. Staff who know what to do in the first hour, who to tell, what not to forward, how to keep the evidence, will do more good than a green tick that says something is real when it is not. And please do not buy something because another school had a scare. That is how you end up with a licence nobody opens. ## Reply 4 by Sam Whitlock, 2026-10-25 @Lena B. agreed on training. I would only add that the training should include a rehearsal, one hour with a made up incident. Plans that have never been walked through fall apart the first time someone has to use them. ## Reply 5 by Marco Ferri, 2026-10-26 Newsroom view. Detection tools are one step in verification, never the whole process. We look at the source, the first appearance online, other copies, metadata, and we call people. A detector score is an input to that, and we log it with the others. For a school: define who verifies, how, and what the threshold for escalating to safeguarding or the police is. A vendor can supply a tool but cannot supply that process. ## Reply 6 by Hana Sato, 2026-10-26 Two things I would add. First, make sure you know which media the tool covers. Many are built for faces in video, some for images, some for audio, and they are not interchangeable. A tool good at face swap video may be useless on a fully generated still. Second, ask whether it was evaluated on images that look like yours, meaning phone photos, screenshots and compressed files that have been shared several times. Performance on pristine lab files tells you very little about those. ## Reply 7 by Kwame A., 2026-10-26 Legal angle, since the incident prompting this involved a student image: a detector will not tell you whether an offence has been committed, and in many places the obligations on a school are about reporting and safeguarding processes, not about proving the image is synthetic. Check what your local guidance says you must do, regardless of whether anyone can prove it is fake. That belongs in the recommendation too. ## Reply 8 by Tomas Reyes (staff), 2026-10-26 Start with the process, not the tool. Ask what the detector detects, on which media and which generators, how it performs on compressed images that have been shared, and what staff will do with its output. Pair any purchase with a response plan, training and a verification routine. The government evidence supports that order. The NSA, FBI and CISA joint sheet [Contextualizing Deepfake Threats to Organizations](https://media.defense.gov/2023/Sep/12/2003298925/-1/-1/0/CSI-DEEPFAKE-THREATS.PDF) describes detection as a cat and mouse game and recommends combining technology with procedure: real-time identity verification for communications (for example a one-time PIN or known personal details), passive detection when judging media that already exists, planning and rehearsing responses, training personnel, and reporting incidents. For passive checks it lists basic steps that cost nothing: keep a copy and hash the original, check the source, run reverse image searches, look at metadata, and examine the media for physical impossibilities. NIST's [AI 100-4 report on synthetic content](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf) is the source for the caution on accuracy. It says detector performance depends on the tool and the dataset, that detectors perform better on images from the generator they were trained on, that cross-generator error rates are substantial after compression and resizing, and that video detector accuracy across studies ranged widely, from 62% to 99% in one 2023 survey. Those are different tests under different conditions, so read them as "it depends", not as a score for any product. For your group: request test conditions in writing, trial on your own files, check where uploaded images are stored, define who verifies and who escalates, and budget for training. Dan's concern about a defensible recommendation is best served by a short document saying what the tool will and will not do. The [thread on whether deepfakes are illegal](/forum/q/is-it-illegal-to-make-or-share-a-deepfake-of-someone/) may help with the safeguarding side. ## Reply 9 by Dan Okafor, 2026-10-26 @Tomas Reyes this is exactly what I can take to leadership: process first, a trial on our own files, a question about where uploads go, and a one page statement of what the tool cannot do. Thank you, and thanks to everyone who answered. --- Cite as: Global 100 Forum, "Our school wants to buy a deepfake detector for staff. What should we look for?", https://global100.org/forum/q/our-school-wants-to-buy-a-deepfake-detector-for-staff-what-should-we/, accessed 2026-11-10.