Deepfakes at work
Why fakes got so cheap, so fast, and the small habits that catch them
The $25 million video call
Before we get to defence, let's calibrate on how good the offence has already got. Pick one of the prompts below and paste it into any current image AI you have handy (ChatGPT with image generation, Gemini, Copilot Designer, whichever your organisation lets you use). Ask for three variations so the pattern is easy to see.
Project the results in the room, then look at them together with the questions underneath.
In early 2024 a finance clerk at the British engineering firm Arup joined what looked like a normal video conference with the CFO and several colleagues at their Hong Kong office. Every colleague on the call was a deepfake. Over the following week she authorised transfers totalling roughly $25 million before the fraud came to light. As things stand it's the largest deepfake-enabled theft on record, and it very obviously won't be the largest for long.
The tools that made this attack possible didn't exist in this form even five years ago. Cloning a voice used to need studio equipment and days of clean source audio. Today it needs a laptop, a small subscription, and about 30 seconds of your voice pulled from a podcast, a work recording, or a voicemail. Convincing video is a couple of years behind on the same curve, and closing quickly.
Over the next hour we'll pull deepfakes apart: what they are, what they already cost, and the small set of verification habits that would have caught nearly every high-profile deepfake attack you'll read about in the next twelve months.
Please generate a photorealistic wire-service photo of the CEO of Amazon walking out of 10 Downing Street with the Prime Minister on a sunny afternoon, press photographers in shot, no visible text.
Please generate a professional LinkedIn-style headshot of a woman in her early forties, navy blazer, warm smile, neutral office background, photorealistic. Then generate two more variations of the same person.
Please generate a candid, slightly blurry smartphone photo of two rival tech CEOs shaking hands at a conference after-party, dim lighting, in the background you can see a bar and a few blurred bystanders, no visible text.
Look at the results together and ask the room:
- Which of these three would you pause on if you were scrolling your feed at speed? Which would you actually forward on?
- What visual details give the fakes away, if any? What would we need to see to be sure they were real?
- If someone in your team clocked one of these as suspect, who would they escalate to, and how quickly?
Turn to whoever is next to you and answer one question: what could a competent fraudster do with two minutes of your voice pulled from any recording you've ever made public? Come back with one worst-case scenario each. This is your baseline exposure, and we'll build the defences on top of it.
What deepfakes actually are, and what's changed
A deepfake is any piece of media (an image, an audio clip, or a video) that has been generated or altered by AI to look convincingly like something, or someone, real. The technique itself isn't new. Doctored photos are as old as photography, and skilled editors have been faking clips for decades. What's new is the collapse in cost and skill.
Ten years ago, faking a passable video of a named executive required a specialist VFX team, weeks of source footage, and a real budget. Today, a convincing 30-second voice clone can be built by any competent hobbyist in an evening. Static image fakes are already effectively free. Live video is a couple of years behind audio on the same curve, and, as the Arup case showed, already good enough to fool a trained finance professional in a Zoom-style call.
That change in cost is the whole story. When faking anyone convincingly costs pennies and takes minutes, the pool of people who can do it goes from 'well-funded specialist units' to 'anyone with a grudge, an angle, or a fraud plan'. This is the shift that turns a technical curiosity into a boardroom risk.
Ask each person to name one publicly available recording of their own voice: a podcast appearance, a webinar, a client demo, a voice note left on a shared channel. Then ask how long they think each recording is, in seconds. Almost everyone will realise they've already produced enough source audio for a decent voice clone. That's the honest starting point for every conversation about deepfake defence at work.
Three flavours of deepfake
Deepfakes at work come in three flavours. Each is used for slightly different attacks, and each has slightly different tells. Before you tap the cards below, have a think: which of the three is your organisation most exposed to right now, and why?
Tap each card to reveal.
Still images
Fake photos of events that didn't happen. The viral image of Pope Francis in a white puffer jacket (March 2023) was one of the first mainstream examples that fooled millions of people at speed. In a business context these turn up as fake screenshots, fake signed contracts, fake 'leaked' photos of executives, and fake product mock-ups used for market manipulation.
Audio
Voice clones built from short samples of real speech. In 2019 the CEO of a UK-based energy firm was defrauded of about £200,000 by a cloned voice on a phone call, believed to be the first documented deepfake-audio fraud. Since then the technique has become effectively free and near-instant. Used for CEO-fraud voice notes, 'family in trouble' scam calls, and impersonation of clients on the phone.
Video
Talking-head fakes, either pre-recorded or, more recently, generated live on a video call. The Arup Hong Kong case in February 2024 (roughly $25 million lost after a video conference where every colleague was faked) is the current textbook example. Ferrari had a very similar attempt in summer 2024, caught only because the impersonated executive was asked a personal question the imposter couldn't answer.
Each flavour rides its own curve of quality and cost. Short audio clips are already essentially indistinguishable from the real thing. Still images are almost there. Live video is close behind. Any defence strategy that only covers one of the three has already lost.
In pairs, spend three minutes sketching your organisation's exposure across the three flavours. Ask: whose face is publicly available in enough quality to fake (executives on YouTube, staff on LinkedIn video posts, spokespeople in news clips)? Whose voice? Which teams routinely receive urgent requests over voice or video? Come back with one weakest-link answer per pair. There are almost always more than people expect.
Why this is a workplace problem, not a Twitter one
For the first couple of years of the current deepfake wave, most of the visible damage was reputational. A doctored photo went viral, a public figure had to issue a statement, the news cycle moved on. Unpleasant, but survivable, and not something most finance directors lost sleep over.
That picture has changed. Deepfakes have started to cost businesses actual money. The 2019 UK energy company case (about £200,000 to a cloned CEO voice) was novel enough to make the front pages. By 2024 the Arup Hong Kong case saw roughly $25 million transferred after a video conference full of deepfaked colleagues. Ferrari fielded a near-identical attempt in the same summer, saved only by an alert executive who asked a personal question the imposter couldn't handle. Insurance underwriters and internal audit teams have quietly moved deepfake risk from 'emerging' onto their live risk registers.
And it doesn't stop at wire fraud. Fake earnings clips have briefly moved share prices. AI voice clones have been used to impersonate candidates in remote job interviews. Fake HR references are already showing up in verification checks. Reputation attacks now include synthetic 'clips' of executives saying things they never said, resurfaced conveniently before an acquisition or a bid. Fake-media risk sits next to phishing and ransomware as a category that any serious organisation now has to plan for.
In groups of three, draft a one-slide brief for your senior leadership team on your organisation's top three deepfake exposures over the next twelve months. Constraint: each exposure must name a business consequence (money, reputation, regulatory), not just a technical possibility. Share two briefs in the room and let the group critique for realism. This is exactly the exercise a mature risk function does once a year, and it's better done now than in the aftermath of an incident.
Four habits that catch nearly everything
You don't need a forensics lab or an in-house AI team to catch most of the deepfakes you'll meet at work. Four small habits, done consistently, will handle the bulk of what actually reaches your inbox, your team's WhatsApp, and your video calls.
Before you tap the cards, have a guess: if you were writing your organisation's one-page verification protocol tomorrow, what would you put on it?
Tap each card to reveal.
Reverse image search first
Google Lens or TinEye will find the earliest known copy of any image on the open web in about ten seconds. If a 'leaked' photo of your CEO has zero prior copies anywhere online, that's your first substantial flag. Real events almost always leave a trail. Fabricated ones don't.
Look for provenance markers
Content Credentials (part of the C2PA open standard, backed by Adobe, Microsoft, the BBC, Reuters, and the New York Times) attach a cryptographic history to real media, showing where the file came from and how it has been edited. Absence isn't proof of fakery, but presence is meaningful evidence of authenticity. Expect these to become as normal as EXIF data over the next few years.
Ask whether the underlying event actually happened
Verify the claim, not just the file. If the photo shows a well-known executive at a controversial event, do multiple credible outlets report the event? Do independent sources place that person there? A real event has a paper trail across many sources. A faked event has just the fake.
Verify through a second channel
This is the golden rule for audio and video impersonation, and the single most valuable habit in this whole lesson. If your 'CEO' rings, texts, or Zooms with an urgent money request, hang up and call her directly on the number already in your phone. Thirty seconds of out-of-band confirmation stops nearly every voice-fraud attack on record, including all the well-known ones.
These four habits fit on a sticker. Get them written into your team's operating procedure, drill them once a quarter, and you'll be ahead of the overwhelming majority of the fakes coming your way. The remaining sliver is what your forensics vendor is for.
In pairs, spend eight minutes turning the four habits into a real one-page verification protocol for a named team in your organisation (finance is the obvious starting point). Each habit needs a named owner and a named threshold: at what value or urgency does the second-channel check become mandatory? Come back with your one page and read the money threshold out loud. If nobody in the room has ever written this down before, that itself is the answer to whether it needed doing.
Bad prompt, good prompt (when you ask AI to help verify)
The four habits are your process. There is also something you can do better at the AI-assist step itself: word the verification prompt properly.
If you paste a suspicious image into ChatGPT (or Gemini, or Claude) and simply ask 'is this real?', you'll get a hedge. The AI can't actually run a reverse image search from your desk, it doesn't have access to your Content Credentials checker, and it doesn't want to give you a wrong yes or no. You'll get 'I can't verify that with certainty', which is honest but useless.
If you paste the same image and ask for a structured verification workflow, you'll get something genuinely useful. Same tool, same file, wildly different output. Here are the two versions side by side.
Is this photo real or AI-generated?
The AI can't establish ground truth from a single image alone and will (correctly) hedge. You'll get a shrug wrapped in polite hedging, which is honest but stops you doing any of the actual verification work. You've spent a call for no gain.
I want to verify whether this photo is authentic. Please: (1) describe any visual artefacts you can see that suggest AI generation (odd hands, jewellery, teeth, background text, lighting or shadow inconsistencies, warped reflections); (2) draft two Google reverse-image-search queries I could run right now; (3) list the earliest plausible publication date given the content of the image, plus 3 to 4 credible outlets I should check for corroborating reports of the underlying event; (4) tell me what a Content Credentials check on this file would (and would not) confirm.
Names the workflow, names the deliverables, and forces the AI to give you an actionable checklist rather than a disclaimer. You still do the verification, but you're now doing it three times faster with a proper set of leads.
In pairs, write a structured verification prompt for one of the following scenarios: (a) a WhatsApp voice note that claims to be from your CEO; (b) a video clip that appears to show a competitor's product exploding; (c) a screenshot of a signed contract, sent by a supplier you've never spoken to. Compare two of the prompts in the room. Look for: is the workflow named? Are the deliverables specific? Does the prompt push the AI toward action or toward opinion? Opinion is useless. Action is the game.
So what would you do?
Friday, 4:47pm. A WhatsApp voice note from your CFO lands in the finance team chat. It's her voice, and you're certain of that: you speak with her twice a week. She sounds like she's on speakerphone in the back of a taxi. 'Really sorry, urgent one. I'm about to board a flight, and I need you to push through the £180,000 payment to the new supplier I mentioned this morning. I know it's outside the approval window, but I've cleared it with Michael, the paperwork will come Monday. Please just do it now, I'll be off signal for the next four hours.'
You don't remember her mentioning a new supplier this morning. You do remember her joking, more than once, about payment approval workflows being annoying but necessary. Nothing on any recent memo mentions a £180,000 payment. But the voice is unmistakable, the ask is urgent, and Michael, the operations lead, is genuinely often in these conversations.
You have the four habits. You have the prompt-compare rule. You need to decide in the next couple of minutes.
Read the four options below out. Ask the room to raise a hand for their pick. Then invite three people to explain: one for the strongest option, one for whichever tempting-but-weak option got the most hands, one for what your own organisation's current process would actually do in practice. Push people to name which of the four habits (or the prompt-compare rule) they are leaning on.