Call Assistant
← Back to blog

When you can't trust your eyes or ears: how to protect yourself from fraud in the age of AI

Not long ago we relied on simple signals. Mum's voice on the phone meant Mum was calling. A video with the CEO meant it was the CEO. A book with an author's name on the cover meant that author wrote it. A thesis with a student's signature meant the student wrote it.

Generative AI has broken these habits. A voice can be cloned from a short recording, a face can be "worn" by someone else on a video call, a book or a dissertation can be generated in an evening, and a photo "from the scene" can be drawn from scratch. Fraud has become cheaper, faster and more convincing.

This article covers the whole range: from school and university work that students did not do themselves, through fake books, dissertations and artworks, to cyber fraud with fake voices, videos and call forwarding. For each situation you will find warning signs, simple rules and tools that help you check. The main idea is simple: when you can't trust your eyes and ears, trust a verification routine.

Part 1. School and university: when the work isn't the student's

The most everyday kind of "fake" is work handed in under one's own name but done by someone else: an older sibling, a paid "writer", and increasingly a chatbot. The teacher receives a neat text, but the student hasn't gained the skills the assignment was meant to build.

Why "just run it through a detector" is a bad idea

AI text detectors estimate probability; they do not establish facts. They are more often wrong on short texts, on texts by people writing in a second language and on very formulaic genres such as instructions or school essays. Accusing a student based on a percentage alone is unfair and risky.

It is more useful to design learning so that faking is hard and pointless:

Tools for teachers and lecturers

If you do need checks, use several signals at once:

What parents can do: ask about what the child understood rather than the grade; ask them to tell you what the essay is about; explain that AI is a learning assistant, not a substitute for learning. A child who uses AI honestly will always be ahead of one who can only copy with it.

Part 2. Fake books, dissertations and research papers

Generated books

Large online marketplaces now carry books written entirely or almost entirely by AI: travel guides, cookbooks, "biographies", mushroom-foraging guides, knock-offs of popular releases with similar titles and covers. Sometimes they carry the name of a real author who was never asked.

How to avoid buying a fake:

Publishers and editors can use the same AI text detectors, such as Originality.ai, Pangram and GPTZero, but again as a signal for manual review rather than an automatic decision.

Dissertations and research papers

In science, fakes are even more dangerous: other researchers cite fabricated results and decisions are made on them. Besides classic plagiarism, today you find:

Tools used by editors, universities and researchers themselves:

Tip for PhD students and researchers: keep raw data, lab notebooks and version history. That is your best defence if your work is ever questioned.

Artworks, illustrations and music

Faking in art has two sides. One is passing off generated work as your own handmade work (in competitions, commissions, marketplaces). The other is generating work "in the style of" a specific artist without consent and selling it as "almost theirs".

There are detectors for images: Hive Moderation, AI or Not, Illuminarty, Sightengine. They give a probability, not a verdict, and are more often wrong on edited or compressed images.

It is more reliable to look at where a file came from. The Content Credentials (C2PA) standard, backed by Adobe, Microsoft, Google, OpenAI, Meta and hundreds of other companies, lets a camera, editor or generator attach a signed record: who created or changed the image, when and with what. Google embeds an invisible SynthID watermark into content from its models and lets you check a file through Gemini. The limitation of both: if there is no label, that does not prove the image is genuine.

Artists who don't want their work copied can use free tools from the University of Chicago: Glaze makes it harder for models to copy a style, and Nightshade makes images "poisonous" for unauthorised training. They don't guarantee full protection, but they raise the cost of theft.

Part 3. Cyber fraud: voice, video, calls

A relative's cloned voice

The scheme is old, the tools are new. A scammer used to call and say "Mum, I've had a car accident, I need money", hoping that in a panic you wouldn't notice a stranger's voice. Today a short recording, for example a video from social media, is enough to clone a voice. The US Federal Trade Commission warns plainly: don't trust the voice, even if it sounds exactly like someone you love.

Warning signs:

What to do:

Tools can help but don't replace these rules. Hiya and Truecaller show who's calling and flag spam and scam numbers; Hiya is also developing real-time AI voice detection. ElevenLabs offers a free classifier that shows the probability a recording was made with ElevenLabs technology (it doesn't recognise other providers' voices). AI Voice Detector analyses uploaded audio. For businesses such as call centres and banks there are specialised systems: Pindrop, Resemble AI (Resemble Detect), Reality Defender.

Deepfake videos and video calls

The best-known case happened in early 2024 in Hong Kong. A finance worker at the engineering firm Arup received an email from the "CFO", had doubts, but then joined a video call where he saw and heard the "CFO" and "colleagues". All of them were deepfakes. After the call he made 15 transfers totalling about 25 million US dollars. Arup confirmed that fake voices and images were used in the fraud.

How to protect yourself:

For checking videos and photos:

Call forwarding and "codes from SMS"

Many banks and messengers confirm sign-in with an SMS code or a call. Scammers don't need to steal your phone, only to make those calls and messages reach them.

How it looks in practice:

What to do:

Phishing messages, sites and "great deals"

AI writes phishing emails without mistakes and in any language, so "an email full of grammar errors is a fake" no longer works. Check the substance, not the language:

Checklist: rules that work even when technology deceives

1. Urgency is a red flag. The harder someone rushes you, the more time you should take to check.

2. Verify through another channel. A call: call back on a known number. An email: phone. A video: message the person in your usual messenger.

3. A family code word and a "two keys" rule at work.

4. No SMS codes, PINs, CVVs or passwords to anyone. Not to the "bank", the "police" or the "operator".

5. Don't dial dictated combinations of stars and hashes. Check forwarding with *#21#.

6. Two-factor protection via an app or key, not SMS.

7. Photos and videos: look for the original source. Reverse search, official channel, metadata.

8. A detector is a hint, not a verdict. Especially when accusing a real person.

9. Less public data means less material for fakes. Limit access to videos with your voice, your date of birth, your children's names.

10. Talk about this with your family, especially older parents and children. The best defence against fraud is someone who knows the scheme in advance.

What to do if you've already been scammed

FAQ

Can you tell for sure that AI wrote a text?

No. Detectors give a probability estimate and make mistakes both ways. It's more reliable to look at the process: drafts, version history, the author's ability to explain their text.

How much audio is needed to clone a voice?

Modern services need only a short clip, and such clips are often taken from social media videos. So don't trust a voice alone, and agree on a code word.

Does asking someone to turn their head during a video call help?

It may expose a simpler deepfake, but it's no guarantee. The most reliable step is to confirm the request through another channel.

What does the **21*number# code I'm asked to dial do?

It turns on unconditional forwarding of all calls to that number. Don't dial it at someone else's request. You can check forwarding with *#21# and cancel all forwarding with ##002#.

Is it safe to upload photos to online checkers?

Not always. Some services are public or share files with partners (FotoForensics, VirusTotal). Don't upload personal photos or documents.

If an image has no AI watermark, is it genuine?

No. Content Credentials and SynthID label only part of the content, and metadata can be stripped. A missing label is not proof.

How can I protect my drawings from style copying?

Glaze and Nightshade make it harder for models to train on your work. Posting images at lower resolution and with a watermark also helps.

Summary

AI didn't invent fraud; it made it cheap and convincing. So protection today depends less on sharp eyes and ears and more on habits: call back on a known number, don't rush, never read out codes, confirm money requests through another channel, find the original source of photos and videos, and judge learning by the process rather than only the result. The tools in this article help, but the strongest protection is agreements between people who trust each other.

Facts checked: 27 October 2026. This article is for information only and is not legal advice. Service terms may change, so check them on the official websites.