GACS will never ask for your seed phrase, private keys, or payment. Always free.
GACS Logo
GACS will never ask for your seed phrase, private keys, or payment. Always free.
GACS Logo

AI Scam Report 2026 — voice clones, deepfake video, and LLM-written fraud

The five ways generative AI is materially changing scams in 2026 — and the defences that still work against each one. Reference page for journalists, researchers, and platform trust-and-safety teams.

1. Voice cloning is now real-time and 3-second-sample

Production voice-cloning APIs in 2025 collapsed sample-length requirements from 30s to under 3s and shifted from batch synthesis to real-time bidirectional conversation. The McAfee 2024 study (cited on /scam-statistics-2026) found that 1 in 4 surveyed adults knew someone who had received a voice-cloned scam call. In 2025 the same study found 1 in 3. The classic 'grandkid in jail' scam is now indistinguishable from the real voice for the average listener.

What works against it: Family verification phrases. Never trust voice alone for money transfers. Call back on a known number. See /spot-x-impersonator pattern #5 — the same logic applies to voice.

2. Video deepfakes have arrived in job interviews

Multiple national-press cases in 2024 and 2025 documented fraudsters using real-time face-swap during Zoom and Teams interviews to either fraudulently obtain remote-work positions or to impersonate hiring managers and steal candidate credentials. The North-Korean IT-worker fraud schemes flagged by the US Treasury and FBI are the highest-profile examples. The technology is now consumer-grade.

What works against it: Ask the candidate or interviewer to turn their head, cover their face with a hand momentarily, or hold up an unusual object — real-time face-swap pipelines still degrade on occlusion. Verify the email domain. Use a known phone number to confirm a hire.

3. LLM-written romance scripts have replaced the broken-English giveaway

For a decade, the easy tell of a romance scam was awkward English. Open-source and commercial LLMs have removed that signal completely. 2025 reports to GACS show scripts now adapt fluently across English, French, German, Spanish, Portuguese, Hindi, and Mandarin — often in the same conversation. The scam structure (slow trust-building, then an investment platform) is unchanged; only the language quality has improved.

What works against it: Focus on the structural tells — never met in person, refuses video, conversation eventually pivots to crypto/investments — instead of language quality. See /pig-butchering-scam for the full pattern.

4. AI-generated profile pictures dominate new impersonator accounts

Across GACS-confirmed impersonator reports submitted in 2025, the majority of profile pictures were synthetic (diffusion-model generated) rather than scraped from the real owner. This breaks reverse-image lookup as a defence and means defenders need to consider account-level signals (join date, follower-to-following ratio, posting cadence) instead.

What works against it: Stop relying on image-only checks. Run the full 6-step verification routine at /verify-any-social-account.

5. AI-written phishing emails get past the obvious filters

Spam filters trained on broken-English Nigerian-prince templates miss LLM-written phishing that is grammatical, brand-tone-matching, and personalised from public LinkedIn data. Microsoft and Google have responded with model-based classifiers, but the arms race continues. The 2025 Verizon DBIR notes that AI-assisted phishing now contributes to the majority of credential-theft incidents in their dataset.

What works against it: Hover over links. Type URLs by hand. Treat every 'verify your account' email as hostile. Use /link-checker for any URL you're unsure about.

The big picture

AI hasn't created new scam categories — it has industrialised the existing ones. The pig-butchering script, the recovery-scam DM, and the fake-support reply are all the same playbook with better text, voice, and pictures. That means the defences are also the same playbook, just harder to execute by eye. Tools that automate the check (like the GACS scanner) become the practical answer for individuals; structural verification routines (like the 6-step check) become the practical answer for everyone else.

Related

Explore the full GACS map