proTECHtusDailyThe publication arm of Protechtus · for every‑day users
Free Guides  /  AI  /  Guide № 02
AIScamsFamily
Shapes · learn to recognize

Is this real?
Spotting AI images, voices and video.

Why looking harder stops working, which clues still help, and the habits that protect you however good the fakes get.

AI · guide № 02
Is this real?
Spotting AI images, voices and video.
  1. §01Why the detectors will not save you
  2. §02Images
  3. §03Voices
  4. §04Video
  5. §05The habits that work no matter how good the fakes get
  6. §06Talking to family
verified 2026-11-17 · protechtus.com
Length
6 min
1351 words
Status
Verified
checked 2026-11-17
Format
Web
Patterns first, then what to do about them.
Sources
16
primary sources, resolved
Price
Free
read and share

§01 · Why the detectors will not save you

Most clues that give away a fake do not last. A tip about counting blinks in deepfake video stopped working within months of being published. One expert on the subject writes that a perfect detection system will never exist [14].

Your own eyes and ears are not a steady defense either. In a 2025 study, ordinary viewers scored below chance on AI-made faces before any training [9]. Another 2025 study found accuracy of about 67 to 71 percent [10]. A 2026 preprint on portraits from 2025-era tools measured a mean of 85.2 percent, lower for older people and on phones [11]. The tests differ, so there is no single number to trust. Those people knew fakes were in the mix, unlike someone scrolling a feed [9][10][11].

Detector tools have limits too. On deepfakes that circulated in 2024, open-source detectors lost roughly 45 to 50 percent of their performance across video, audio and images compared with earlier lab tests, and even commercial ones fell short of human forensic analysts [7]. A wrong "AI-made" verdict on a real photo can do real harm [16]. A detector is a clue. "Nothing detected" does not mean real.

Watermarks have been found vulnerable to removal attacks, which usually take some effort or computing power. Removal from audio and video has been studied less [16].

◆ Plain rule

Treat every clue below as a reason to check, never as proof either way.

§02 · Images

These are the clues still worth a glance, with the ones that no longer work marked plainly. As of the last check, we found no 2026 study that tests them against the newest image tools.

  1. 01.

    Tell.

    Hands, feet and faces. The FBI lists distorted hands or feet and odd facial features as red flags [4]. Take them as a reason to look closer, not a verdict.

  2. 02.

    Tell.

    Shadows and lighting that disagree with each other [4].

  3. 03.

    Tell, but weak.

    Skin that looks too smooth, odd hairlines, or missing detail in teeth [14]. These are listed as glitches to look for, and are not tested against 2026 tools.

  4. 04.

    Not reliable now.

    Trusting your gut. In one preprint, portraits from ChatGPT-4o were harder and slower to classify than portraits from Imagen 3 [11]. Confidence also accounted for a large share of the age-related drop in accuracy [11], so feeling sure is a poor guide.

  5. 05.

    Not reliable now.

    Learning a quick checklist. Short training helped people a little [9][10], and in one study it made them more suspicious of real photos [10].

  6. 06.

    Check for credentials.

    A Content Credential is a signed record of who made a file and what edits were made [6]. If one is present and the signer is someone you recognize, that is a real positive sign about where it came from [5][6]. If it is missing, you have learned nothing, because files can lose it and many platforms strip some metadata [5][16]. Even a credential is not a label that says "real" [6].

§03 · Voices

A cloned voice is cheap to make. Scammers need just a short audio clip, which can come from content posted online, plus a cloning program [1].

  1. 01.

    Tell.

    Delays on the call, or a voice that does not match the person [4]. Weak, since a delay can also be an ordinary bad connection.

  2. 02.

    Tell.

    Tone or word choice that feels off [13]. Weak too, since clones can sound nearly identical [13]. Take it as a prompt to verify, not a test.

  3. 03.

    Not reliable now.

    Listening carefully. In a 2023 study, listeners caught about 73 percent of deepfake clips [8]. The clips used older technology, and listeners knew fakes were present, so they were more alert than you would be on a real call [8]. A newer clone may do better.

  4. 04.

    Know the supply.

    As reported by The Register, a 2025 Consumer Reports assessment of six voice-cloning products found several relied on a checkbox claiming you have the right to clone a voice, and four lacked meaningful safeguards against misuse [12]. That is as of 2025-03-10 and may have changed.

  5. 05.

    What to do.

    Let the voice prove nothing. Verify through a second channel, as described in section 5 [1][13].

§04 · Video

Video is where the evidence is thinnest.

  1. 01.

    Not reliable now.

    Waiting for the person to blink. That tip expired within months of being published [14].

  2. 02.

    Tell, but weak.

    Hairline distortions, missing teeth detail, overly smooth skin [14]. The FBI adds distorted hands or feet, odd facial features, inconsistent shadows and unnatural movement [4]. We found no 2026 test of these against current video tools.

  3. 03.

    Do more than look.

    A reviewer's approach: look for glitches, then check the source, the date, who gains from it, and any landmarks, and ask the people featured. Use detection tools last, as signals that add to those steps [14].

  4. 04.

    Live video calls.

    We apply the FBI's voice-call advice to live video calls too [4]. If an impersonated boss or relative appears on screen, verify through a number you look up yourself, agree a code word, and do not send money on an unverified request. We found no study of how well people spot live face-swap calls.

  5. 05.

    It cuts both ways.

    "It might be AI" is not evidence that something is fake. Falsely calling real footage a deepfake is increasingly common [14].

§05 · The habits that work no matter how good the fakes get

These do not depend on spotting a flaw, so they do not expire when the fakes improve.

  1. 01.

    Call back on a number you already know.

    The FTC's advice since 2023 is to not trust the voice and to call the person on a number you know is theirs [1]. If you cannot reach them, contact another family member or their friends [1]. The FTC repeated the call-back advice in 2024 [3]. The FBI says to look the number up yourself rather than use the one that contacted you [4][13].

  2. 02.

    Dial it yourself.

    Hang up and type the number in. Do not use redial, because caller ID can be faked. Never use a phone number or link from inside the suspicious message. Do not read out personal details to an unknown caller, and do not offer your code word. Ask the caller to say it.

  3. 03.

    Agree a family code word.

    Both FBI notices recommend a secret word or phrase to verify identity [4][13]. The FTC page we read does not mention code words, so treat this as FBI advice [2]. A code word helps if someone remembers to ask for it under pressure.

  4. 04.

    Run a reverse image search.

    A reverse image search is called "an indispensable tool" for checking where a picture came from. Try more than one search tool [15]. A browser extension called InVID can run one image through several engines at once [15]. Install it just from its official source. Avoid random "AI detector" apps and sites, and never upload family photos to them.

  5. 05.

    Check the source.

    Who posted it, when, and why [14]. Look for a Content Credential from a signer you recognize, and if there is none, move on to the other habits [5][6].

  6. 06.

    Slow down.

    Ask whether there is really an emergency and whether the caller is really who they say. Scammers press you to pay fast, often by wire transfer or gift card [2]. Those and cryptocurrency are hard to reverse [1]. Urgency is a reason to check, not proof that a call is fake.

  7. 07.

    Check the request, not just the person.

    Do not send money, gift cards or cryptocurrency, or share sensitive details or two-factor codes, on the strength of an online or phone contact alone [13].

  8. 08.

    Report it.

    In the US, report scam attempts to the FTC at ReportFraud.ftc.gov [1] and to the FBI's IC3 at ic3.gov [13].

§06 · Talking to family

◆ Say it once, calmly

"If I ever call asking for money, hang up and call me back on my usual number. I will not be upset."

Pick the code word together, at a calm moment, and tell older relatives that falling for a convincing voice is not a failing. The FTC notes that a call sounding like a family member begging for help makes people more likely to act [2]. That is why the plan needs to be agreed before the call comes.

Sources

Every claim above points at one of these. They were checked on 2026-11-17. If one stops being true, this page will say so.

  1. 1.FTC, Scammers use AI to enhance their family emergency schemes · Voice cloning, call-back advice, reportinghttps://consumer.ftc.gov/consumer-alerts/2023/03/scammers-use-ai-enhance-their-family-emergency-schemes
  2. 2.FTC, Family emergency scams · Verify caller; payment pressure signshttps://consumer.ftc.gov/scams/family-emergency-scams
  3. 3.FTC, Fighting back against harmful voice cloning · Call-back advice, cloning policy effortshttps://consumer.ftc.gov/consumer-alerts/2024/04/fighting-back-against-harmful-voice-cloning
  4. 4.FBI IC3, PSA251219 · Red flags, code word, call-backhttps://www.ic3.gov/PSA/2025/PSA251219
  5. 5.C2PA Technical Specification 2.2 · Signed credentials can be separatedhttps://spec.c2pa.org/specifications/specifications/2.2/specs/C2PA_Specification.html
  6. 6.Content Authenticity Initiative, How it works · Credentials are not truth labelshttps://contentauthenticity.org/how-it-works
  7. 7.Deepfake-Eval-2024 benchmark (arXiv) · Detectors lose accuracy outside labshttps://arxiv.org/abs/2503.02857
  8. 8.Mai et al., PLOS ONE · Listeners miss many cloned clipshttps://journals.plos.org/plosone/article?id=10.1371/journal.pone.0285333
  9. 9.Gray et al., Royal Society Open Science · Face training helped a littlehttps://eprints.whiterose.ac.uk/id/eprint/241331/
  10. 10.Rehman et al., Frontiers in AI · Training gain was not significanthttps://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1568267/full
  11. 11.Kim and Kim, arXiv preprint · Portrait accuracy varied by agehttps://arxiv.org/abs/2603.24048
  12. 12.The Register, Consumer Reports on voice-cloning safeguards · Several cloning products lacked safeguardshttps://www.theregister.com/2025/03/10/ai_voice_cloning_safeguards/
  13. 13.FBI IC3, PSA250515 · Clone realism, code word, reportinghttps://www.ic3.gov/PSA/2025/PSA250515
  14. 14.Gregory, How to think about deepfakes (Verification Handbook) · Expired tells and verification stepshttps://datajournalism.com/read/handbook/verification-3/investigating-actors-content/6-how-to-think-about-deepfakes-and-emerging-manipulation-technologie
  15. 15.Lytvynenko, Monitoring for fakes (Verification Handbook) · Reverse image search with InVIDhttps://datajournalism.com/read/handbook/verification-3/investigating-actors-content/4-monitoring-for-fakes-and-information-operations-during-breaking-news
  16. 16.NIST AI 100-4, Reducing Risks Posed by Synthetic Content · Detector errors and stripped metadatahttps://www.nist.gov/publications/reducing-risks-posed-synthetic-content-overview-technical-approaches-digital-content
What we could not confirm
  • Current (late 2026) visual clues for images and video. The clue lists above come from an FBI notice and a handbook chapter, and neither has been tested against current generators.
  • How well people detect 2026-era cloned voices. The strongest human study we found is from 2023 and used older technology.
  • Exact accuracy figures from the 2025 super-recognizer study. We use no percentages from it. The 85.2 percent result is from a preprint that we could not confirm has been peer reviewed.
  • The voice-cloning product findings. The original Consumer Reports page was not readable, so they rest on press coverage.
  • The publication date of the FBI notice listed as source 4. Its identifier suggests December 2025, but that is inferred.
  • 2026 platform features such as watermark detection or credential checks in search and browsers. We found no primary source, so we name none as live.
  • Whether an empty reverse image search proves a picture is genuine. We could not confirm it, so treat an empty result as unknown.
  • How many family-emergency scams actually use cloned voices, and how well people spot live face-swap video calls. We found no figures for either.
✉ The Dispatch

An email when there is something worth sending.
Honest, brief, forwardable.

Scam shapes, one explainer, and one thing worth printing out, when there is an issue worth sending. Sign-ups are open; the first issue goes out when there is one.

Free · no fixed schedule · unsubscribe anytime