↑

How to spot AI fakes and AI bots.

Field-tested techniques to check whether text, images, voices, videos, website visitors, or chat partners are AI-generated, and how to build verification into your business.

Free Guide

Detecting artificial intelligence (AI) content starts with a reality check.

There is no single tool that reliably tells you whether something was made by AI. Even OpenAI withdrew its own AI-text classifier in 2023, citing its low rate of accuracy.[1] The dependable approach is the one digital forensic investigators use: combine provenance, technical evidence, context, and out-of-band verification, then weigh the results together.

The SIFT habit (from digital-literacy educator Mike Caulfield): Stop before reacting, Investigate the source, Find better coverage, and Trace claims, quotes, and media back to the original. It works on AI fakes and human ones alike.
Deepfakes Explained

What deepfakes are, and why they matter to small businesses.

A deepfake is audio, video, or an image created or altered by AI to make someone appear to say or do something they never did. The technology is cheap and widely available, and criminals use it mostly for one thing: getting money or access. In 2024, an employee at global engineering firm Arup sent about US$25 million to fraudsters after a video call in which every other participant, including the “chief financial officer,” was a deepfake (CNN, May 2024).

Show detailsHide details
TypeWhat it isHow it is used against businesses
Voice cloneAI copies a voice from a short recording, such as a podcast, webinar, or voicemail greeting“Urgent” calls from the boss requesting wire transfers, gift cards, or password resets
Face swap or puppet videoA real person's face is mapped onto someone else in live or recorded videoFake video calls with executives; fake job candidates in remote interviews
Lip-sync videoReal video of a person, with new AI-generated words and matching mouth movementsFake statements from leaders posted to damage reputation or move stock prices
Fully synthetic person or imageA face, document, or scene that never existedFake profiles, fake identity documents, fake invoices, and fake reviews
AI-written messagesFluent, personalized emails and texts in any languageMore convincing phishing and business email compromise with fewer spelling mistakes
The key lesson: you cannot rely on spotting a good deepfake by eye or ear. Protect the process: call-back verification, two-person approval for payments, and code words for executives. Download our free Deepfake & AI Fraud Verification Checklist (PDF).
Text

Checking whether text was written by a large language model (LLM).

Show detailsHide details

Signals worth noticing

  • Polished but generic writing with few specific, checkable details
  • Citations, quotes, statistics, or case law that do not exist when you look them up
  • Uniform sentence rhythm, tidy summaries, and the same structure in every paragraph
  • Confident claims about recent events the writer could not know
  • Leftover chatbot phrasing, such as "as an AI language model" or "certainly, here is"

Better evidence than a detector score

  • Verify every reference. Fabricated sources are the strongest single tell.
  • Look at process evidence: document version history, drafts, and edit timestamps.
  • Ask follow-up questions about the content; authors can explain their reasoning.
  • Check for watermarks where vendors offer them, such as Google's SynthID for text from Google models.[3]
Caution: AI-text detectors produce false positives, and they are known to misflag writing by non-native English speakers. Never discipline an employee, student, or vendor based on a detector score alone.
Images

Spotting AI-generated and AI-edited images.

Show detailsHide details

Look closely

  • Hands, teeth, ears, and jewelry with extra, merged, or missing parts
  • Garbled or nonsensical text on signs, badges, and documents
  • Shadows and reflections that do not match the light sources
  • Backgrounds that melt, repeat, or bend straight lines
  • Skin that is too smooth, and hair that blends into the background

The newest generators make fewer of these mistakes, so a clean image is not proof of authenticity.

Check provenance and metadata

  • Content Credentials: upload the file to contentcredentials.org/verify to read Coalition for Content Provenance and Authenticity (C2PA) data showing how it was made and edited.[2]
  • SynthID Detector: Google's portal checks for watermarks embedded by Google AI tools.[3]
  • Reverse image search: Google Lens, Bing Visual Search, or TinEye can find the original or earlier versions.
  • Metadata: tools such as ExifTool show camera model, software, and timestamps. Missing metadata proves nothing, because most social platforms strip it.
  • Suspicious files and links: check them against dozens of antivirus engines at VirusTotal. For confidential files, search by the file's hash instead of uploading it.
  • Error level analysis: FotoForensics can highlight areas edited differently from the rest of an image. Interpret with care.
Video & Deepfakes

Checking video and live video calls.

Show detailsHide details

Recorded video

  • Lip movements that drift out of sync with speech
  • Unnatural blinking, frozen expressions, or stiff head movement
  • Flicker or blur around the hairline, jaw, glasses, and earrings
  • Lighting on the face that does not match the room
  • Use the free InVID-WeVerify plugin to split video into keyframes and reverse-search them.

Live video calls

  • Ask the person to turn their head fully sideways, or pass a hand in front of their face.
  • Ask a question only the real person would know, and that is not on social media.
  • Watch for call lag, mismatched voice tone, and pressure to act quickly or secretly.[4]
  • End the call and call back on a number you already have on file.
Voice Cloning

Stopping AI voice-clone fraud.

A few seconds of public audio is enough to clone a voice convincingly. The Federal Bureau of Investigation (FBI) warns that AI voice clones can sound nearly identical to the real person.[4] Process controls beat ears.

  • Callback rule: verify any request for money, credentials, or data by calling a known number, never the number that called you.
  • Code words: agree on a verbal passphrase for executives, finance staff, and family members.
  • Dual approval: require two people to approve payments, wire changes, and new vendor bank details.
  • Slow down urgency: train staff that "urgent and confidential" is a warning sign, not an instruction.
  • Limit voice samples: be thoughtful about long public recordings of executives where they are not needed.
Chat & Social Media

Is the account or chat partner a bot or a person?

Show detailsHide details

Common signs

  • Long, polished replies that arrive instantly at any hour
  • Answers that never reference shared history or specific details you mentioned earlier
  • New accounts with stock or AI-generated profile photos and generic bios
  • Networks of accounts posting near-identical wording

Simple tests

  • Ask about something local, current, or personal that requires real experience.
  • Ask the same question twice in different words and compare the answers.
  • Ask directly. In California, bots used to sell products or influence votes must disclose that they are bots, and the European Union (EU) AI Act adds AI-interaction transparency duties.[7]
  • Move important conversations to a verified channel such as a video call or known email address.
AI Checkers

AI detectors and checkers: what they can and cannot tell you.

Detection tools look for statistical patterns, hidden watermarks, or provenance records. They are useful clues, but every one of them produces false alarms and misses. Researchers at Stanford found popular AI-text detectors frequently misflagged essays by non-native English writers as AI-generated (Liang et al., 2023), and OpenAI withdrew its own text classifier for low accuracy.

Show detailsHide details
ContentExample toolsHow they workLimits
TextGPTZero, Originality.ai, Copyleaks, TurnitinScore how predictable the wording is compared with known AI writingFalse positives on simple or non-native writing; easily fooled by light editing
Images and videoHive, Reality Defender, SensityMachine learning models trained to spot generation artifactsStruggle with new generators, compression, and screenshots
WatermarksGoogle SynthID DetectorLook for invisible watermarks embedded by a specific AI providerOnly detect content from providers that watermark
ProvenanceContent Credentials VerifyRead signed records of how a file was created and editedMissing when platforms strip metadata or creators do not sign
Voice and callsReality DefenderAnalyze audio for synthetic speech in real timeEnterprise products; process controls still essential

Use checkers wisely

  • Use them to decide what deserves a closer look, not to reach a verdict
  • Combine at least two methods, such as provenance plus reverse image search
  • Document results if a decision or investigation depends on them
  • Learn from real cases in the AI Incident Database

Never

  • Discipline an employee, student, or applicant based only on a detector score
  • Upload confidential documents to a free online checker
  • Treat “no watermark found” as proof something is real

Tools are listed as examples, not endorsements; most offer free trials or tiers with paid plans.

Your Website: Front End

Detecting AI bots and automated agents on your website.

AI crawlers, scrapers, and browser-driving AI agents now make up a meaningful share of web traffic. Some announce themselves; others pretend to be people.

Show detailsHide details

What you can see in the browser and analytics

  • Visits with no mouse movement, scrolling, or time on page, or impossibly fast form fills
  • Spikes from data-center networks or unusual countries, at odd hours
  • Headless-browser signals, such as automation flags set in the browser
  • Form submissions that fill hidden "honeypot" fields a human would never see

Defenses

  • Add honeypot fields and server-side validation to forms
  • Use a privacy-friendly challenge such as Cloudflare Turnstile on forms and logins
  • Enable bot management or a web application firewall (WAF) with rate limiting
  • Publish AI-crawler preferences in robots.txt, knowing that compliance is voluntary
Common AI crawler (user agent)Operator and purposeHonors robots.txt?
GPTBotOpenAI, collects content that may be used to train modelsYes[5]
OAI-SearchBotOpenAI, surfaces sites in ChatGPT search resultsYes[5]
ChatGPT-UserOpenAI, fetches a page when a user asks ChatGPT toMay not apply, because a user initiated it[5]
ClaudeBotAnthropic, model training crawlerYes, per the operator
Google-ExtendedGoogle, a robots.txt token that controls use of content for Gemini trainingYes, per the operator
PerplexityBotPerplexity, search indexingYes, per the operator
CCBotCommon Crawl, open web dataset widely used for AI trainingYes, per the operator
Watch for spoofing: anyone can fake a user-agent string. Confirm that a "GPTBot" or similar visitor really comes from the operator by checking its Internet Protocol (IP) address against the operator's published IP ranges (OpenAI publishes JSON lists for each crawler) or by a reverse Domain Name System (DNS) lookup.[5]
Your Systems: Back End

Finding bots and AI agents in server logs and APIs.

This section is written for your information technology (IT) team or managed service provider. Share it with them.

Show detailsHide details
  • Request patterns: steady, high-rate requests, sequential URL walking, and no loading of images, CSS, or JavaScript.
  • Network origin: group traffic by autonomous system number (ASN) to spot cloud and hosting providers posing as home users.
  • Fingerprints: Transport Layer Security (TLS) fingerprints such as JA3 and JA4, plus header order, often reveal automation tools behind a browser-like user agent.
  • Application programming interface (API) abuse: watch for credential stuffing, scraping through APIs, and token sharing across many IP addresses.
  • Agent behavior: AI agents that drive a real browser can look human, so watch for unusually consistent timing and task-only navigation.
  • Centralize logs in your security information and event management (SIEM) tool and alert on spikes, new ASNs, and failed-login bursts.
Vendors & Products

Is a product or vendor quietly using an LLM?

Many software products have added AI features that send your data to a model provider. Here is how to find out.

  • Read the privacy policy, terms, and subprocessor list for AI model providers such as OpenAI, Anthropic, Google, or Microsoft Azure OpenAI.
  • Open your browser's developer tools (Network tab) while using the feature and look for calls to AI APIs or streaming responses.
  • Notice word-by-word "typing" output and slightly different answers to the same question, both typical of LLMs.
  • Ask the vendor directly in your security questionnaire: which models, where data is processed, retention, training use, and whether you can opt out.
  • Request an AI bill of materials (AI-BOM) or model card for critical systems.
Verification Playbook

A five-step check before you trust, share, or act.

1. Pause

Strong emotion and urgency are the attacker's tools. Slow down.

2. Source

Who published it first? Is the account, domain, or number real?

3. Provenance

Check Content Credentials, watermarks, metadata, and reverse search.

4. Corroborate

Find independent, reputable coverage or a second channel.

5. Verify out of band

Call back on a known number before money, data, or access changes hands.

For organizations, the National Institute of Standards and Technology (NIST) report AI 100-4: Reducing Risks Posed by Synthetic Content describes provenance, watermarking, and detection approaches in more depth.[6]

Nexus Security Advisors can run deepfake and voice-clone fraud tabletop exercises, write verification and payment-approval procedures, review website bot defenses, and support digital forensic analysis of suspected synthetic media.

Sources

References

Sources cited on this page (last reviewed October 31, 2026)
  1. Search Engine Land, OpenAI's AI Text Classifier no longer available due to "low rate of accuracy", July 2023. searchengineland.com
  2. Content Authenticity Initiative, Using the Verify tool. opensource.contentauthenticity.org
  3. Google, SynthID Detector: a new portal to help identify AI-generated content. blog.google
  4. FBI Internet Crime Complaint Center (IC3), Public Service Announcement I-072026-PSA, July 20, 2026. ic3.gov
  5. OpenAI, Overview of OpenAI Crawlers. developers.openai.com
  6. NIST, AI 100-4: Reducing Risks Posed by Synthetic Content, November 2024. doi.org/10.6028/NIST.AI.100-4
  7. Orrick, EU AI Act Update: Digital Omnibus Finalizes 8 Compliance Changes, July 2026. orrick.com

Related free guides: AI Explained for Business Leaders · Responsible & Ethical AI

Let's Talk

Start with a free 30-minute discovery call.

Tell us your industry, size, and the security or compliance challenge in front of you. We'll tell you plainly whether we're a fit.

We reply within one business day.