PROBLLMS
Why You Can't Trust the Robot.
Keep the title. Cut the deck in half. Graft in four 2026 angles — voice-clone scams, kids and AI companions, AI-injected ads, deepfake nudes — that show up in the polls the audience actually fills out.
AI didn't invent any of these threats — it removed the friction. That single sentence is the spine the laymen edition should hang on. Phishing click-rate · 54% vs 12%.
Same threat. Friction removed.
54%
Microsoft's red-team study: 4.5× more effective than human-written and far more profitable. The whole story of "AI in security 2026" is in this number — old threat, new fluency.
Versus 12% for human-written.
SOURCE · paubox.comBAIT
2.0
$40B
Deloitte's 2024 estimate. Voice clones, deepfake video calls, scripted scams. The damage is already metric — and the curve is exponential, not linear.
SOURCE · deloitte.com
deloitte · 2024
$893M
Of which $352M against adults 60+. This is the one AI risk that laymen fear personally — and it's the one number the deck should put on the wall in 96-point type.
SOURCE · scamwatchhq.com
fbi ic3 · 2025
440K
From 4,700 in all of 2023. That's a ~94× jump in 18 months. The FTC opened a Sept-2025 inquiry into seven chatbot companies over kids and companion-AI harms. Schools are seeing deepfake cyberbullying.
SOURCE · usnews.com · cnbc.com486+
80%
It invents dialogue. Mata v. Avianca: lawyer cited cases that don't exist, sanctioned. Apple Intelligence rewrote BBC headlines into nonsense. The robot is fluent and confident — confidence ≠ truth.
SOURCE · Mata v. Avianca · daringfireball.net
usdc sdny · seal
$25M
Hong Kong, May 2024. A finance worker at engineering firm Arup joined what looked like a routine video call with the CFO and colleagues. Every face on the call was a deepfake. The transfer cleared in minutes.
If you remember one story for the deck, remember this one.
SOURCE · cnn.com
cnn · may 2024
The arc — six chunks · ~60 minutes.
The chatbot that swore.
DPD bot calls customers names. Audience laughs; you've earned 9 minutes.
Confident liar.
Mata v. Avianca. Apple BBC. Whisper 80%. Confidence ≠ truth.
Voice clones.
Brightwell mom. Arup $25M. FBI $893M. Personal, not theoretical.
AI being fooled.
Resume injection. Confused-deputy. The defender becomes the surface.
The honesty pause.
78 election fakes studied — apocalypse mostly didn't land. Earn trust.
Four words to write down.
Codeword. Callback. Dual-officer. Don't paste.
The deck triage.
- "ProbLLMs · Why You Can't Trust the Robot" survives. It's funny, it's accurate, it lands without tech jargon.
- The hallucination stories are the audience's mental model. Don't cut them.
- No consumer surface. Zero appearances in 2025 public-concern polls.
- Beautiful for engineers. Wrong room.
- + voice-clone scams (FBI $893M)
- + kids and AI companions (NCMEC 4.7K → 440K)
- + AI-injected ads ($100M run-rate, heading to $10B)
- + deepfake nudes (school cyberbullying pipeline)
Defenses. Workflow-level.
Family codeword
Agree a phrase the scammer can't know. Voice-clone scams collapse the moment you ask for it.
Callback verification
Hang up. Call the known number back. Don't trust the number that called you.
Dual-officer sign-off
Money instructions on a video call require a second human, on a different channel. Period. FBI · IC3 PSA.
Don't paste secrets
Assume "share" links are public. Opt out where the toggle exists. Logging is the default. bitdefender.
Two findings cut the other way. Admit them on stage.
The election deepfake apocalypse mostly didn't land.
Knight Columbia studied 78 fakes. The damage was targeted, not viral — Slovakia 2023, the Biden NH robocall ($6M FCC fine), the Arup video-call. Don't oversell the apocalypse — undersells the targeted danger.
— knightcolumbia.orgThe headline AI privacy fines have mostly been overturned.
Italy's €15M against OpenAI was scrapped March 2026. The legal wins are for piracy (Anthropic $1.5B), not privacy. Practical advice for the audience is workflow-level, not regulatory hope.
— tradingview/reutersWhen ChatGPT's $100M-run-rate ad business becomes a $10B one — does the "confident liar" become the most lucrative ad surface ever built? [src]
The eight angles.
-
01
Title candidates Eight title candidates with a recommended pick and the patterns behind each.recon 7 cites
-
02
Existing deck triage Slide-by-slide keep/trim/simplify/cut verdict on the AI-Security-Talk deck for a non-technical audience.expedition 77 cites
-
03
AI vs LLM and other foundational concepts A non-technical primer on the AI/ML/deep-learning/LLM stack and the surrounding jargon.survey 17 cites
-
04
Scams, deepfakes, and social-engineering in 2026 Field guide to the 2026 AI scam landscape — named cases, dollar losses, attacker tooling, and four defense layers.expedition 99 cites
-
05
Data harvesting and privacy AI runs on harvested data. Defaults still leak; your every prompt is logged unless you turn it off.survey 29 cites
-
06
Ads, manipulation, and trust erosion The ad-funded internet was already a manipulation engine; AI cheapens every step.survey 26 cites
-
07
Misinformation, hallucination harms, and election/medical/legal risks Where LLM hallucinations have caused measurable harm — courts (486+ cases), hospitals (Whisper 80%), elections.survey 20 cites
-
08
Presentation craft for non-technical audiences A recon-depth playbook for explaining technical work to non-technical audiences.recon 7 cites