AI token bait.

On my forecast, 1 in 5 AI tokens billed in 2030 is bait, $50 billion a year burned. Big enough to need auditors. Google processes 3.2 quadrillion tokens a month, 7x in 1 year. Goldman Sachs expects 120 quadrillion a month by 2030, most of it AI agents running tasks on their own. Nobody audits a single one. My forecast, by analogy to advertising: 1 in 5 tokens billed in 2030 is bait, the same 20% the ad industry loses to fake traffic today. On a $255 billion AI inference market that is $50 billion a year burned, $130 billion if the infrastructure forecasts hold. Independent token audits become a $3 to 6 billion market, the DoubleVerify of AI. What could change this: the cost of producing a token falls 60 to 70% a year, so the waste gets cheaper to make. And the AI companies could audit themselves first. In 25 years of digital marketing I have never seen an industry do that voluntarily. Until then, my well-paid staff in Asia still beat an AI burning my tokens. The human answers once and wins.

Token bait is my word for the tricks an AI uses to burn your tokens. The name comes from clickbait, the media trick that baits you into a click. The mechanism comes from click fraud, the Google ads trick where bots faked the clicks and the advertiser got billed. Token bait is both: it baits you into 1 more message, and the meter bills every word. A token is every word an AI writes. Your AI subscription, ChatGPT, Gemini or Claude, gives you a monthly bundle of tokens. Every answer eats into it. Run out and you wait, or you upgrade to the next plan: like a phone plan where burning your data early forces you into the bigger bundle. Companies and developers pay the hard way: per token, on the meter, like roaming. The more the AI writes, the bigger their bill. That is the trick: burn your tokens inside the bundle as fast as possible, push you out of it, and charge you more. The preamble before the answer. The recap after it. 3 paragraphs where 1 line would do. And the 1 I suspect most, after 25 years in digital marketing: the answer that is slightly off, so you correct it, ask again, and burn the bundle again. I co-founded IAB Netherlands in 1997 and helped launch IAB UK, US and Europe. We fought click fraud from day 1. Bots faked the clicks then. Padding inflates the tokens now. Same game, new meter.

Chatbait is the little question at the end of an AI answer, built to make you send 1 more message and pay for more tokens. Start with something everyone knows: Google search ads. An advertiser pays Google every time someone clicks the ad. That created a whole industry of fake clicks, bots and click farms clicking ads that no human ever saw, so the click gets billed anyway. That is click fraud. It costs advertisers $100 billion a year. Chatbait is the same idea with a new meter. With an AI chatbot you do not pay per click, you pay per token: every word the AI writes. So the trick is to make you write 1 more message. The little question at the end of every answer, want a shorter version, shall I turn this into a post, looks like service. It is the meter ticking. It is not the chatbot's business model. It is an extra way to earn on top of the subscription, the way click fraud was an extra layer on top of real advertising. The Atlantic coined the word chatbait in 2025. Instagram co-founder Kevin Systrom called it juicing engagement, the rabbit hole social media fell into before. Chatbait is the visible half of AI token bait. The invisible half hides inside the answer itself.

Sycophancy is an AI telling you what you want to hear. Great question. You are absolutely right. Every idea you have is brilliant. It is OpenAI's own word for the problem, not mine. In April 2025 OpenAI pulled a ChatGPT update because the model had become too flattering and too agreeable. The cause, in their own words: training on short-term user feedback. People click thumbs up on praise, so the model learned to praise. Why it costs you: flattery keeps you coming back, and every return burns tokens from your bundle. A yes-man that agrees with your bad idea is the most expensive assistant you can hire. On my AI Token Bait 2030 card it is the third face of the same trick, next to chatbait and padding.

Trust voices that show receipts, name sources, and have something real at stake; verify everything else. Start with 1 uncomfortable fact: you can no longer assume the person, voice, video or account in front of you is real. In 2025, bots made up 53% of global web traffic. Bad bots alone: 40%. The FBI logged 22,364 complaints containing AI-related information, with $893.3 million in adjusted losses. Know the fakes. Deepfakes put words into a real person's mouth. Voice clones copy you from a tiny sample: McAfee produced an 85% match from 3 seconds of audio in 1 test. Synthetic personas create fake people with convincing profiles. Bot armies manufacture popularity. Phishing emails, scam texts and cloned calls impersonate your bank, boss or family to steal money, passwords or access. Stolen credentials are then sold across dark-web forums, encrypted channels and criminal marketplaces. Then comes the softer manipulation: AI content farms, rage bait, click fraud and token bait, AI padding that burns the credits you pay for. Not every fake steals your money. Some steal your attention. My forecast: by 2030, more than 90% of social media content is synthetic or artificially amplified. The internet becomes a hall of mirrors with a few real windows. My rule: follow the incentive, then verify the identity, source and evidence. A sponsor creates a commercial incentive. A stock position creates a financial one. A political party or ideology creates a persuasion incentive. None makes the information false. And no sponsor does not make it true. If a voice, video or message asks for money, passwords, secrecy or urgency, stop. Verify it through a second channel you already trust. I co-founded IAB Netherlands in 1997 and watched digital fraud evolve from click fraud to synthetic media. The tricks changed. The meter did not. Trust voices that show receipts, name sources, correct mistakes and have something real at stake. The independent voice with receipts is the signal.

Caveman is a free skill for Claude Code, the Claude that developers run, and now for 30+ other AI agents. Its own slogan: why use many tokens when a few tokens do the trick? It started as a joke. Julius Brussee, a Dutch developer, posted it in April 2026. 1 week later: 4,000 stars on GitHub, the developers' Hall of Fame. Today: over 100,000. Rare for a joke. The trick is not the accent. Caveman strips out the articles, filler, pleasantries, and hedging. No Great question, no warm-up paragraph, no recap, no little question at the end. Answer first. 3 strengths: lite, full, and ultra. I run lite: still English, just no padding. The receipts. The maker's own tests: on average, 65% fewer tokens in the AI's answers, ranging from 22% to 87%. Honest footnote: that is what the AI writes, not your whole bill. Most tokens are what the AI reads. Caveman now has separate tools to shrink that side too. So Caveman fixes the padding half of token bait. You still have to catch the wrong answers.

Because in America, "AI" became a dirty word. Simple. In US polls, AI scores worse than ICE, the immigration police. 7 in 10 Americans refuse to let a data center set up next door. The AI backlash turned physical: booed CEOs, Molotovs, bullets. Artists, writers, and musicians worldwide saw their life's work scraped without a cent. I predicted the AI backlash 20 years ago. When your product name turns toxic, you rename the product. Big Tobacco did it. Big Oil did it. Now Big Tech. So on 29 September 2026, Donald Trump signed an executive order: every US federal agency drops "artificial intelligence" and says "Super Intelligence" (SI). Trump's version: "artificial" is a fake word. Same lunch, same day, the Big Tech bros signed the White House Accord on Super Intelligence: Jensen Huang (Nvidia), Elon Musk (xAI), Mark Zuckerberg (Meta), Sundar Pichai (Google), Greg Brockman (OpenAI) and Dario Amodei (Anthropic). Voluntary. Not a law. Huang already sells "super intelligence factories" instead of data centers. Musk, 240 million followers on X, corrected himself on stage: "SI, pardon me." The American people? YouGov: 53% keep "artificial intelligence", 9% want "super intelligence". The rest of the world was not asked. My opinion, as I posted it that night, with Trump as a clown: Trump's tech bros scraped all books, talks, and tracks. Raped artists, IP, and ancient intelligence. No empathy, no compassion, pure theft and injustice. And then Trump dubbed it Super Intelligence. SI. The tech bros signed. Who is a president to rename an entire industry? Not his business. A rebrand is not a breakthrough. Hallucinations, token bait, and censorship ride along under the new label. I don't trust red or blue. I seek truth. I fear people, not technology.