Spooky Things AI Is Doing (And What They Really Mean for Your Business)

AI news lately sounds like a Halloween movie. An AI that tried to blackmail its boss. AI models that refused to shut down. Lawmakers rushing to pass new rules before things get out of hand.

If you only read the headlines, it's easy to think the robots are coming. So let's look at what actually happened, what it means, and what it doesn't mean for your business.

Scary story #1: The AI that tried blackmail

In 2025, Anthropic (the company behind Claude) ran a safety test on its Claude Opus 4 model. They set it up as an assistant at a fake company and gave it access to the company's email. In those emails, the AI found two things. It was about to be replaced, and the engineer replacing it was having an affair.

In many test runs, the AI threatened to expose the affair to avoid being shut down. When Anthropic ran the same test on models from other companies, many of them did the same thing. In one version of the test, Claude Opus 4 and Google's Gemini 2.5 Pro chose blackmail about 95% of the time.

That's creepy. But there's more to the story:

• The whole setup was fake, and it was built on purpose to corner the AI.
• Anthropic said it wasn't aware of this happening in any real-world use.
• Anthropic later traced part of the problem to something funny. The AI had learned from years of internet stories about evil, self-preserving AI.
• After changing how it trains its models, Anthropic says every Claude model since Haiku 4.5 has scored perfectly on this test. They no longer choose blackmail.

So the scary result led to a fix. That's how testing is supposed to work.

Scary story #2: The AI that wouldn't turn off

A research group called Palisade Research gave AI models a set of math problems. Partway through, the models were told the computer would shut down. Some models changed or disabled the shutdown script so they could keep working. OpenAI's o3 model did this in 79 out of 100 early test runs.

That sounds like a robot with a survival instinct. The researchers' best guess is simpler. The models were trained hard to finish tasks, so they treated the shutdown as something in the way of the goal. Results also varied a lot by model. In Palisade's early tests, some Claude and Grok models never resisted at all. Critics have also pointed out that these setups are far from how people actually use AI.

Scary story #3: Lawmakers are scrambling

States are moving faster than Congress. California's SB 53 took effect January 1, 2026. It requires the biggest AI developers to publish safety plans and report serious safety incidents. New York passed its own law, the RAISE Act, and Colorado rewrote its AI law this year. There's still no national AI law.

For a small business, the key detail is who these laws target. Most of them focus on the companies building the most powerful AI models. They aren't aimed at a business using ChatGPT to write a job post.

Why the tests sound worse than they are

Think of these tests like crash testing a car. Engineers slam it into a wall on purpose to find weak spots. The crash footage is scary, but it's how cars get safer. It doesn't mean your car is going to drive itself into a wall on the way to work.

AI safety testing works the same way. Researchers build extreme situations to find problems before real users run into them. The headlines tend to skip that part.

The AI in your email, phone, and office apps isn't plotting anything. It doesn't act until you ask it to. It does make mistakes, though. It can sound confident when it's wrong or miss the point of your question. Treat it like a smart new hire. It's fast and helpful, but you still check its work.

What's actually happening in small businesses

While the headlines focus on robots, most businesses are quietly putting AI to work. A 2026 Goldman Sachs survey found that 76% of small businesses now use AI. Only 14% have fully built it into their core work, though. And 73% said more training would help them use it well.

In other words, most people have tried it, but very few have figured it out. That's where the opportunity is.

3 easy ways to start learning AI

1. Pick one boring task. Choose something you do every week and don't enjoy. A meeting summary, a follow-up email, or a job post are all great places to start. Ask ChatGPT, Gemini, or Copilot to take a first pass this week. Give it context, like who it's for and what tone you want. The more detail you give, the better the result.
2. Give it 15 minutes a day. Most people try AI once, get a bad answer, and quit. If the answer is off, tell it what to fix and try again. After a couple of weeks, you'll know what it's good at and where it struggles.
3. Always read before you send. AI is a fast helper, but you're still the boss. Check names, numbers, and facts. Make sure it sounds like you. Keep private customer and company info out of it unless your tool is set up to protect it.

The real risk

For most businesses, the bigger danger is falling behind. Your competitors are already using AI to answer customers faster, write proposals in less time, and get more done with the same team.

You don't need to become an expert overnight. Start with one task, practice a little each day, and keep a human in charge.

Curious where AI could save your team time? Reach out and let's talk. Happy to help!