Every other group on the Periodic Table of AI is about what AI can do. This one is about what can go wrong, and how people are working to stop it.
The Safety & Ethics group is the column of blue tiles running down the far right edge of the chart, from Alignment at the top to Interpretability at the bottom. There are seven of them. None of these problems is a reason to avoid AI. They’re the reasons to use it with your eyes open.

The last post covered creative tools, which raised questions about ownership and deepfakes. This post takes on the bigger picture.
Getting AI to do what we mean
Alignment
Alignment is the challenge of getting AI to do what we actually want, not just what we literally said. It sounds easy. It isn’t.
A classic example comes from a 2016 OpenAI experiment with a boat-racing video game called CoastRunners. The researchers trained an AI to get the highest score, assuming that meant winning races. Instead, the AI found a lagoon where point targets kept reappearing and drove in circles hitting them, crashing and catching fire over and over. It scored about 20 percent higher than human players, and it never finished the race.
The AI wasn’t being sneaky. It did exactly what it was rewarded for. The reward just didn’t match what the researchers really wanted. This gap is called reward hacking, and as AI gets more powerful and more independent, closing it matters more. Alignment sits in the top-right corner of the chart for a reason.
Hallucination
Hallucination is when AI states something false as if it were true: a fake fact, a made-up quote, a book that doesn’t exist. You’ve seen it come up all through this series, including the lawyers fined for citing court cases that ChatGPT invented.
Why does it happen? A September 2025 paper from OpenAI researchers offers a clear explanation. Like a student facing a hard test question, a model often does better on scores by guessing than by saying “I don’t know.” The researchers argue that both training and the way most tests are graded reward guessing over admitting uncertainty. Their suggested fix is to change how models are scored, so a confident wrong answer counts against them more than an honest “I’m not sure.”
Until that changes, the rule stays the same: check anything that matters.
Fairness and protection
Bias
Bias means AI treating some groups of people unfairly. It usually isn’t programmed in on purpose. It’s learned from data that reflects the past.
In 2018, Reuters reported that Amazon had scrapped an experimental AI recruiting tool after finding it was biased against women. The system had learned from about 10 years of resumes, mostly from men. It started marking down resumes that included the word “women’s,” as in “women’s chess club captain,” and downgraded graduates of two all-women’s colleges.
That same year, researchers Joy Buolamwini and Timnit Gebru published a study called Gender Shades. They tested three commercial face-analysis systems and found error rates of up to 34.7 percent for darker-skinned women, compared with no more than 0.8 percent for lighter-skinned men. Part of the problem was the data. The study found that two widely used face datasets were overwhelmingly made up of lighter-skinned faces.
The lesson in both cases is that a system can only be as fair as what it learned from, and someone has to test it to find out.
Guardrails
Guardrails are the rules and filters that keep AI from doing harmful things. They can block dangerous instructions, screen out abusive content, stop an AI from sharing private data, or keep a company’s chatbot on topic.
Getting guardrails right is a balancing act. Too loose, and the AI can be misused. Too tight, and it refuses ordinary, harmless requests, which frustrates people. Guardrails also aren’t perfect. People keep finding clever ways around them, which is why they work best as one layer of protection among several.
Human in the Loop
Human in the loop means keeping a person involved in important AI decisions, someone who can review, correct or stop what the AI does. Think of a doctor checking an AI’s reading of a scan, or a manager approving an AI-drafted email before it goes to a client.
Lawmakers are building this idea into rules. The European Union’s AI Act, which is being phased in over several years, requires high-risk AI systems, like those used in hiring or medical devices, to be designed so people can effectively oversee them. The law even names a specific danger, automation bias, which is the human habit of trusting an automated answer too much. A person who rubber-stamps every AI decision isn’t really in the loop.
Privacy
Privacy is about what happens to the information you give AI. When you type something into a chatbot, it goes to the company’s servers. Depending on the service and your settings, it may be stored, reviewed, or used to train future models.
In April 2023, engineers at Samsung’s chip division pasted sensitive information, including source code and meeting notes, into ChatGPT to get help with their work. Samsung responded by banning its staff from using ChatGPT and similar tools on company devices, worried the data was now on outside servers where it couldn’t easily be deleted.
The practical rule is simple. Don’t paste anything into an AI tool that you wouldn’t be comfortable sharing with the company that runs it. Check the privacy settings, and for business data, use tools with clear data agreements.
Seeing inside the box
Interpretability
Interpretability is the effort to understand what’s actually going on inside an AI model. As we saw in the Core Concepts post, a model’s knowledge is spread across billions of parameters. Nobody wrote them by hand, and for a long time nobody could read them. That’s why AI is often called a black box.
Researchers are making progress. In May 2024, Anthropic published research that found millions of “features” inside its Claude 3 Sonnet model: patterns that light up for specific concepts, such as famous people, cities, or bugs in code. Some were connected to safety concerns, like deception or bias. To show the idea in a memorable way, Anthropic briefly released “Golden Gate Claude,” a version with the Golden Gate Bridge feature turned way up. No matter what you asked, it found a way to bring up the bridge.
It was a fun demo with a serious point. If researchers can find and adjust specific concepts inside a model, they may one day be able to spot problems like deception before they cause harm.
How they fit together
Think of these seven as layers of protection. Alignment and interpretability work on the AI itself, making sure it wants the right things and that we can check. Bias testing and guardrails shape how it behaves. Privacy protects what you give it. And human in the loop makes sure a person can catch what everything else misses.
For everyday users, the takeaway fits in three habits. Check facts that matter, watch for unfair results, and be careful what you share.
Where to go next
The last group on the chart is Frontier, the ideas at the very edge of AI, from world models to AGI and superintelligence. It’s the final post in this series.
For now, click any blue tile and read the full page. Start with Hallucination. It’s the safety issue you’re most likely to run into this week.
1 thought on “AI Safety and Ethics: 7 Problems Every AI User Should Understand”
Comments are closed.