Risks that are concrete and present today
Confidently wrong information. AI systems generate plausible text, not verified facts. People acting on incorrect AI output — medical, legal, financial — is a real, current harm, not a hypothetical one.
Privacy of what you type. Anything you paste into a third-party AI tool goes to that company's servers. Whether it's used for training, how long it's kept, and who can see it varies by provider and plan. Treat this like any other cloud service: read the terms before pasting anything sensitive.
Bias reflected from training data. Models learn patterns from data produced by people, including the skewed and unfair patterns. This shows up in real outputs, and matters most when AI is used in consequential decisions.
Convincing fake content. Generated text, images, audio, and video are now good enough to deceive people. This is already being used for fraud and misinformation.
Over-reliance. Using AI in place of developing your own judgment or skills has costs that only show up later.
Risks that are genuinely debated
Longer-term questions — about highly capable autonomous systems, or large-scale economic disruption — are taken seriously by serious researchers, and are also genuinely uncertain. Reasonable, well-informed people disagree substantially about both the likelihood and the timeline.
The honest position is that these are open questions, not settled facts in either direction. Anyone presenting them as certain — in either direction — is going beyond the evidence.
What's mostly hype
Framing that treats current AI systems as conscious, self-aware, or possessing intentions. Current language models predict likely text. That's a genuinely powerful capability, and it isn't the same thing as understanding, wanting, or planning.
Confusing these makes the real risks harder to see clearly, because attention goes to the wrong things.
What you can practically do
- Verify anything important. Especially facts, numbers, and citations.
- Don't paste sensitive data into tools you haven't checked the terms for.
- Be skeptical of media that provokes a strong reaction — that's exactly what generated fakes are built to do.
- Keep developing your own skills rather than fully outsourcing them.
None of this requires alarm. It requires the same ordinary caution you'd apply to any powerful tool.
Related reading
AI Hallucinations Explained covers the confidently-wrong problem in more depth, and How Does AI Actually Work? covers why prediction-based systems behave this way.
Risks worth your attention today
The public conversation splits between science-fiction scenarios and dismissal, and the genuinely important risks sit in between — unglamorous, present, and affecting people now.
Confident wrong answers at scale. The most common real harm is mundane: someone acts on a fabricated fact. It has produced invented legal citations filed in court, incorrect medical information, and business decisions built on numbers no source supports.
Automated decisions nobody can explain. Systems used in hiring, lending, and benefits can produce outcomes with no available reason. When the decision affects someone's life, "the model said so" is not an answer.
Bias learned from data. Models reproduce patterns in their training data, including discriminatory ones. This is not a bug to be patched; it is a property of learning from a record of how things have been.
Privacy through carelessness. Pasting confidential material into a consumer AI tool is now one of the most common ways sensitive information leaks out of organizations, usually by well-meaning people who did not think of it as sending data anywhere.
Convincing synthetic media. Voice and video generation is good enough for fraud, and the most frequent version is not political — it is a phone call from a familiar voice asking for money.
Where the hype outruns the evidence
Claims that a chatbot is conscious, that general intelligence is imminent, or that current systems have goals of their own are not supported by how these systems work. A model predicting likely text is doing something genuinely impressive and genuinely different from wanting anything.
Sweeping predictions about mass unemployment deserve similar caution. The effects on work are real but uneven, and confident timelines have a poor track record. The opposite claim — that none of this matters — is equally unsupported.
Both extremes have the same practical cost: they distract from the risks that are actually affecting people this year.
What you can control
Verify anything factual before acting on it. Never paste confidential material into a tool your employer has not approved. Be skeptical of urgent requests arriving by voice or video, and confirm through another channel. Assume public tools may retain what you send unless their terms clearly say otherwise.
None of that requires expertise — just habits. How to Fact-Check AI Output covers the verification side, and Using AI in Everyday Work includes a full lesson on what should never go into these tools.