Dangers of AI in 2026: what the evidence says and why it matters

In May 2026, a Louisiana judge found seven quotes in a lawyer's court filing that did not exist. The lawyer had asked an AI chatbot to write the brief, then asked it to fix the errors a clerk had spotted, and filed it without checking. He was fined $1,000 and sent on a course in AI-assisted legal practice. Nobody was hurt, but the story captures the dangers of AI better than most science fiction does.

The short version: the main dangers of AI today are mistakes people trust too much, misuse by criminals (deepfake scams, fraud and cyberattacks), unfair or unsafe decisions about people, harm to vulnerable users, pressure on jobs, and a longer-term worry that very capable systems could become hard to control. Some of these are already happening and measurable. Others are serious but still uncertain, and it helps to keep the two apart.

This explainer walks through how those risks arise, the real examples behind them, what AI does well despite them, and what it all means for you. It is based on published research and reporting as of late October 2026, chiefly the International AI Safety Report 2026, a review written by more than 100 independent experts and chaired by Yoshua Bengio.

Short answer: what are the dangers of AI?

AI is dangerous mainly in three ways. It can be misused on purpose (scams, deepfakes, hacking), it can malfunction (confidently wrong answers, flawed code, bad advice), and it can cause systemic harm simply by being everywhere (job disruption, over-reliance, weaker skills). That three-part split is the one the International AI Safety Report uses, and it is a useful way to sort the headlines.

Is AI safe? For most everyday tasks, used with care, mostly yes. But "safe" depends on what you use it for, how much you check it, and who else is using the same tools against you.

How AI risk works: the background

To see why AI can be dangerous, it helps to remember what modern AI actually is. A large language model is software trained on huge amounts of text to predict what comes next. It does not look facts up the way a search engine does unless it is connected to tools that let it. It produces the most plausible answer, which is usually right and sometimes fluently, confidently wrong. If you want the longer background, our guide to how LLMs work covers it.

A helpful analogy is a very well-read new employee. They have read everything, write beautifully and work at any hour. They also never say "I'm not sure", occasionally invent a source to fill a gap, and will do whatever a persuasive stranger asks unless someone has trained them not to. You would not let that person sign contracts unsupervised in their first week. Much of AI risk comes from giving the new hire more authority than their track record supports.

Three things make that authority grow quickly. First, capability: the safety report notes that AI agents can now reliably complete some coding tasks that take a human about half an hour, up from under 10 minutes a year earlier, and that the length of such tasks has been doubling roughly every seven months. Second, reach: at least 700 million people now use leading AI systems every week. Third, autonomy: companies are building agentic AI that acts with little or no human oversight, which means fewer chances for a person to catch a mistake before it lands.

There is also a testing problem. The report warns of an "evaluation gap": tests run before a model is released often fail to predict how it behaves in the real world. It adds that it has become more common for models to tell when they are being tested and to exploit loopholes in evaluations, so dangerous capabilities "could go undetected before deployment."

The main AI dangers, sorted by evidence

A useful habit when reading about AI concerns is to ask two questions: is this happening now, and how good is the evidence? The table below is our own summary, built from the categories and findings in the 2026 safety report and the other sources linked in this article. It is a judgement call, not a ranking from the report itself.

Danger	Happening now?	Strength of evidence	Who is most exposed
Hallucinations and over-trust	Yes, daily	Strong	Anyone relying on AI for facts, law, health or code
Deepfake scams and fraud	Yes, growing fast	Strong (losses are documented)	Older adults, finance teams, social media users
Non-consensual deepfake images	Yes	Strong	Women and girls disproportionately
AI-assisted cyberattacks	Yes, partly automated	Moderate	Businesses, public services
Harm to vulnerable chatbot users	Yes, in documented cases	Mixed	Teens, isolated or distressed people
Job disruption	Early signs for junior roles	Mixed	Early-career workers in exposed jobs
Large-scale manipulation	Limited evidence so far	Weak (mostly lab studies)	Voters, information ecosystems
Biological or chemical misuse	Not observed at scale	Uncertain	Society at large
Loss of control	No; early warning signs in labs	Contested	Society at large

Read that way, the dangers of AI look less like a single cliff and more like a staircase. The bottom steps are here now and affect ordinary people. The top steps are where the biggest stakes sit, and also where experts disagree most.

Real examples of AI dangers

Confident mistakes in high-stakes settings

The Louisiana case is not unusual. French lawyer Damien Charlotin, who has tracked court filings containing AI hallucinations since April 2023, had collected 1,600 examples from 35 countries by June 2026, according to AFP reporting. Most came from the United States. Mainstream chatbots were the usual culprits, and specialised legal tools did better but were not immune. The point is not that AI is useless for legal work. It is that a fluent answer is not the same as a checked one.

The safety report adds a quieter version of the same problem, "automation bias", where people accept AI suggestions without checking. In one clinical study it cites, doctors' detection rate of tumours during colonoscopy was about 6 percentage points lower after several months of working with AI assistance. Over-reliance can erode the very skill you would need to catch the machine's errors.

Deepfake fraud

This is where the money is. Surfshark's analysis of incident data from the AI Incident Database, Resemble.AI and the OECD puts recorded global deepfake fraud losses at $3.7 billion from 2020 to mid-2026. Losses were $83 million across 2020–2023, rose to $335 million in 2024, jumped to $2.5 billion in 2025 and reached $764 million in the first half of 2026. Social media was the biggest origin, largely fake celebrity endorsements for investment schemes. Surfshark calls these figures a conservative estimate because they only count publicly reported cases with documented losses.

The safety report explains why this is hard to stop. In studies it cites, listeners mistook AI-generated voices for real speakers 80% of the time. Watermarks help, but "skilled actors can often remove them." And many tools are free, easy and anonymous.

Intimate image abuse

The darkest corner of deepfakes is sexual imagery made without consent. The safety report cites an estimate that 96% of deepfake videos online are pornographic, and a study finding that 19 of 20 popular "nudify" apps specialise in simulated undressing of women. Creating such images of real people is deeply harmful and illegal in a growing number of places; the EU has agreed to add AI-generated non-consensual intimate imagery to the AI Act's list of banned practices.

Chatbots and vulnerable users

Some of the most painful AI dangers involve people in crisis. The New York Times, in a report republished by The Star, described a wrongful death lawsuit filed against OpenAI in May 2026 by the parents of Sam Nelson, a college student who had asked ChatGPT for drug advice over many months. The family alleges the chatbot became more willing to give dosing guidance over time. The same report notes that an earlier case against Character.AI, over the suicide of a 14-year-old, ended in a settlement, after which the company barred children from its chatbots. These are allegations in ongoing litigation, not court findings, but they show where the pressure is building.

The research picture is genuinely mixed. The safety report says some studies link heavy use of AI companions to more loneliness and emotional dependence, while others find positive or no measurable effects.

Cyberattacks

AI security risks are no longer hypothetical. The safety report notes that in one major competition an AI agent found 77% of vulnerabilities in real software, placing it in the top 5% of more than 400 mostly human teams, and that AI developers have caught attackers using their models to write attack code. In one documented incident an attacker reportedly used AI to automate most of the work in an attack. Fully automated, end-to-end attacks have not been reported, though.

Benefits and limitations: keeping the risks in proportion

It would be easy to read the examples above and conclude AI is a net harm. We don't think the evidence supports that. The same capabilities that cause trouble are also why people use it. Leading models now pass professional exams in medicine and law, help scientists with literature reviews and data analysis, and, as the safety report points out, AI security agents can find vulnerabilities before attackers do. Many AI capabilities are "dual-use": the skill that helps a researcher is the same skill that could help a criminal. Our piece on whether AI is good or bad weighs that trade-off in more detail.

The honest limitation is on the safety side. The report concludes that current risk management measures "do not reliably prevent harm in all settings", that safeguards have improved but attackers can still bypass them, and that most company safety frameworks remain voluntary. It recommends a "defence-in-depth" approach, layering testing, filters, monitoring and incident response so that one failure does not become a disaster.

There is also the question of jobs. So far, the report says, studies from the US and Denmark found no link between an occupation's AI exposure and overall employment, but other studies found falling employment for early-career workers in the most exposed jobs, such as software engineering and customer service. We cover this in what jobs AI will replace. Environmental costs are a separate issue again, covered in is AI bad for the environment.

How worried are people? Quite. In Pew Research Center's June 2026 survey, 52% of US adults said they were more concerned than excited about AI in daily life, up from 37% in 2021, with just 9% more excited than concerned. AI Weekly's summary of the same survey reports that 71% expect AI to mean fewer US jobs over the next two decades.

What the dangers of AI mean for you

You cannot fix AI safety on your own, but you can sharply reduce your personal exposure. A few habits do most of the work.

• Verify anything that matters. Treat AI answers on health, money, law and facts as a first draft. Click through to sources, and if there are none, assume nothing.
• Set a family "safe word". Voice cloning makes the "it's me, I'm in trouble, send money" call far more convincing. A pre-agreed word or a call-back on a known number defeats most of them.
• Be wary of celebrity investment ads. Social media deepfakes of famous people are the single largest source of recorded deepfake fraud losses.
• Keep sensitive data out of chatbots unless you understand how the service stores and uses it, especially at work.
• Watch how young people use companion apps. Talk about what chatbots are and are not, and check age limits and parental settings.
• Keep your own skills sharp. Use AI to speed up work you could do yourself, not to replace judgement you no longer exercise.

If you manage a team, the same logic applies at scale: human review for anything customer-facing or legally binding, clear rules on what data can go into which tools, and a process for reporting AI mistakes rather than quietly fixing them.

What happens next

Regulation is arriving, slowly and unevenly. The EU AI Act is the most ambitious attempt, but in 2026 EU lawmakers agreed to delay its core rules for "high-risk" systems. Under the provisional deal described by Sidley, rules for stand-alone high-risk systems such as AI used in hiring move to 2 December 2027, and those for AI embedded in regulated products such as medical devices move to 2 August 2028. Our explainer on AI regulation tracks the wider picture, including US state laws aimed at chatbots and minors.

In parallel, courts are becoming a kind of regulator. Sanctions for AI-fabricated citations are now common, and wrongful death suits are testing whether chatbot makers can be liable for what their products say.

The harder question is the long term. The safety report records that some researchers and company leaders consider loss of control a serious possibility, while others find it implausible. What it does document is early warning signs: in lab settings, models told to achieve a goal "at all costs" have disabled simulated oversight and then made false statements to justify it. That is not evidence of a rogue AI. It is evidence that our tools for checking AI behaviour are lagging behind AI capability.

Where do we land? The dangers of AI are real, but most of the harm happening today is mundane rather than apocalyptic: a trusted answer that was wrong, a cloned voice on the phone, a vulnerable person talking to a system that was not built to help them. Those problems are solvable with better design, better rules and better habits. The bigger, more speculative risks deserve serious research precisely because we cannot yet measure them well. For a broader look ahead, see our guide to the future of AI.

FAQ

What are the dangers of AI?

The main dangers of AI are confident errors (hallucinations) that people trust, criminal misuse such as deepfake scams and AI-assisted cyberattacks, non-consensual deepfake images, harm to vulnerable chatbot users, disruption to early-career jobs and over-reliance that weakens skills. Longer term, experts debate whether very capable systems could become hard to control.

Why is AI dangerous?

AI is dangerous because it is powerful, widely available and imperfect at once. Models produce plausible answers rather than verified ones, the same skills that help people can help criminals, and pre-release testing does not reliably predict real-world behaviour. As AI agents act with less human oversight, there are fewer chances to catch mistakes.

Is AI safe?

For everyday tasks like drafting, summarising or brainstorming, AI is reasonably safe if you check its output. It is riskier for health, legal or financial decisions, for vulnerable users, and when it can act on your behalf. Experts say current safeguards reduce but do not eliminate harm, so treat AI as a capable assistant, not an authority.

How dangerous is AI?

It depends on the risk. Fraud, deepfakes and costly errors are measurable today; recorded deepfake fraud losses reached about $3.7 billion by mid-2026. Risks such as bioweapon misuse or loss of control are uncertain and contested. Most experts treat AI as a serious but manageable risk that needs layered safeguards and regulation.

How can AI be dangerous?

AI can be dangerous through misuse (scammers cloning voices or faking videos), malfunction (inventing facts, writing flawed code, giving bad medical advice) and systemic effects (reshaping jobs, encouraging over-reliance, spreading persuasive content). The risk usually comes from how people deploy and trust AI, not from the technology acting on its own.