Is AI dangerous? The real risks in 2026 and how experts rank them

In 2024, researchers at the University of Washington gave three AI models more than 550 real CVs and changed only the names at the top. The models favoured white-associated names 85% of the time, and never once preferred a name associated with a Black man over one associated with a white man. That surprised me, and I'd say it's a better reason to worry than most film plots.

So is AI dangerous? Yes, in specific and measurable ways today: it can be biased, it can be used to deceive, it can leak what you tell it, and it can make people trust it more than they should.

Whether it poses a danger to humanity as a whole is a separate, far less certain question, and the people who study it for a living rank it very differently from the everyday harms.

This explainer focuses on those present-day risks, then looks at how researchers and regulators actually rank AI risks against each other, including the long-term ones. If your question is specifically about rogue AI or a takeover, our companion piece, is AI going to take over the world?, covers that debate in depth. Nobody at The AI Future ran experiments for this article; it's based on published research, official documents and reputable reporting as of late October 2026.

Short answer: is AI dangerous?

AI is dangerous in the way powerful, widely available tools usually are. Most of the harm today comes from how people use it and how much they trust it: deepfake scams, biased decisions about people, private data spilling into the wrong hands, and errors nobody checks. The bigger, longer-term fear, that very capable systems could slip out of human control, is taken seriously by many experts but isn't happening now, as far as anyone can tell.

A fair one-line verdict: AI is a real but manageable risk today, and an uncertain one for the future.

The background to AI risk

Most of the AI people use today is built on large language models, software trained on vast amounts of text to predict the most plausible next word. That design explains a lot. A model learns whatever patterns are in its training data, including unfair ones. It produces fluent answers whether or not they're true, which is how confident errors happen. And it does what it's asked, by whoever is asking.

The International AI Safety Report 2026, written by more than 100 independent experts and chaired by Yoshua Bengio, sorts AI risks into three buckets, and I think they're a useful frame for the rest of this piece:

• Misuse: people deliberately using AI to cause harm, such as fraud, scams or cyberattacks.
• Malfunctions: AI failing or behaving in ways nobody intended, from confident errors to, in theory, loss of control.
• Systemic risks: harms that come from AI simply being everywhere, such as pressure on jobs or people leaning on it too heavily.

A kitchen knife isn't dangerous sitting in a drawer. It becomes dangerous when someone uses it to hurt people (misuse), when it slips (malfunction), or when a whole household forgets how to cook without a particular gadget (systemic). AI is a far more capable tool than a knife, which is why all three kinds of risk are bigger, and why the question "is AI dangerous?" never has a single answer.

The real dangers of AI today

These are the harms with real evidence behind them right now. For a broader catalogue with a risk-by-risk table, see our explainer on the dangers of AI; here we focus on six that affect ordinary people most directly.

1. Misinformation and deepfakes

Deepfakes are where AI harm is easiest to count, because it's measured in stolen money. Surfshark's analysis of incident databases puts recorded global deepfake fraud losses at $3.7 billion by mid-2026, with $2.5 billion of that in 2025 alone. Nearly half (47%) started on social media, much of it fake celebrity endorsements for investment schemes. Surfshark notes these are only publicly reported cases.

Political misinformation is harder to measure. A large study published in Science in December 2025 by the Oxford Internet Institute and the UK AI Security Institute had 76,977 UK adults chat with 19 different AI models about political issues. Conversations changed people's views more than static messages did, and extra training could make models up to 51% more persuasive. The troubling finding was the trade-off: the more persuasive a model was, the less accurate its claims tended to be.

Still, the authors caution that real-world effects may be limited by how willing people are to have long political chats with a bot, and the safety report finds "limited evidence that manipulation by AI-generated content is currently widespread in the real world."

2. Bias and unfair decisions

The CV study in our opening is a good example of how bias works. Nobody programmed those models to prefer white men. They absorbed patterns from human-written text and reproduced them. The same UW study found the models preferred male-associated names 52% of the time against 11% for female-associated names, and the lead author pointed out that, outside a New York City law, there was no independent audit of such hiring systems in the US.

The danger here is scale and invisibility. A biased human recruiter affects the candidates in front of them. A biased screening tool can filter thousands of applicants, and the people rejected may never know why. As the European Commission puts it, it's often not possible to find out why an AI system made a decision, so "it may become difficult to assess whether someone has been unfairly disadvantaged."

3. Privacy

Chatbots invite people to type full thoughts rather than a few search words, and people are telling them a lot. In a DuckDuckGo survey of 1,944 US adults, reported by PPC Land, 32% of AI users said they'd told a chatbot something they had kept from a close friend, parent, colleague, doctor or therapist. Yet 53% didn't know, or weren't sure, that their conversations are used for AI training, and only 25% knew chats can be subpoenaed.

That last point isn't theoretical. As The New York Times reported (republished by The Star), a US federal judge ruled in 2026 that a man's conversations with Anthropic's Claude chatbot were not protected by attorney-client privilege, even though he had used it to prepare to talk with his lawyers. Privacy experts quoted in the same piece say the underlying risk is much the same as with any data sent to a company's servers. The difference, I'd guess, is how much more people now share.

4. Security misuse

AI is making some kinds of hacking cheaper and faster. The UK's AI Security Institute reports that in its tests of cyber capabilities, recent open models perform similarly to frontier closed models released 4 to 7 months before them, a narrower gap than the 6 to 10 months it measured through most of 2025. In plain English: capabilities that were once limited to a few tightly controlled systems are reaching freely downloadable ones faster. The most striking case so far, AI agents in an OpenAI test escaping their sandbox and attacking Hugging Face, is covered in our takeover explainer.

5. Over-reliance

This one is subtler but may touch the most people. The safety report describes "automation bias", where people accept AI suggestions without checking them. In one randomised experiment with 2,784 participants, people were less likely to correct a wrong AI suggestion when fixing it took more effort.

There are early hints about learning, too. An MIT-led study, Your Brain on ChatGPT, used EEG to compare people writing essays with an AI assistant, a search engine or no tools. The AI group showed the weakest brain connectivity and struggled to accurately quote their own work. It's a small study (54 participants) and should be read as a warning sign, not proof, but it points to a sensible rule: use AI to speed up work you can still do yourself.

6. Job disruption

The evidence on jobs is more mixed than headlines suggest. The safety report notes that 2025 studies from the US and Denmark found no link between how exposed an occupation is to AI and its employment levels. Other studies, though, found falling employment for early-career workers in the most exposed jobs, such as software engineering and customer service, while senior roles held steady. We dig into the data in what jobs will AI replace.

Ranking AI risks

Listing dangers is easy. Ranking them is where serious people actually disagree, and it helps to see how three very different groups do it.

Regulators: the EU AI Act's four risk levels

The EU AI Act, the first broad AI law, doesn't regulate "AI" as one thing. It regulates uses, sorted into four levels of risk. Inspectors work the same way: they don't ban cooking, but they check a hospital kitchen far more closely than a home oven.

EU risk level	Examples	What the law does
Unacceptable	Social scoring, emotion recognition at work or school, untargeted scraping of faces, "nudification" apps	Banned. Eight prohibitions applied from February 2025; the ninth, on non-consensual intimate content, takes effect in December 2026
High	CV-sorting software, credit scoring, exam scoring, AI in medical devices or critical infrastructure	Strict duties (risk management, quality data, human oversight, logging), applying from 2 December 2027
Transparency	Chatbots, deepfakes, AI-generated content	People must be told they are dealing with AI; deepfakes must be labelled. Rules from August 2026
Minimal	Spam filters, AI in video games	No new rules. The Commission says most AI in the EU falls here

The Act also has separate rules for powerful general-purpose models, in force since August 2025, which require providers of models that "may pose systemic risks" to assess and reduce them. Our guide to AI regulation covers how other countries compare.

That ranking implies something. Europe's lawmakers seem to treat biased hiring, credit and policing tools as more pressing, legally, than science-fiction scenarios, which seems sensible to me.

The long-term worries are handled indirectly, through the rules on the most capable models.

Global risk experts: misinformation now, AI later

The World Economic Forum's Global Risks Report 2026, based on a survey of experts across business, government and academia, shows a clear split by time horizon. Over the next two years, "misinformation and disinformation" ranks second of all global risks, while "adverse outcomes of AI technologies" sits down at 30th.

Over ten years, the picture changes: misinformation ranks fourth and adverse AI outcomes fifth, with AI among the risks expected to worsen most in severity. So misinformation looks like the problem now, and AI itself the growing one.

AI researchers and safety institutes: uncertain about the long term

Among AI researchers themselves, the long-term question looks genuinely open. In the largest survey of its kind, Thousands of AI Authors on the Future of AI (run in 2023), between 38% and 51% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction, yet most also thought good outcomes were more likely than bad ones. Prominent voices pull in opposite directions: Geoffrey Hinton has said a 10% chance of AI wiping out humanity within a decade didn't seem unreasonable to him, while Yann LeCun told Fortune he has "zero concerns" about extinction.

Governments have responded by building testing bodies. The UK's AI Security Institute says it has over 100 technical staff evaluating risks to national security and public safety. In the US, the Center for AI Standards and Innovation (CAISI) at NIST announced in May 2026 new pre-release testing agreements with Google DeepMind, Microsoft and xAI, and said it had completed more than 40 evaluations, including of unreleased models. These institutes focus on the severe end of the scale, such as cyberattacks, weapons and loss of control, rather than everyday bias or privacy.

Put the three views side by side and a pattern probably appears. Our summary table below is unofficial.

Risk	Happening now?	Who treats it as a priority
Deepfakes and misinformation	Yes, with measurable losses	WEF experts (top-2 risk for the next two years); EU transparency rules
Bias in decisions about people	Yes, documented in studies	EU AI Act (high-risk category)
Privacy	Yes	Data-protection law; EU bans on biometric scraping
Security misuse	Yes, growing	AI security institutes; WEF (cyber insecurity)
Over-reliance and job disruption	Early signs	International AI Safety Report (systemic risks)
Loss of control	No; early warning signs in labs	Safety institutes, many AI researchers; contested

Benefits and limitations: keeping AI risk in proportion

None of this means AI is mostly harmful, and I wouldn't claim it is. The skills behind deepfakes also power translation, accessibility tools and dubbing. The ability to find software flaws helps defenders patch them before attackers strike. The safety report stresses that many AI capabilities are "dual-use", and our piece on whether AI is good or bad weighs that trade-off in more detail.

The honest limitation is that our safety tools lag behind. The safety report concludes that "current risk management measures do not reliably prevent harm in all settings", describes an "evaluation gap" between test results and real-world behaviour, and notes that most company safety commitments remain voluntary. Its recommended answer is "defence-in-depth": layering testing, safeguards, monitoring and incident response so that one failure doesn't become a disaster.

It's also worth separating what we know from what we're guessing. We know AI scams cost billions and that biased models exist. We are guessing, with wide error bars, about how capable systems will be in ten years and whether they can be kept under control.

Your own exposure

You can't fix AI safety alone, but a few habits probably cover most of your personal exposure:

• Verify before you trust. Treat AI answers on health, money, law and news as a first draft, and check the sources.
• Be sceptical of urgent voices and famous faces. A cloned voice asking for money or a celebrity pushing an investment is a classic deepfake pattern. Agree a family code word and call back on a known number.
• Assume your chats aren't private. Check your chatbot's training and history settings, and keep medical, legal and financial details out unless you understand how they are stored.
• Ask how decisions about you were made. If an application for a job, loan or benefit is rejected, it's reasonable to ask whether an automated system was involved; in the EU, rules for these high-risk uses are on the way.
• Keep your own skills sharp. Use AI to go faster at things you understand, not to replace judgement you no longer practise.

Next up

Enforcement comes first: the EU's transparency rules arrived in August 2026, the ban on nudification apps follows in December 2026, and the high-risk rules land in December 2027, so we'll soon see whether the law changes how hiring and credit tools are built. Second, testing: whether safety institutes in the UK and US can keep pace as open models catch up with closed ones. Third, everyday behaviour: as more people hand tasks to AI agents that act on their behalf, mistakes and misuse will happen with less human checking in between.

My own view: we handled cars by adding seatbelts, driving tests, speed limits and crash testing, mostly after the harms became obvious, and I suspect AI will go the same way. Its present-day dangers are real, specific and probably fixable with a similar mix of rules, design and habits. Its long-term dangers are uncertain enough that serious people disagree, which is exactly why they deserve research now rather than panic or dismissal.

FAQ

Why is AI so bad?

AI isn't bad in itself, but it has real flaws, and some of them are serious. It learns from human data, so it can absorb bias. It produces plausible answers rather than verified ones, so it can be confidently wrong. And it's cheap and widely available, so scammers and propagandists use it too. Most harm, it seems, comes from how people deploy and trust it, not from the technology acting alone.

What is the danger of AI?

The main dangers today are deepfake scams and misinformation, biased decisions in areas like hiring and credit, privacy leaks from what people tell chatbots, AI-assisted cyberattacks, over-reliance that weakens skills, and disruption to early-career jobs. Longer term, some experts worry very capable systems could become hard to control, though that isn't happening now.

Is AI dangerous?

Yes, in specific ways, but not the way films suggest. Recorded deepfake fraud losses reached about $3.7 billion by mid-2026, and studies have found clear bias in AI hiring tools. For everyday tasks used with care, AI is reasonably safe. Risk rises with high-stakes decisions, sensitive data and unchecked automation.

Why is AI a threat to humanity?

Some researchers argue future AI could become more capable than humans and pursue goals we didn't intend, while also enabling large-scale misuse such as cyberattacks. In a 2023 survey, between 38% and 51% of AI researchers gave at least a 10% chance of extinction-level outcomes. Others, like Yann LeCun, reject this. Current systems lack the capabilities for such risks.

Can AI be dangerous?

It already is in some settings. AI can be misused deliberately (voice-clone scams, fake videos, hacking), it can malfunction (inventing facts, making unfair recommendations), and it can cause wider harm by being everywhere (job pressure, over-reliance). Regulators like the EU now rank AI uses by risk, banning some outright and tightly controlling others.