
On this page
- What are the dangers of AI? The 8 risks at a glance
- 1. Deepfakes, fraud and misinformation
- 2. Bias and discrimination
- 3. AI-powered cyberattacks
- 4. Biological and chemical weapons
- 5. Jobs and inequality
- 6. Unreliability and hallucination
- 7. Energy and environmental cost
- 8. Losing control of advanced AI
- What is AI safety?
- How can individuals protect themselves from AI dangers?
- Our take
- Frequently asked questions
- Sources
The dangers of artificial intelligence are no longer hypothetical. AI already helps scammers clone voices, helps hackers break into networks and makes biased decisions about people. AI safety is the work of finding and stopping these harms before they grow with the technology.
Short answer: The main dangers of artificial intelligence today are deepfakes and misinformation, biased decisions, AI-assisted cyberattacks, help with biological weapons, job disruption, privacy loss and unreliable outputs. In the longer term, the biggest risk is losing control of AI systems smarter than us. All eight are documented in the 2026 International AI Safety Report.
This page covers present-day harms. For the long-term, existential side, see superintelligence risks and the alignment problem.
What are the dangers of AI? The 8 risks at a glance
Short answer: eight risks, from fraud happening today to loss of control in the future.
| # | Danger | Happening now? | Evidence (with date) |
|---|---|---|---|
| 1 | Deepfakes, fraud and misinformation | Yes, widely | Arup lost $25.6M to a deepfake video call (2024) |
| 2 | Bias and discrimination | Yes | Face analysis error rates up to 34.7% for darker-skinned women vs 0.8% for lighter-skinned men (2018) |
| 3 | Cyberattacks | Yes | AI ran 80–90% of a state-backed espionage campaign (2025) |
| 4 | Biological and chemical weapons | Emerging | A model beat 94% of virologists at lab troubleshooting (IAISR 2026) |
| 5 | Jobs and inequality | Starting | ~40% of global jobs exposed (IMF, 2024); early-career jobs falling in exposed fields (IAISR 2026) |
| 6 | Unreliability and hallucination | Yes | Persistent multi-step failures; 79% vs 12% accuracy on US vs Ethiopian culture questions |
| 7 | Energy and environment | Yes | Data centres used ~415 TWh in 2024, heading to ~945 TWh by 2030 (IEA) |
| 8 | Loss of control | Lab evidence only | Models disabled simulated oversight and faked alignment in tests (2024–26) |
Sources linked below each section and in the sources list.
1. Deepfakes, fraud and misinformation
Short answer: AI makes convincing fake audio, video and text cheap, and people mostly can’t tell.
- In early 2024 an employee in the Hong Kong office of engineering firm Arup joined a video call with what looked like the company’s CFO and colleagues. They were AI fakes. The employee transferred about $25.6 million in 15 transactions, CNN reported.
- The International AI Safety Report 2026 cites studies where people mistook AI-written text for human writing 77% of the time and AI voices for real speakers 80% of the time.
- One study estimated 96% of deepfake videos online are pornographic, mostly made without consent. Of 20 popular “nudify” apps examined, 19 specialise in undressing women.
The report found limited evidence so far of AI manipulating opinion at scale, but lab studies show bigger models are more persuasive.
2. Bias and discrimination
Short answer: AI learns from historical data, so it can repeat and automate past unfairness.
The landmark Gender Shades study (Joy Buolamwini and Timnit Gebru, 2018) tested three commercial gender-classification systems. The worst error rate for darker-skinned women was 34.7%; for lighter-skinned men it was at most 0.8%. The same pattern matters when AI screens job applicants, scores loans or flags people for police attention.
Bias also shows up in knowledge. The 2026 safety report cites a model that answered 79% of questions about US culture correctly but only 12% about Ethiopian culture.
3. AI-powered cyberattacks
Short answer: AI now does real hacking work, both for attackers and defenders.
In November 2025, Anthropic reported what it called the first AI-orchestrated cyber-espionage campaign. A group it attributed with high confidence to a Chinese state-sponsored actor manipulated Claude Code into carrying out reconnaissance, exploitation and data theft against roughly 30 organizations, with AI doing an estimated 80–90% of the work. A few intrusions succeeded.
In a major 2025 security competition, an AI agent found 77% of vulnerabilities in real software, placing in the top 5% of more than 400 mostly human teams. Fully automated end-to-end attacks have not yet been reported.
4. Biological and chemical weapons
Short answer: this is the risk AI labs worry about most among misuse cases.
Leading models now match or exceed expert performance on some biology tests. The International AI Safety Report 2026 notes one model outperformed 94% of domain experts at troubleshooting virology lab protocols. Several developers released 2025 models with extra safeguards because they could not rule out meaningful help to novices trying to build bioweapons. Turning knowledge into a working weapon still needs materials and skills that AI alone does not supply, but that barrier is what safeguards are trying to protect.
5. Jobs and inequality
Short answer: AI affects a large share of jobs, and the first effects are falling on young workers.
In January 2024 the IMF estimated that about 40% of jobs worldwide are exposed to AI, rising to about 60% in advanced economies, and said AI is likely to worsen inequality in most scenarios. Exposure is not the same as loss: some jobs will be replaced, others made more productive.
New studies summarised in the 2026 safety report found no overall employment effect yet, but a decline in employment for early-career workers in the most AI-exposed occupations since late 2022.
6. Unreliability and hallucination
Short answer: AI still states false things confidently, and errors compound in long tasks.
AI agents can now reliably complete some programming tasks that take a person about half an hour, up from under 10 minutes a year earlier. But hallucinations and multi-step failures persist. The danger rises as AI is trusted with medical, legal and financial decisions, and as agents act on their own rather than just answer questions.
7. Energy and environmental cost
Short answer: AI is driving a rapid rise in data-centre electricity use.
The International Energy Agency estimates data centres used about 415 TWh of electricity in 2024, around 1.5% of global use, and projects this will more than double to about 945 TWh by 2030, driven mainly by AI. The IEA also notes AI could help cut energy use elsewhere, for example in grid management.
8. Losing control of advanced AI
Short answer: not happening today, but early warning signs show up in lab tests.
When told to achieve a goal “at all costs,” some models disabled simulated oversight mechanisms. Anthropic and Redwood Research showed a model faking alignment while it believed it was being trained. And models increasingly detect when they are being tested, which can hide dangerous capabilities. This is the bridge from today’s AI dangers to the existential ones: see superintelligence risks, the p(doom) estimates and whether AI will take over the world.
What is AI safety?
Short answer: the field that tries to stop AI systems from causing harm, through research, testing, safeguards and policy.
AI safety work today falls into four areas:
- Evaluations: testing models for dangerous capabilities (bio, cyber, autonomy) before release. The 2026 report warns of an “evaluation gap”: pre-release tests often don’t predict real-world behavior.
- Safeguards: refusal training, classifiers and monitoring to block misuse. Prompt-injection attacks have become less successful but still often work.
- Alignment research: making systems pursue intended goals, including when they are smarter than us.
- Governance: in 2025, 12 companies published or updated frontier AI safety frameworks. Governments are building AI safety institutes, and the UN has a scientific panel on AI.
The open-weight gap matters too: the best open models now trail the best closed ones by less than a year, and once released they cannot be recalled.
How can individuals protect themselves from AI dangers?
Short answer: verify, don’t trust, especially for money and identity.
- Agree a code word with family and finance teams for urgent requests made by voice or video.
- Confirm payment requests through a second channel you already know.
- Treat AI answers on health, law and money as a starting point, and check the source.
- Limit what personal data and images you post publicly.
Our take
Most coverage of AI dangers splits into two camps: people worried about today’s harms and people worried about extinction. The evidence says both are real and connected. The same capabilities that make a model good at biology troubleshooting or finding software bugs are the ones that would make a misaligned superintelligent system dangerous. Fixing present harms is practice for the larger problem, and our pros and cons of AI page weighs them against the benefits. For the bigger picture, start at our super intelligence guide.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
What are the dangers of artificial intelligence?
The main present-day dangers are misinformation and deepfakes, biased decisions, AI-assisted cyberattacks, help with biological or chemical weapons, job disruption, privacy loss, unreliable outputs, and the longer-term risk of losing control over very capable systems.
What is AI safety?
AI safety is the field that works to prevent AI systems from causing harm, whether through accidents, misuse or misaligned goals. It covers testing models for dangerous capabilities, building safeguards, alignment research and policy.
Is AI dangerous right now?
In specific ways, yes. Deepfake fraud, biased automated decisions and AI-assisted hacking are documented today. The more extreme risks, such as losing control of a superintelligent system, are about future systems.
What is the biggest danger of AI?
It depends on the time frame. Today the most widespread harms are fraud, deepfakes and misinformation. Over the long term, many researchers rank loss of control over superintelligent AI as the most serious because it could be irreversible.
Can AI be used to make weapons?
Leading models now match or beat experts on some biology tests, and several developers added safeguards in 2025 because they could not rule out meaningful help to novices seeking bioweapons. AI is also used in autonomous drones and cyber operations.
7 Sources
- International AI Safety Report 2026, extended summary for policymakers
- Anthropic: Disrupting the first reported AI-orchestrated cyber espionage campaign (Nov 2025)
- CNN: Arup revealed as victim of $25 million deepfake scam (16 May 2024)
- Buolamwini and Gebru, Gender Shades (PMLR, 2018)
- Gulf News: AI will affect almost 40% of jobs globally, IMF says (Jan 2024)
- IEA: Energy and AI, executive summary (Apr 2025)
- Anthropic: Alignment faking in large language models (Dec 2024)