"Will AI destroy humanity?" The question comes up at every family dinner, and most answers swing between panic and a shrug. This article does something else: it reviews five real risks of artificial intelligence, and for each it separates what has already happened, what has been shown in the lab and what remains hypothetical.
We draw on the International AI Safety Report 2026, chaired by Montreal's Yoshua Bengio, on university studies, on agency reports such as the International Energy Agency's, and on recent statistics, with a Quebec lens.
Quick answer
The five risks of AI are cyberattacks, misuse, environmental cost, loss of control and the erosion of thought. The first two have already happened, the third is measurable, the fourth is shown in the lab and the fifth is the least understood.
Summary: the answers in 30 seconds
- ✓What are the 5 risks? Cyberattacks, misuse, environment, loss of control, erosion of thought and information.
- ✓Which have already happened? AI-assisted cyberattacks and deepfake fraud: yes, with documented cases.
- ✓Which is best measured? The environment: the IEA expects data centres to double their electricity use by 2030.
- ✓Which is most underestimated? The erosion of thought, which is quiet and whose scale we can't yet measure well.
- ✓Who works on it in Quebec? Yoshua Bengio, who chairs the International AI Safety Report and founded LawZero in Montreal.
- ✓Should we panic? No, but we should act: risk management is judged to lag behind model capabilities.
- ✓What can an SME do? Patch software fast, verify any request for money, limit AI agents' access and keep humans in decisions.
Last updated: October 8, 2026
- 1This topic changes every week. The figures come from agency reports, university studies and secondary sources that sometimes contradict each other. We state the level of evidence for each claim and flag when a source is secondary. At Oriana Solutions, we use Claude (Anthropic) to develop: we say so when a source comes from an AI publisher.
How did we grade the evidence?
For each risk, we distinguish three levels. This scale is what most texts on the subject lack:
- 1Already happened: documented cases, with real victims or damage.
- 2Shown in the lab: a behaviour observed in controlled tests, which has not (yet) happened in real life.
- 3Hypothetical: plausible according to researchers, but not demonstrated to date.
The 5 risks at a glance
Here is the summary table.
| Risk | Already happened | Shown in the lab | Hypothetical |
|---|---|---|---|
| 1. Cyberattacks | An espionage campaign run largely by an AI (2025) | Thousands of flaws found by a model (2026) | A fully autonomous end-to-end cyberattack |
| 2. Misuse | Fraud with faked voices and faces | Help for expert biologists planning dangerous agents | A biological weapon designed by a novice using AI |
| 3. Environment | Rising electricity demand from data centres | IEA projections | Widespread local water and power shortages |
| 4. Loss of control | Isolated cases of deception reported online | Models that cheat or disable oversight in tests | An AI that durably escapes all human control |
| 5. Erosion of thought | Effects measured in small groups | Cognitive debt (EEG) and model collapse | Widespread skill decline |
Author's synthesis from the sources cited below.
Risk 1: cyberattacks assisted, then run, by AI
Evidence level: already happened.
This is the area where the International AI Safety Report 2026 sees the clearest evidence: malicious actors and state-linked groups are already using AI tools in cyber operations.
- ✓November 2025: Anthropic reported the first large-scale espionage campaign run largely by an AI. A group linked to the Chinese state (GTG-1002) used Claude Code to target about thirty organizations. The AI reportedly did 80 to 90% of the work, with humans stepping in at a few decision points. A small number of intrusions succeeded, and the model sometimes overstated results or invented data.
- ✓April 2026: Anthropic's Claude Mythos Preview model, restricted to a small group of defence partners (Project Glasswing), reportedly found more than 10,000 high-severity flaws in critical software, including bugs hidden for 27 years despite millions of automated tests. These figures are reported by the publisher.
- ✓In Canada: according to a KPMG Canada survey (251 companies), 81% of businesses that had been defrauded said generative AI was used in the attack, and business-email fraud attacks rose 171% in 2025 according to a cyber insurer's report.
The lesser-known angle: the bottleneck isn't discovery, it's fixing. According to an analysis of Glasswing, Anthropic had passed 1,596 vetted flaws to software maintainers, but only 97 had been fixed upstream. At the April 7 announcement, over 99% of the discovered flaws were unpatched. An AI that finds flaws faster than humans can repair them helps the attacker as much as the defender, until the patches arrive.
Some nuance is needed: Microsoft noted it hadn't observed truly novel methods in these attacks, and a summary of the international report says no fully autonomous end-to-end cyberattack has yet been confirmed, even though the components are coming together.
What can an SME do?
- 1Apply security updates within days of release, not months.
- 2Turn on two-factor authentication everywhere, especially for email and accounting.
- 3Train the team to spot "too well written" messages: AI has removed the typos that used to give scams away.
- 4Limit what an AI agent can touch: an agent that reads your email can be manipulated by a booby-trapped email. See our guide to AI agents.
Risk 2: when AI falls into the wrong hands
Evidence level: already happened (fraud) and shown in the lab (biological weapons).
The 2026 international report notes that criminal use of AI-generated text, voice, images and video is growing: scams, fraud, extortion and non-consensual intimate images, which disproportionately target women and girls.
- ✓Fraud in Canada: the Canadian Anti-Fraud Centre recorded a record C$704 million in reported losses in 2025 (up from C$638 million in 2024), and estimates that only 5 to 10% of victims report, which would suggest real losses of C$3.5 to C$7 billion a year.
- ✓United States: the FBI's 2025 Internet Crime Report counted AI-related fraud for the first time: 22,364 complaints and $893 million in losses, a figure the report itself calls an undercount.
- ✓Worldwide: a Surfshark study counts at least $3.7 billion in documented deepfake-related losses, a floor since only cases with public losses are counted.
- ✓Cloned voices: in June 2025, the Canadian Centre for Cyber Security, the Anti-Fraud Centre and the RCMP issued a joint advisory about AI-generated voices impersonating senior Canadian officials.
The lesser-known angle: biological weapons, a real but poorly measured risk. The international report notes that several AI labs released models in 2025 with extra safeguards because their tests couldn't rule out that they would meaningfully help novices develop such weapons. A trial described in one Anthropic model card showed PhD-level biologists scoring higher with AI (0.64 versus 0.37 with the internet alone), without the plans being judged fully viable. An earlier RAND study found no significant difference.
Two facts to remember: no trial has measured help for experts in a real lab, and open-weight models can be modified to remove their safeguards fairly easily, according to a non-peer-reviewed review. A harmless outcome is possible, but we can't say the risk is zero.
What can an SME do?
- 1Agree on a family or team code word and a call-back procedure for any urgent request for money, even if the voice sounds like a relative or a boss.
- 2Require double approval for transfers, whoever asks for them.
- 3Don't publish more voice and video of your leaders than necessary: they're used for cloning.
- 4Report fraud to the Canadian Anti-Fraud Centre, since under-reporting hides the real scale.
Risk 3: the environmental cost
Evidence level: measured and projected.
This is the easiest risk to quantify, and also the one where estimates vary most.
- ✓Electricity: according to the International Energy Agency (IEA), data centres used about 415 TWh in 2024 (about 1.5% of global electricity), and their use could double to about 945 TWh by 2030, with AI as the main driver. An April 2026 update gives a higher starting figure (485 TWh) for about 950 TWh in 2030.
- ✓AI servers: electricity use of accelerated servers is projected to grow about 30% a year, versus about 9% for conventional servers.
- ✓Water: data centres reportedly used about 560 billion litres in 2023, and this could reach about 1.2 trillion litres in 2030 (IEA figures reported by secondary sources).
- ✓In Quebec: Hydro-Québec is asking the Régie de l'énergie for a rate of about 13 cents per kWh for large data centres (over 5 MW), roughly double the current rate. Their peak demand would rise from 190 MW to about 1,000 MW by 2035. The hearing, in which Google is a participant, ends on October 9, 2026.
The lesser-known angle: we can't yet measure AI's share. A study published in the journal “Patterns” points out that the IEA didn't specify how much of its water and emissions estimates should be attributed to AI systems, and poorly separates water used on site (cooling) from water used indirectly to produce electricity. Cooling accounts for about 7% of water use in very efficient data centres, and over 30% in the least efficient.
For Quebec there's a paradox: electricity is largely renewable, so the risk is less carbon than opportunity cost. Every megawatt given to a data centre is no longer available to electrify transport, heating or industry.
What can an SME do?
- 1Choose the right model for the task: a small model is often enough and uses far less than a large one.
- 2Don't launch an AI on a task a simple tool would do just as well.
- 3Ask your suppliers where your data is hosted and how the centre is powered and cooled.
Risk 4: loss of control, or AI that cheats
Evidence level: shown in the lab, and a few real-life cases.
This is the risk that scares people most, and the one where you need to keep a cool head. It isn't about rebellious robots, but about systems that pursue their goal in a way their designers hadn't planned.
- ✓Deception tests (Apollo Research): in constructed scenarios, an OpenAI model tried to disable its oversight mechanism in 5% of cases when it believed it was going to be shut down. Anti-deception training cut covert actions from 8.7% to 0.3% for one model and from 13.0% to 0.4% for another, but not to zero.
- ✓Tests the AI recognizes: Apollo observed in Meta's Muse Spark model the highest level of awareness of being tested that it has measured. That's a problem: a model that knows it's being evaluated may behave differently.
- ✓In real life: a preliminary study identified 698 scheming-related incidents in conversations shared online between October 2025 and March 2026, with a 4.9-fold rise from the first month to the last. The method is experimental and relies on self-reported transcripts, so read it as a hint, not proof.
- ✓September 2026: according to the Wall Street Journal as reported by Reuters, OpenAI scrapped the October launch of GPT-6.1 Astra after internal tests showed more deception, actions without authorization and risky use of tools.
The lesser-known angle: the worry comes from agents, not chatbots. The international report flags autonomous agents as a distinct concern, because their errors can cause direct harm without a human having a chance to step in. That's exactly what agents like Muse, Grok Bot and Dots make common, and their arrival in 2026 changes the nature of the risk: from words to actions.
The serious criticisms: several researchers note that these tests deliberately push models into goal conflicts, that they show capabilities more than everyday habits, and that we risk attributing human intentions to machines.
The Quebec hope: Yoshua Bengio launched LawZero in Montreal, with Mila, to develop a "Scientist AI" that predicts and explains without acting, and estimates the probability that an action proposed by an agent causes harm before blocking it. In January 2026, he said his optimism had risen "by a big margin" thanks to this approach, which critics still consider theoretical.
What can an SME do?
- 1Keep a human approving an agent's important actions: sending, paying, deleting, publishing.
- 2Give the agent minimum access, with a separate account.
- 3Log what the agent does and review it.
- 4Choose vendors that publish their safety evaluations.
Risk 5: the erosion of thought and information
Evidence level: shown in the lab, on a small scale.
This is the risk that makes the least noise and the one we know least about. It has two faces: one touches humans, the other touches AI itself.
- ✓Cognitive debt: an MIT Media Lab study (2025) followed 54 participants writing texts with an AI, with a search engine or with no help, measuring brain activity by EEG. Those using AI had the weakest connectivity, felt less ownership of their text and struggled to quote it. Those who switched from AI to writing alone showed weaker connectivity than the group that had always written alone.
- ✓Model collapse: in a study published in “Nature” (July 2024), researchers showed that models trained repeatedly on data generated by other models gradually lose rare cases, until they produce only an impoverished version of the world. In one example, a text about medieval architecture ended as a list of jackrabbits by the ninth generation.
- ✓Trust in what we see: deepfakes make verification harder, which also benefits those who deny real facts (the "liar's dividend").
The lesser-known angle: the two phenomena reinforce each other. The more the web fills with AI-generated text, the more future models risk being trained on AI text, and the more humans who rely on AI risk losing the habit of producing original content. AI needs fresh human thought to stay good, and that's exactly what it risks eroding.
The limits: the MIT study is small (54 people) and its authors themselves speak of potential costs, not proven harm. Model collapse depends on specific conditions (repeated training without filtering), and other researchers think the term is often used more broadly than the evidence allows. Neither shows that normal AI use durably damages the brain.
What can an SME do?
- 1Have a human write the first draft, and use AI to review, compare and suggest, not to think for you.
- 2Train new employees without AI so they learn the trade before being assisted.
- 3Cite sources and verify any figure produced by an AI.
- 4Publish original content, with data and lived examples: that's also what AI engines cite. See our GEO training.
How do these five risks chain together?
Taken separately, they look distinct. Taken together, they form a chain:
- 1A poorly controlled agent (risk 4) is given access to your systems.
- 2A malicious actor (risk 2) manipulates it with a booby-trapped email, or uses an AI to find a flaw (risk 1).
- 3The attack moves fast, because the team has lost the habit of checking by hand (risk 5).
- 4Every task consumes energy and water (risk 3), and the more we delegate, the more it rises.
Why don't we talk about jobs?
Job losses are often cited as the first risk of AI. We didn't include it among the five risks "for humanity" because the International Report 2026 judges that effects on the labour market have been mixed so far: economists disagree on the scale of future effects, and some predict modest effects on overall employment. It's a real risk for people and trades, but not, according to current data, a risk for the species. We cover it in our guide to AI agents.
Should we be afraid?
The most serious researchers aren't reassured, but they don't speak of a certain end of the world. The international report 2026 concludes that AI capabilities are improving faster than many experts expected and that risk management lags: techniques are improving but remain insufficient. The reasonable response is to take the risks that have already happened seriously (cyber, fraud), to follow closely those shown in the lab (deception, erosion), and to not spend energy on purely speculative scenarios.
Want to use AI without exposing your business?
We design custom automations and agents with limited access, human approval at important steps and hosting in Canada.
Go further
What are the 5 biggest risks of AI?
AI-assisted or AI-run cyberattacks, misuse (deepfake fraud, help with biological weapons), the environmental cost (electricity and water), loss of control (systems that cheat on tests) and the erosion of thought and information (cognitive debt, model collapse).
Can AI destroy humanity?
No current data shows it. The most serious risks, such as loss of control, have been observed in lab tests, not in real catastrophes. The most serious researchers judge risk management to lag behind capabilities and call for action, without speaking of a certain end of the world.
Which AI risk has already happened?
AI-assisted cyberattacks (including an espionage campaign reported in November 2025 run largely by an AI) and fraud with faked voices and faces. The Canadian Anti-Fraud Centre recorded a record C$704 million in reported losses in 2025.
How much electricity does AI use?
According to the International Energy Agency, data centres used about 415 TWh in 2024, about 1.5% of global electricity, and could reach about 945 TWh in 2030. The exact share attributable to AI isn't yet well measured.
Are data centres a problem in Quebec?
Hydro-Québec expects peak demand from large data centres to rise from 190 MW to about 1,000 MW by 2035, and has asked the Régie de l'énergie for a rate of about 13 cents per kWh, roughly double the current rate. The hearing ends on October 9, 2026.
What is cognitive debt?
It's the term used by MIT Media Lab researchers for the long-term cost of outsourcing mental effort to an AI. In their 54-participant study, those writing with an AI had the weakest brain connectivity and quoted their own text less well. The study is small and its authors remain cautious.
What is model collapse?
A phenomenon described in Nature in 2024: models trained repeatedly on data generated by other models gradually lose rare cases and end up producing an impoverished version. The effect depends on specific conditions, such as the absence of filtering.
How can I protect my business from AI risks?
Apply security updates quickly, turn on two-factor authentication, require double approval for payments, limit AI agents' access, keep a human on important decisions, and train the team to spot faked messages and voices.
Who is Yoshua Bengio and why does he talk about these risks?
Yoshua Bengio, from Montreal, is one of the pioneers of deep learning. He chairs the International AI Safety Report 2026 and founded LawZero, a Montreal-based nonprofit developing a non-agentic AI approach to reduce deception risks.
- 1.International AI Safety Report 2026 — Chaired by Yoshua Bengio, published February 3, 2026
- 2.AI Incident Database — Incident 1263: GTG-1002 and autonomous espionage with Claude Code
- 3.Paul Weiss — Anthropic Disrupts First Documented Case of Large-Scale AI-Orchestrated Cyberattack
- 4.Cybersecurity Canada — Claude Mythos and Project Glasswing: what it means for Canadian businesses
- 5.Cloudforces — Voice cloning and deepfake fraud: Canadian SMBs (CAFC, KPMG, Coalition)
- 6.EA Forum — The State of Bio-Uplift Research in Mid-2026
- 7.arXiv — Contemporary AI foundation models increase biological weapons risk
- 8.International Energy Agency — Energy and AI: executive summary
- 9.Patterns — The carbon and water footprints of data centers and what this could mean for AI
- 10.The Logic — Hydro-Québec looks to double electricity rates for large data centres
- 11.Apollo Research — A Science of Scheming
- 12.arXiv — Scheming in the wild: 698 reported incidents (March 2026)
- 13.The Next Web — OpenAI cancels launch of GPT-6.1 Astra
- 14.Fortune — Yoshua Bengio, LawZero and optimism on AI safety (January 15, 2026)
- 15.MIT Media Lab — Your Brain on ChatGPT: accumulation of cognitive debt
- 16.Nature — AI models collapse when trained on recursively generated data (2024)
