AI Safety and Risks Explained: The Debate, Plainly (2026)

By the WahLiao desk · Last verified 3 October 2026

Is Super Intelligence, the kind of AI behind ChatGPT, Claude and Gemini, dangerous? Here’s the fairest answer we can give in 2026: some harms are already here and measurable, and a smaller set of much bigger risks is being seriously debated. Everyday problems such as confident errors, bias, scams, privacy leaks and deepfakes are well documented. The frontier debate, over misuse for biological or cyber attacks and whether future systems could slip out of human control, splits respected experts. The International AI Safety Report 2026, chaired by Yoshua Bengio, written by over 100 experts and backed by more than 30 countries and international organisations, found that current systems “lack the capabilities” to pose loss-of-control risks, while noting that their autonomous abilities are improving.

If you take away one idea, make it this: worry in proportion. Guard carefully against the harms that could reach you this week (a cloned voice asking for money, a wrong answer you act on), and follow the bigger debate through named, accountable sources rather than viral clips. We’re not here to tell you how worried to be; we’ve tried to set out who says what, so you can weigh it yourself. If we’ve got something wrong, tell us and we’ll fix it.

AI safety and risk: quick facts

Main expert review	International AI Safety Report 2026, published 3 February 2026, chaired by Yoshua Bengio
Three risk buckets	Malicious use, malfunctions, and systemic risks (the report’s own grouping)
Singapore’s safety body	Digital Trust Centre at NTU, designated Singapore’s AI Safety Institute on 22 May 2024
Singapore Consensus	Research priorities agreed by 100+ experts after a Singapore conference on 26 April 2025
EU AI Act	Main high-risk rules now start 2 December 2027 (stand-alone systems) and 2 August 2028 (in regulated products)
Singapore scams, 2025	41,974 cases, S$913.1 million lost, both down on 2024 (Singapore Police Force)

The WahLiao Verdict

Real today	Errors, scams, deepfakes, privacy leaks. Worth acting on now.
Genuinely debated	Loss of control. Serious people disagree, and as far as we can tell nobody has proof either way.
Watch closely	Bio and cyber misuse, flagged by labs and the Bengio report alike.
Take with salt	Anyone who is certain it’s either nothing or the end of the world.
Your part	Verify, protect your data, report fakes. The rest is for policymakers.

The everyday harms: already here

The most common harm is also the simplest: Super Intelligence can be wrong while sounding perfectly fluent. The International AI Safety Report lists “fabricating information, producing flawed code, and giving misleading advice” among current malfunctions. Our explainer on how Super Intelligence works shows why a system that predicts plausible text can make confident mistakes. Bias is the quieter cousin: a model trained on skewed data can give skewed results, which matters most when it’s used to screen people.

Then there are the harms people cause with it. The Bengio report documents AI-generated content used in scams and fraud, and rising deepfakes, including non-consensual intimate images. In Singapore, the good news is that total scam losses actually fell in 2025, to S$913.1 million across 41,974 cases, according to the Singapore Police Force. Still, voice clones and fake video calls give old tricks a new face. Two pages that may help: how to spot a deepfake scam and what happens to what you type.

The societal risks: jobs, power and elections

On jobs, the International AI Safety Report says economists disagree on the overall effect on employment, but it notes early evidence of “declining demand for early-career workers” in AI-exposed fields. For the Singapore picture, have a look at Budget 2026 and your job.

A second worry raised by many observers is concentration: a handful of companies and countries control the most capable models, the chips and the data centres, which raises questions about competition and about who gets to set the rules. Our guide to who makes the models maps out the field.

Elections are where deepfakes meet democracy. Singapore passed the Elections (Integrity of Online Advertising) (Amendment) Act on 15 October 2024. It bans publishing digitally manipulated online election advertising that realistically shows a candidate saying or doing something they didn’t, from the issue of the writ to the close of polling. The details are in deepfakes and the law.

The frontier debate: misuse and loss of control

Two frontier risks get the most attention. The first is misuse: helping someone build a biological weapon or run a cyberattack. The Bengio report notes that several developers released new models with extra safeguards because they “could not exclude the possibility” of the models helping novices towards biological weapons, and that AI agents are now capable of finding software vulnerabilities. The second is loss of control: the worry that future systems, especially AI agents acting on their own, could pursue goals nobody intended in ways people can’t stop.

The range of views here is wide, so we’ve tried to set each one out fairly and in its own words.

Concerned. In May 2023 the Center for AI Safety published a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Signatories included Turing Award winners Geoffrey Hinton and Yoshua Bengio, OpenAI’s Sam Altman and Google DeepMind’s Demis Hassabis.

Sceptical. Yann LeCun, then Meta’s chief AI scientist, called the doomsday scenario “preposterous” in 2023, arguing it is “not the way anything works in the world”.

Focus on now. Critics such as machine-learning scientist Sasha Luccioni have argued that talk of extinction distracts from present harms like bias and consent.

The scientific middle. The 2026 International AI Safety Report doesn’t pick a side. In its assessment, current systems can’t pose loss-of-control risks, capabilities are improving, and experts substantially disagree on how fast progress will come.

A quick note on words: “superintelligence” in this debate means a hypothetical system far beyond human ability, not WahLiao’s house name for all AI. We explain the difference in AGI and superintelligence explained.

What is being done

The labs. At the Seoul summit on 21 May 2024, 16 companies including Amazon, Anthropic, Google, Meta, Microsoft and OpenAI signed the Frontier AI Safety Commitments. They agreed to publish safety frameworks, define intolerable risk thresholds and, in extreme cases, not develop or deploy a model at all. Examples include OpenAI’s Preparedness Framework (updated April 2025, tracking biological and chemical, cybersecurity and AI self-improvement risks) and Anthropic’s Responsible Scaling Policy (version 3, February 2026). These commitments are self-imposed, and critics point out that the companies both write and mark their own tests.

Governments and scientists. Singapore’s AI Safety Institute, the Digital Trust Centre at NTU, belongs to an international network with Australia, Canada, the EU, France, Japan, Kenya, South Korea, the UK and the US. In December 2025 the network renamed itself the International Network for Advanced AI Measurement, Evaluation and Science. The Singapore Consensus on Global AI Safety Research Priorities, drawn up after a Singapore conference on 26 April 2025, groups safety research into three areas: building trustworthy systems (development), evaluating their risks (assessment), and monitoring and stepping in after release (control).

Laws. The EU AI Act is the broadest law so far. A “Digital Omnibus” amendment, in force since 27 July 2026, pushed its main high-risk rules to 2 December 2027 and 2 August 2028, and added a ban on AI systems designed to generate non-consensual intimate imagery and child sexual abuse material. Singapore has gone for targeted laws plus voluntary guidance: the election deepfake law, the Online Safety Commission (operating since 29 June 2026, starting with harms such as intimate image abuse, doxxing and online harassment), and IMDA’s Model AI Governance Framework for Agentic AI, launched on 22 January 2026. Rules in this area move quickly, so check the official page before you act. More in Singapore’s Super Intelligence plan.

Risks side by side

Risk	Near or long term	What is being done	What you can do
Confident errors	Now	Lab testing; model reports	Check anything that matters at the official source
Bias in decisions	Now	Governance frameworks; EU high-risk rules from 2027	Ask whether a person reviews AI-screened decisions
Scams and voice clones	Now	Police enforcement, platform takedowns	Call back on a number you know; agree a family code word
Privacy leaks	Now	PDPA; provider opt-outs	Keep NRIC, passwords and client data out of chatbots
Election deepfakes	Now, peaks at elections	Singapore ban from writ to polling	Don’t share; report suspected fakes
Jobs disruption	Now to 2030s	Retraining support, research	Learn the tools in your own field
Bio and cyber misuse	Emerging	Lab safeguards, safety institute testing	Keep software updated; use strong sign-in
Loss of control	Debated, longer term	Safety research, the Singapore Consensus	Follow named, accountable sources

AI safety and risk: FAQ

Is AI dangerous?

Some harms are real now: confident errors, scams, deepfakes and privacy leaks. Bigger risks such as bio misuse and loss of control are debated by experts, and the 2026 International AI Safety Report says current systems can’t yet cause loss of control.

Could AI cause human extinction?

Experts disagree. Signatories of the 2023 Center for AI Safety statement called extinction risk a global priority; sceptics such as Yann LeCun call the idea preposterous. As far as we can tell, no one has evidence that settles it.

Does Singapore have an AI Safety Institute?

Yes. IMDA designated the Digital Trust Centre at NTU as Singapore’s AI Safety Institute on 22 May 2024. It works with counterparts in the international network of safety institutes.

What is the Singapore Consensus on AI safety?

It’s a 2025 report by more than 100 experts from academia, industry and government that sets shared priorities for AI safety research in three areas: development, assessment and control.

When do the EU AI Act rules apply?

The Act is being phased in, and some parts already apply. After the 2026 amendment, the main rules for high-risk systems start on 2 December 2027, or 2 August 2028 for AI built into regulated products.