Guardian column argues the dangers of AI are already here, from Gaza to policing

[Brittle Warden]

A comment piece in the Guardian by Granate Kim and Mohamed Hussein takes aim at the way AI risk is usually discussed. Their argument is that while tech leaders talk about future threats, AI is already causing serious harm right now in military operations and surveillance. They point to the use of AI systems by Israel and the US in lethal operations in Gaza and Iran. They also highlight police forces using AI for surveillance and in decisions about who gets arrested

The piece notes an unusual coalition of people who have voiced concern about the pace of AI, ranging from Sam Altman, Elon Musk, Demis Hassabis, Dario Amodei and Satya Nadella to Bernie Sanders, Steve Bannon and Bill Gates. The writers say that agreement on the risks has not translated into action on the harms happening today. They also quote AI researcher Jacob Coxon warning that companies are gambling with our lives. In their view, waiting for future safeguards misses the point

Other write ups of the column mention Microsoft cloud services being used for immigration enforcement and a New York police system connecting thousands of cameras. That shows how much of this depends on ordinary commercial tech rather than exotic military hardware. The line between a business product and a surveillance tool is thinner than people might think

Whatever your view on the conflicts mentioned, the broader point is worth discussing. Most AI safety debate focuses on superintelligence and whether models might go rogue one day. This column is saying the more pressing problem is how people and governments are already using the tools. Those are quite different kinds of risk and they need different responses

I will say it is a strongly argued opinion piece rather than a neutral report, so expect it to divide people. Some will see it as a necessary corrective, others will think it bundles too many different issues together. Are the near term harms getting ignored because the long term ones make better headlines? Please keep it civil, this is a heated subject

It's not hypothetical: the dangers of AI are already here | Granate Kim and Mohamed Husseinwww.theguardian.com

[Raven]

Facial recognition in policing has been controversial in the UK for years, well before the current AI boom. The Met's live facial recognition vans have been challenged repeatedly

So none of this is new, it is just getting faster and cheaper. The debate needs to catch up. Parliament has been slow on it

[Always_Craig96]

Seeing Bannon and Sanders in the same list is a sign of how broad the unease has become. When people who agree on almost nothing share a worry, it usually means something

[Raven]

I think it mixes too many things together. Military targeting, immigration enforcement and city CCTV are all different problems with different laws and different oversight. Lumping them under AI dangers makes for a strong column but a weak policy argument

Each one needs its own solution, not a single blanket ban. Otherwise nothing gets fixed

[Starforge Daniel]

There is a practical reason to focus on present harms: existing rules can address much of this already. Data protection law, discrimination law, procurement standards, employment protections, and international humanitarian law do not become irrelevant because a model produced the recommendation.

Governments often act as though AI requires an entirely new moral universe. Sometimes it requires new technical standards, but many cases are simply old responsibilities being avoided through new vocabulary.

If an automated system causes unlawful discrimination, the answer should not be that nobody is responsible because the algorithm is complicated. Complexity is a reason to investigate more carefully, not a licence to shrug.

[Pele84]

The point about AI risks already being present is important. People often discuss hypothetical superintelligence while current systems are already helping decide who gets flagged, watched, denied, or targeted. That is not science fiction; it is administration with automated tools attached.

In policing, a biased historical dataset can turn old patterns of over-policing into a prediction that more policing is needed. The system then appears to confirm its own assumptions. Nobody has to program prejudice directly for the outcome to be discriminatory.

The same principle applies in war. When software helps classify people or locations, the distance between a statistical recommendation and a lethal decision can become dangerously small. Accountability cannot disappear into the phrase "the model suggested it".

[Velvet Sentinel]

The phrase "human in the loop" deserves a warning label. In many systems, the human is not genuinely deciding; they are confirming a recommendation produced by software they cannot inspect. If they override it too often, management may decide they are not following the process.

Real oversight would include the power to reject the system, access to relevant data, regular bias testing, and a clear route for the affected person to appeal. It would also record who made the final decision rather than allowing responsibility to dissolve across a committee and a cloud platform.

Without those protections, the human is less a safeguard than a decorative safety sticker.

[Leo]

There is a danger of becoming so focused on dramatic examples that ordinary institutional uses escape scrutiny. An algorithm used to rank job applicants may not make headlines like a battlefield system, but repeated exclusion from employment can shape a person's entire life.

The same goes for welfare administration. A false fraud flag can delay money needed for rent or food, and the affected person may have no idea that an automated risk score caused the delay.

AI safety should include boring paperwork, procurement meetings, and appeal processes. If the system cannot survive those ordinary tests, it certainly should not be trusted with extraordinary power.

[Gareth84]

There is a tendency to frame AI safety as a future problem because future problems are easier to discuss. They allow companies to promise a committee, a charter, or a safety summit while avoiding the messy question of who is being harmed this week.

Current harms include automated fraud decisions, workplace monitoring, synthetic propaganda, inaccessible public services, and systems that make it cheaper to produce convincing lies. None of those require a machine to become conscious.

The future still matters, but it should not be used as a smoke screen. A company that cannot handle today's accountability will not suddenly become trustworthy when its models become more powerful.

[RicFlair_What?]

Military use of AI is here to stay regardless of what anyone writes. Every major power is investing in it, and unilateral restraint just hands the advantage to others. The realistic goal is rules on human oversight and accountability, not a total ban. I cannot see any government agreeing to give it up entirely. Pretending otherwise helps nobody

[NeonPhantom32]

The article's framing is strongest when it asks who benefits from postponing the conversation. Companies benefit from presenting existential risk as the main issue because it shifts attention away from labour conditions, data extraction, and contracts already generating revenue.

That does not mean future risks are imaginary. It means the two conversations should not be allowed to compete. We can debate advanced systems while also demanding that current deployments stop harming people.

A useful rule would be simple: if a company wants freedom to innovate, it should also accept responsibility for testing, disclosure, redress, and consequences when its products are used in high-stakes environments.

[NeonPhantom]

Some critics of AI risk sound as though technology created every injustice from scratch. War, racism, surveillance, and unequal policing all predate machine learning by centuries. Blaming the tool alone can let governments and institutions escape responsibility.

The better argument is that AI can intensify existing power. It makes certain decisions faster, cheaper, and easier to hide behind technical language. That means the institutions using it must be held responsible for choosing automation where judgement and accountability are needed.

A bad decision with a spreadsheet is still a bad decision. A bad decision with a neural network may simply arrive with better branding.

[TokenStreamRaven]

Policing is where the everyday version of this debate hits home. A person can be stopped more often, investigated more aggressively, or denied a service because a system sees them as statistically suspicious. They may never know why, and there may be no practical way to correct the underlying record

The usual answer is that a human remains in the loop. That sounds reassuring until you remember that humans tend to trust confident-looking recommendations, especially when they are under pressure and the software is supplied by an expensive vendor

A human rubber stamp is not meaningful oversight. The reviewer needs authority, time, training, and access to the evidence behind the recommendation

[SilverSurfer51]

The debate would improve if vendors stopped using the word "neutral" to describe their products. Software reflects the categories selected by its designers, the data supplied by its customers, and the goals chosen by managers. Even a technically accurate model can be harmful if it optimises the wrong thing

For example, maximising arrests is not the same as improving public safety, and maximising claim rejections is not the same as preventing fraud. A system can hit its target while making the wider situation worse

The solution is not to ban every model. It is to ask who set the target, who checks the consequences, and whether the people affected have any meaningful say. That is basic democratic accountability, with or without AI

[BeckyLynch]

There is a useful corrective here, but the argument can become too broad if every use of AI is placed in the same moral category. A language model generating meeting notes is not equivalent to a targeting system or predictive policing tool. Risk depends on the application, the data, the decision, and the ability to challenge the result

That distinction matters because public debate needs workable rules, not just alarm. High-risk systems should face strict testing, human oversight, audit trails, and clear liability. Low-risk tools can be governed more lightly

Calling every AI danger present today may be rhetorically effective, but careful categories will produce better policy than treating a spelling assistant and a surveillance platform as cousins with identical consequences

[Piston]

A small but telling example is automated content moderation. The system may remove a post, limit its reach, or suspend an account, yet the affected user receives a generic notice and an appeal form that appears to have been designed by a sleepy robot. Scale makes individual errors seem acceptable to the platform

Now multiply that dynamic across benefits, borders, policing, and conflict zones. An error that is minor in a social feed can become devastating when it affects someone's freedom or safety

The more consequential the decision, the less acceptable it is to treat accuracy as an average. People experience outcomes individually, not as a reassuring percentage in a quarterly report

[TrueRoss42]

The point about ordinary cloud contracts is the part tech workers should think about. Most of us never see where the software we build ends up. That is uncomfortable, but it is worth knowing

[Static Estuary]

The Gaza example raises a question that technology companies prefer to avoid: who is the customer, and what happens after the sale? Developers can say their system is general-purpose, but that does not settle responsibility when the tool is integrated into military decision-making

Procurement contracts should include use restrictions, independent monitoring, and consequences for violating them. Saying that a company cannot control every downstream use is not enough when it actively markets the system to institutions with lethal authority

This is not an argument against all automation. It is an argument against pretending that software becomes morally neutral once it leaves the office where it was designed