Has AI Gone Too Far?

The question is asked constantly and almost always answered badly, because two entirely different arguments get bundled together. One is about existential risk from some future machine intelligence. The other is about the concrete damage being done right now by systems already in production. Only the second is worth most people's attention, and it receives far less of it.

Start with what is genuinely alarming, and none of it involves a robot uprising.

Synthetic media is the clearest harm. Voice cloning from thirty seconds of audio is trivially available, and it is already being used to defraud elderly people by imitating a grandchild in trouble and to authorise fraudulent payments by imitating a chief executive. Non-consensual imagery generated from ordinary photographs has produced real and widespread harm, overwhelmingly to women and girls, including in schools. This is not a speculative risk. It is a present crime wave with inadequate enforcement.

Automated decision-making is the second. Systems are being used to filter job applications, price insurance, allocate benefits and assess risk in criminal justice, and they encode whatever bias sat in the training data. The individual affected frequently cannot see that a model was involved, cannot see the reasoning, and has no meaningful route to appeal. That is a serious erosion of due process, and it is happening quietly in administrative back offices rather than dramatically.

The third is the collapse of provenance. When any text, photograph or recording might be synthetic, the value of evidence itself declines. The danger is not that people believe fakes; it is that they stop believing anything, which is a far more useful condition for the dishonest.

The fourth is concentration. The compute, the data and the capital required to build frontier systems sit with a very small number of companies, and the infrastructure now being built consumes serious quantities of electricity and water in places that were not consulted about it. Whatever one thinks of the technology, that is a genuine question of public accountability.

Now the other side, because the panic is frequently misdirected.

The claim that these systems are about to become conscious and hostile is not supported by anything in how they actually work. They are extraordinary statistical predictors, and prediction at scale produces behaviour that looks like understanding without being it. Treating a language model as an emerging mind leads to bad policy, because it directs attention away from the person who deployed it and towards an imaginary agent with no legal address.

The claim that all creative and professional work is finished is also overstated. What is genuinely threatened is the middle of every distribution — competent, generic, undifferentiated output. Work that depends on judgement, accountability, relationships and physical presence is not being replaced. But I would not be complacent, because a great many careers are built precisely in that middle band, and the transition will be painful whether or not the endpoint is benign.

So has it gone too far? In some specific places, unambiguously yes. Deployment has outrun consent. Systems are being used on people who never agreed to it, in decisions that affect their livelihood, with no explanation and no appeal. That is the line that has been crossed, and it is a governance failure rather than a technological one.

What would sensible regulation look like? Provenance marking for synthetic media, with real penalties for stripping it. A legal right to know when an automated system materially affected a decision about you, and a right to human review. Liability that attaches to the deployer, not diffused into the model. Consent and compensation for the data used in training. Transparency on energy and water use for large facilities. None of that requires slowing down research; it requires making somebody answerable.

My own position, having used these tools every working day for several years, is that the technology is more useful than the sceptics allow and far less magical than the marketing suggests. It is a power tool. We do not ban power tools; we require guards on them, and we hold the operator responsible. That is the conversation we should be having, and it is more urgent than the science fiction one.