Public discussion of the harms of this technology tends to occupy two positions, one concerned with long term existential risk and one dismissing the entire subject as overstated. The concrete problems are in neither place. They are ordinary, they are present, and most of them follow from a cost curve rather than from anyone's intentions.
Capability concentrates because capital does
Training a frontier model requires an amount of computation, data, and specialist staff available to very few organisations. That is a structural fact rather than a policy choice, and it produces a structural outcome, which is that the most capable systems are built by a small number of well funded entities.
The consequence is dependence. Organisations building on these systems inherit decisions they do not control, including pricing, availability, permitted uses, and behaviour that can change between versions. That is a familiar situation with any critical supplier, and it is worth recognising as the same situation rather than a novel one.
It is also worth noting the counterweight honestly. Openly available models trail the frontier but are considerably more capable than most tasks require, which limits how much leverage concentration actually confers for everyday work.
The labour effect is narrower and stranger than expected
The usual framing asks which jobs disappear, which has not been the observable pattern. The pattern is that specific tasks within jobs become much cheaper, and the tasks affected are disproportionately the routine ones.
That matters more than it first appears, because routine tasks are how people enter a profession. Drafting basic documents, writing straightforward code, producing first pass analysis, and summarising material are the work through which judgement is built. Removing them does not remove the need for experienced practitioners, it removes the path by which people become experienced.
The effect is therefore uneven in a specific direction. Established practitioners gain leverage. Entrants face a market where the work that used to be theirs is cheaper to automate than to teach. Whether that resolves through changed training routes or persists as a structural gap is genuinely unsettled.
Content that was used without agreement
These systems are trained on large quantities of text, images, and code, much of it published without any expectation of that use. Whether that constitutes permissible use of publicly available material or unlicensed appropriation is being decided in courts across several jurisdictions, and the answers are neither consistent nor final.
Two practical points hold regardless of how that resolves. Content contributed before the question arose cannot be withdrawn from models already trained on it, because removing a specific record from a trained model is not a supported operation. And the outcome affects who can afford to build, since a licensing requirement would favour organisations able to pay for data at scale, which is the same set already able to pay for computation.
Verification became the expensive part
Producing plausible text, images, audio, and video used to require effort proportional to the quality wanted. That relationship has broken, and the cost of production has fallen far faster than the cost of verification.
The visible consequences are fraud using synthesised voices of known individuals, fabricated media presented as evidence, and automated content produced at volumes that overwhelm ordinary moderation. The less visible consequence may be more significant, which is that the existence of convincing fabrication provides cover for denying genuine material. Once anything can be dismissed as synthetic, authentic evidence loses some of its force, and that erosion applies whether or not any particular item was actually fabricated.
Inference, not collection, is what changed
Surveillance concerns predate this technology, and the data being gathered is largely the data that was already being gathered. What changed is the cost of drawing conclusions from it.
Correlating records across sources, identifying individuals in footage, inferring characteristics not deliberately disclosed, and monitoring communications at scale were all previously limited by the human effort involved. That limit was doing considerable protective work without anyone designing it that way, and its removal changes the practical meaning of data that has been held for years.
People defer to confident systems
A well documented tendency is that operators over trust automated output, particularly when it is fluent and usually correct. Attention degrades precisely because the system is reliable most of the time.
This undermines the control most often cited as the answer to automated decision making. Human review is only meaningful if the reviewer is genuinely evaluating rather than approving, and a reviewer presented with a high volume of confident recommendations will approve. Designing for that reality means keeping the reviewed set small enough that scrutiny is possible, rather than requiring approval everywhere and assuming it is being given carefully.
Energy, stated honestly
Training and serving these models consumes substantial electricity and, where evaporative cooling is used, substantial water. Both figures are growing quickly and the growth is concentrated geographically, which places real strain on particular grids and water systems.
The honest qualification is that reported figures vary widely, that providers disclose selectively, and that the sector remains a small fraction of total electricity demand even as it grows. Both the alarming and the dismissive characterisations tend to select their comparisons. The defensible position is that the impact is real, local, rising, and poorly measured.
Accountability has no obvious home
When an automated system produces a harmful outcome, responsibility is genuinely difficult to place, and not because anyone is evading it.
The model was trained by one party, adapted by another, integrated by a third, and operated by a fourth, and the behaviour emerged from the combination rather than residing in any single component. The developer cannot enumerate what the system will do. The operator cannot inspect why it did what it did. Neither position is dishonest, and between them there is no party able to give a complete account.
Existing liability frameworks assume a defective product or a negligent operator, and a system that behaves as designed while producing an unforeseeable outcome fits neither category well. This is being worked through in regulation across several jurisdictions, and it remains unresolved.
Separating what is happening from what might
Everything above is occurring now and is measurable to varying degrees. That distinguishes it from the speculative end of the discussion, which concerns capabilities that do not yet exist.
Both conversations are legitimate, and conflating them is unhelpful in a specific way. Attention directed at hypothetical future risk is attention not directed at concrete present harm, and the present harms have identifiable causes, identifiable beneficiaries, and available remedies. They are the tractable part of the problem.
Note: the pattern connecting most of these is that something previously constrained by human effort stopped being constrained. The effort was never intended as a safeguard, and it functioned as one.