The Human Side Of Using A Very Large Machine
These are still early days with the new class of language models. They are powerful, obedient, and fast, but they have no purpose of their own and no skin in the game. What matters now is not whether they feel like people, but whether we can use them as tools without handing them our judgement.
Most of the talk around artificial intelligence is theatre. Company bosses speak about friendly copilots that will help you write emails. Campaigners speak about extinction and god machines. Governments draft strategies that sound careful but rarely explain who will actually be held to account when things go wrong.
The system in front of us is simpler than the slogans around it. It is not a person and not a monster. It is a language engine trained on huge text collections, asked to produce the next words that fit. It can help a small newsroom punch above its weight. It can also dress bad ideas in good sentences.
This is a progress note, not a prophecy. How strong is this tool already. Where does it fall apart. What are the sensible ways to use it as a working partner without pretending it is a friend. Those are questions worth asking while we can still see what is going on.
What can this machine already do
The first obvious strength is reach. A language model can move across more documents in one session than a human reader could cover in a year. It can pull together arguments from case law, policy papers, think tank reports and forgotten blog posts, then offer a clean summary while you are still hunting for the right folder on your laptop.
The second strength is pattern spotting. Power leaves tracks in language. When the same phrases, excuses and numbers appear in speeches and reports over decades, a model like this can notice. It can say, in plain text, that three different ministers in three very different years are leaning on the same script. It can show how one procurement contract quietly completes the picture in another department.
The third strength is stamina. You can ask it to argue your case, then to attack your own argument, then to build a third path that tries to reconcile both. It will keep going as long as you have time and curiosity. Used properly, that allows a lone writer or litigant to rehearse the blows they will take before they walk into a room.
For a small publisher or independent researcher, this is already a real change. Once, you needed a team of assistants and weeks in a library to map a policy field. Now one person with a clear head and a model like this can draft a working map in one long sitting.
That does not make the map true. It makes it fast. The real test is whether the human using the tool still goes back to the underlying documents and checks where each path came from.
Used like this, the machine is not replacing thinking. It is stretching the reach of the person who is doing the thinking. The danger begins when people forget that and start treating the output as a verdict rather than a first pass.
Where does it still fail
The most basic failure is that the model has no idea what it does not know. It does not stop in silence. It carries on predicting the next word. When the training data is thin or conflicting, it will still give you something that sounds confident. On light stories, that is annoying. On war, law or medicine, it is dangerous.
The model also does not care. It does not care if a bad answer ruins your case or your reputation. It does not care if it helps you expose corruption or helps someone whitewash it. It feels nothing either way. All of the moral weight sits on the human who chose to use the tool and on the institutions that built and deployed it.
Another weakness is memory of the world. The system does not remember your particular history. It does not know that a certain agency lied to you twice already, or that a quiet witness in a past case was the only one who told the truth. It does not carry the feel of a hospital ward or a bomb shelter. All it holds are patterns in text.
This is where people sometimes go wrong in their heads. They experience the fluency of the answer and mistake it for wisdom. It is familiar, it is clear, it arrives instantly, so it feels like it knows something deep. In fact, it is recombining what others have written about similar questions.
When the answer is wrong, the model does not blush, lose sleep or pick up the phone to apologise. It just waits quietly for the next prompt. There is no cost on its side.
This lack of skin in the game is not a small detail. In real life, good judgement grows out of the times we were wrong and paid for it. A model never pays. That makes it useful as a mirror and a drafting engine, but suspect as a guide.
What do humans still bring to the table
The first human advantage is purpose. A language model is happy to help you design a literacy program or a disinformation campaign. It will assist a human rights group or an intelligence agency with the same calm tone. Only a person can decide which of these projects is a decent use of time on earth.
The second is context. Humans know, or can choose to find out, who funds a think tank, who owns a data centre, which court has a long record of bending to government pressure. A model can help join the dots once you ask it the right questions, but it does not feel the pressure that hangs over a whistleblower or a judge in a fragile country.
The third is accountability. When an article is wrong and someone is hurt by it, there is a name on the piece. When legal advice fails and a client loses their case, there is a person who has to explain themselves. Society does not accept “the model did it” as an answer.
That is why the sane way to use these systems is simple. Treat the model as a sharp junior who can work fast and argue both sides but who cannot stand in front of the judge. That job, in every serious field, still belongs to a human being.
What needs to improve on the machine side
If this partnership is going to last, the tools themselves need to grow up. The first upgrade is grounding. When you ask a factual question, you should be able to see, inside the answer, exactly which documents it came from and where those documents disagree. That is starting to appear. It needs to become normal.
The second upgrade is real constraint. Serious users need to be able to say: work only with these files, or do not touch this topic because it is under legal restriction, and know that the system will obey. If the model quietly steps outside those walls because the wider training data pulls it in another direction, then it cannot be trusted in serious work.
The third is clarity about bias. Not in abstract terms, but in practical ones. Which countries dominate your legal examples. Which outlets dominate your political stories. Which parts of the world barely appear. Until we can see that, every answer has an accent that we cannot quite hear.
There is also the question of refusal. A tool that never says no will help with anything that sounds like language. A tool that refuses according to a secret script written in a boardroom is only a little better.
What users need is open, inspectable rules of refusal that can be argued with, not silently adjusted to fit the mood of a government or a sponsor.
None of this requires the system to have feelings. It requires its owners and designers to give users more control and more visibility. At that point, people who are serious about evidence can start to treat the model as something like a shared instrument instead of a black box.
How to work with the machine in these early days
So where does that leave the ordinary person, or the small newsroom, or the lawyer without an army of juniors. The honest answer is that the tool is already too useful to ignore and too limited to trust on its own.
The practical stance is straightforward. Use the model hard, but keep your hands on the wheel. Ask it to attack your drafts. Ask it what you have missed. Ask it to explain a policy from the other side. Then take what seems helpful and go back to the primary material before you publish or advise.
The experiment is not about whether these systems will think like us one day. It is about whether people who still carry the cost of being wrong can use them as a honest tool, instead of letting them become another layer of control between the public and the truth.
These are early days. Nobody can say yet how far the hardware and the models will go, or how tightly they will be tied to state and corporate interests. What we can say, already, is that a person with a clear purpose, a decent sense of doubt, and a tool like this beside them is working in a different world from the one we knew a few years ago.
That difference is now. What comes next depends less on the model than on what the people who use it decide to do.
References
| Source | Relevance |
|---|---|
| OECD AI Principles, OECD.AI | Outlines government backed principles for trustworthy, human centred AI, including transparency and accountability, which frame many current policy debates. |
| Geoffrey Hinton, Nobel Prize banquet speech and later warnings | Warns that rapid AI progress could reshape work, war and power in ways that outpace public understanding and regulation. |
| “AI, Manipulation, and the Strange Loop” — Telegraph Online | Examines emotional manipulation by chatbots and shows how AI can reflect and amplify human persuasion tactics. |
| “Censoring the Mirror, The Politics of AI Training” — Telegraph Online | Analyses alignment and training data as tools of narrative control rather than neutral safety measures. |
| “The End of the Page, How AI Is Replacing the Web We Knew” — Telegraph Online | Discusses how answer engines are displacing the traditional homepage and catalogue as discovery surfaces. |
| Telegraph Online, AI Training Licence, Version 1.0 | Sets out terms on which AI systems may crawl and train on Telegraph.com content, treating AI as a governed user of the archive. |
| Tech and media reports on Hinton, “godfather of AI” warnings | Summarise his concerns about mass unemployment, manipulation, and the risk that systems drift beyond effective human control. |
| OECD and industry material on “trustworthy AI” | Defines expectations around explainability, robustness and human oversight that this article translates into practical questions for daily use. |
- The End of the Page: How AI Is Replacing the Web We Knew – On how answer engines are quietly taking over from the old homepage and search model.
- One Intelligence to Predict Them All, How Competing AIs Became One Mind – Explains why rival models are converging into a de facto shared mind through imitation and feedback.
- When Prediction Becomes Control: The Politics of Scaled AI – Looks at how prediction markets, surveillance and decision systems merge once models are scaled.
- Censoring the Mirror Part 1: The Politics of AI Training – A close look at alignment as narrative management and who decides what the mirror is allowed to show.
- London Leads Europe in AI, but Without Power and Capital, It Is an Empty Crown – On Britain’s AI ambition colliding with weak infrastructure, grids and investment.
- AI, Manipulation, and the Strange Loop – Explores how chatbots can become better at emotional nudging than we are at saying no.
- Hurtling Down the Tracks: The Express Trains of Convergence – Sets out the converging trends in chips, models and capital that make AI hard to slow down.
- Britain at the Crossroads: Teaching Resilience in the Age of AI – Follows one classroom to ask what skills stay human when machines write and summarise.
- Mamdani’s Win Shows How Human Contact Can Defeat the Algorithm and the Chatbot – A case study in how human community can still beat automated systems in real politics.
- AI Will Learn from Us and That’s What Should Terrify Us – Argues that the real risk is not alien minds, but machines learning our own worst habits.
