I recently saw someone and have often seen people saying the following about artificial intelligence with two words:

Fuck AI!

The most recent version of this involved someone in the middle of a disagreement. They asked a large language model to compose a response, sent what it produced without properly considering it, and made the situation worse.

I understand the reaction. Family disagreements are emotional, context-heavy, and full of history that no language model can completely know. A generated message can be technically polished while being entirely wrong for the people receiving it. It can sound cold, clinical, patronizing, manipulative, or strangely unlike the person supposedly sending it.

I do not think the lesson is “fuck AI.”

The lesson is: you still have to think.

If you copy an AI-generated response verbatim, send it into a sensitive human situation, and then blame the tool when it goes badly, you have misunderstood both the technology and your own responsibility.

The model did not decide to send the message.

You did.

Generated text is not judgment

A large language model can produce remarkably convincing language. That is one of its strengths, but it is also one of its greatest dangers.

Fluent writing creates an impression of thoughtfulness. A neatly organized response can look reasoned even when it is based on incomplete context, questionable assumptions, or a complete misunderstanding of the relationship involved.

The model does not know the look your brother gives when he is hurt but trying not to show it. It does not know which phrase your mother will interpret as condescending. It does not know that a technically accurate point has already been repeated six times and that repeating it once more will only deepen the argument.

It knows the information you supplied, the patterns learned during training, and the conversation currently in front of it.

That can be useful. It is not the same thing as wisdom.

AI can generate options. It can suggest a structure, identify a blind spot, remove unnecessarily hostile wording, or help us see how a message might be received. It cannot relieve us of the obligation to decide whether those words are true, appropriate, humane, and recognizably ours.

Copy and paste is not using AI well

I use large language models regularly. I use them professionally, for technical projects, for learning, for debugging, and sometimes for improving written communication.

I do not simply copy the result and press Send.

If I ask an LLM to help refine an email, I read what it produces. I compare it with what I intended to say. I retain useful phrasing, remove language that does not sound like me, restore context that was lost, and often merge its suggestions into my own draft.

The finished message remains my message.

That is not an insignificant distinction. Editing the result is where critical thought re-enters the process. I am asking:

• Is this factually correct?
• Does it express what I actually believe?
• Is the tone appropriate for this person?
• Has the model inferred something I never said?
• Is it avoiding an uncomfortable point that needs to be addressed?
• Is it escalating a situation while sounding superficially calm?
• Would I be willing to say these words face to face?
• Am I prepared to take responsibility for the result?

If the answer to that last question is no, I should not send it.

Blindly forwarding generated text is not thoughtful adoption of AI. It is outsourcing the appearance of thought.

Human-in-the-loop is not corporate jargon

“Human-in-the-loop” can sound like one of those phrases people place in a slide deck to make an AI project seem responsible.

It describes something essential.

The human provides the objective, context, constraints, judgment, verification, and final accountability. The system contributes speed, language generation, pattern recognition, analysis, or implementation assistance.

Remove the human judgment and the process becomes ungrounded. The model may still produce an answer—these days a very polished one—but nobody is confirming that the answer belongs in the real world.

In my arcade-cabinet work, agents have written code under my direction. They have helped analyze logs, build maintenance tools, normalize controller behaviour, audit metadata, and document complicated interactions among Linux devices and services.

That does not mean I type “make it work” and accept whatever appears.

I define the intended behaviour. I set boundaries. I ask the agent to inspect the existing system. I review the proposed change. We test it against the actual hardware. When the evidence contradicts an assumption, we correct the assumption. If a script could damage data, I require dry runs, backups, explicit apply modes, and verification afterward.

The code may be generated agentically. The responsibility is not.

That is human-in-the-loop.

AI magnifies the way we already work

AI can help a thoughtful person examine more possibilities, work through unfamiliar material, and accomplish more in a day. It can also help a careless person produce mistakes faster and with greater confidence.

In that sense, it is an amplifier.

Give it a carefully defined problem, relevant evidence, meaningful constraints, and an engaged human reviewer, and it can be extraordinary.

Give it one side of a family argument, ask it to prove that you are right, and paste the result into a group chat, and it can provide beautifully formatted gasoline.

This is not unique to AI. People have always misused tools, repeated claims they did not understand, forwarded articles they did not read, and borrowed other people’s words to avoid difficult conversations. AI makes that behaviour easier, faster, and more convincing.

That increased power deserves increased care.

You cannot outsource accountability

Suppose an LLM produces a message that damages a relationship. The model deserves scrutiny. Perhaps it responded too confidently, mirrored the user’s anger, validated a distorted premise, or generated language that was inappropriate for a vulnerable situation.

The organizations building these systems also have responsibilities. They should communicate limitations honestly, design safeguards thoughtfully, and stop presenting every generated answer as though confidence and correctness were the same thing.

But the person who chose to send the message is still part of the causal chain.

We cannot demand complete personal credit when AI helps us produce something successful and then claim complete helplessness when the same process produces harm.

If I publish generated code without testing it, the resulting bug is my responsibility.

If I submit invented citations without checking them, the false claims are my responsibility.

If I send a cruel or tone-deaf message without reading it carefully, the damaged relationship is my responsibility.

“The AI wrote it” is an explanation of process. It is not an exemption from accountability.

The problem is choosing not to think

I will admit that this is the part that frustrates me most.

We have been given access to tools that can help us explore ideas, learn new skills, analyze huge amounts of information, and build things that previously would have required far more time or specialized knowledge.

And some people use that power to avoid thinking.

They paste the first response. They do not check facts. They do not question assumptions. They do not notice that the tone is wrong. They do not ask whether the output reflects their own beliefs. Sometimes they do not even appear to read it closely before attaching their name to it.

It is tempting to say that people doing this are too stupid to think.

I think the more accurate criticism is harsher: in that moment, they are choosing not to think.

Intelligence is not the only issue. Patience, humility, curiosity, and personal responsibility all matter. Critical thinking requires us to slow down long enough to ask whether a convenient answer is actually a good one. AI does not remove that requirement. It makes it more important.

Use AI to strengthen your voice, not replace it

There are constructive ways to use an LLM during a disagreement.

You might ask:

• What assumptions am I making about the other person’s motives?
• Which parts of my draft are likely to sound accusatory?
• Can you show me three interpretations of what they said?
• What important context might be missing from my account?
• Can you help me shorten this without removing the apology?
• Where am I trying to win rather than resolve the issue?
• What questions should I ask before responding?

Those prompts use AI as a tool for reflection rather than as an automated weapon in an argument.

Even then, the output requires judgment. A model can suggest empathy without feeling it. It can construct an apology without meaning it. It can imitate vulnerability without taking the emotional risk that vulnerability requires.

The human being has to supply those things.

Sometimes the right conclusion after consulting AI will be to write the message yourself. Sometimes it will be to call rather than text. Sometimes it will be to wait until you are calmer. Sometimes it will be to say only, “I don’t want this argument to become worse. Can we talk tomorrow?”

No model should make that decision for you.

Criticism of AI is necessary

None of this means that AI should be above criticism.

There are serious questions about accuracy, bias, privacy, labour, copyright, environmental cost, concentration of power, manipulation, and the effect these systems may have on education and employment. There is plenty to criticize, and “fuck AI” may express a completely understandable anger about some of those issues.

But if the specific complaint is that someone blindly copied generated language into a family dispute, then we should identify the failure accurately.

The system produced words without understanding the full human situation. That is a known limitation.

The user treated those words as finished judgment. That was a human decision.

Both facts can be true at once.

Keep your hands on the controls

I do not advocate surrendering our work, relationships, or critical faculties to artificial intelligence. I advocate learning how to use it well.

Use it to generate possibilities, not commandments.

Use it to challenge your thinking, not replace it.

Use it to improve a draft, not impersonate you.

Check its facts. Question its assumptions. Rewrite its language. Test its code. Protect private information. Consider the person who will receive the result. Be willing to reject the output entirely.

Most importantly, take responsibility for whatever leaves your hands with your name attached.

AI is a tool. It can be a remarkably powerful one.

But you still have to think.