The Hardest Problem of the AI Revolution Might Be Trust

I have started noticing AI in places where nobody says AI was involved.

Sometimes it is an emoji appearing at the beginning of every other paragraph. Sometimes it is a strangely polished rhythm that does not quite sound like the person who posted it. And sometimes it is that very particular long dash that suddenly seems to be everywhere. None of these prove anything, of course. People use emojis. People use em dashes. People occasionally write suspiciously well.

But I find it interesting that we have already started developing an instinct for this. We look at a piece of text and wonder not only who wrote this?, but what wrote this? Recently I noticed these little fingerprints in completely unrelated places: a political post and the title of a video by a popular food creator. They were not bad pieces of content. Nothing was necessarily wrong with them. Yet my brain immediately did the same thing it increasingly does elsewhere.

Was this written with AI?

That small hesitation points toward a much bigger problem.

I don’t think energy is the hardest AI problem

For years, when people have asked me what worries me about the rapid growth of AI, energy consumption has rarely been at the top of my list. That does not mean it is irrelevant. Training and running large models require enormous amounts of computation, electricity, cooling and infrastructure. If AI usage continues growing at its current pace, we will need better chips, better data centres, more electricity generation, better grids and probably technologies that do not exist at scale today.

But this is a category of problem we understand. We are engineers. Computers have always consumed a ridiculous amount of energy relative to the task humans used to perform without them. We built machines, networks, power plants, cooling systems and semiconductor fabs because the economic value of computation justified the infrastructure around it.

There was even a wonderfully early-internet example called Blackle: essentially a black version of Google created around the idea that displaying darker pixels might save electricity. The details turned out to depend heavily on display technology, but I still love the idea. We were already worrying about the energy consumption of search engines before smartphones had taken over the world.

If AI becomes economically valuable enough, we will build infrastructure for it too.

Maybe that means new kinds of power generation, much more solar, better batteries, more efficient accelerators, better models and better cooling. Probably all of the above. These are difficult, expensive and politically complicated problems.

But they are engineering problems.

Trust is different.

We do not yet have a power grid for truth.

My sister just scrolls past it

I was talking with my sister recently about AI-generated content when she casually mentioned something that stayed with me. When she sees a social media post marked with something like contains AI-generated content, she usually just scrolls past it.

Immediately.

She does not investigate which model was used, check whether AI generated the entire thing or merely adjusted a photo, or estimate whether the person behind the post used the tool responsibly. She simply removes the source from consideration.

At first I argued against this. If using AI automatically makes something less trustworthy, where exactly do we draw the line? Is Grammarly cheating? Is spellcheck? What about autocomplete? Should we stop using computers entirely and return to writing letters with quills on handmade paper?

Perhaps the only truly authentic communication is delivered by pigeon.

It is an easy argument to make, especially for someone who works with technology. Tools have always extended human capability. But the more I thought about it, the more I realised that her reaction was not irrational at all.

She was doing risk management.

The label introduced an unknown number of additional variables into the creation of the content, and she had almost no information about any of them. Ignoring the content was simply cheaper than evaluating the new trust chain.

And that is where I think both of us were partly right.

The trust chain just became much longer

Before generative AI, if somebody I trusted made a claim about a subject they understood, the trust relationship was relatively simple: there was a source, a human and a claim. Reality was never actually that clean. People misunderstand sources, misremember facts, have biases and occasionally just make things up. But the mental model worked reasonably well.

Now imagine somebody publishes an answer they produced with an AI assistant.

Behind that answer might be training data, a particular model and model version, system instructions, retrieved documents, search results, conversation history, a user prompt, tool calls, generated text and, finally, a human deciding that the output looks correct enough to publish.

The reader sees almost none of this.

If I tell you, “I checked this with AI”, you now have to trust more than me. You also have to trust my ability to use AI. Did I choose an appropriate model and give it the relevant context? Did I phrase the question well? Did I ask it to verify its assumptions? Did I check the sources it provided and notice when it confidently invented something? Did I even understand the subject well enough to recognise a bad answer?

AI literacy is becoming part of source credibility.

When someone uses AI, trusting their answer increasingly means trusting their ability to use AI.

For people working in security or distributed systems, this feels strangely familiar. We have spent decades building systems around chains of trust. Certificates do not magically make communication trustworthy. Identity providers do not magically make users trustworthy. A signed package is not automatically good software. Instead, we construct systems where one trusted component can make claims about another, and where those relationships can be inspected and verified.

When those assumptions become too dangerous, we move toward zero trust:

Verify explicitly.

Trust as little as necessary.

Assume boundaries can fail.

This used to sound like infrastructure architecture. Now ordinary people are being forced to apply a crude version of the same reasoning while scrolling Instagram.

“Contains AI” is probably the wrong question

This is also why I increasingly dislike the binary distinction between AI-generated and not AI-generated. It hides more than it explains.

Imagine two people writing an article.

The first person writes every sentence themselves, then asks an AI tool to fix three grammar mistakes. The second person types:

Write me a convincing article about this subject.

They copy the result without reading the cited sources and publish it under their name. Both may have “used AI”.

Those are obviously not equivalent processes.

Between them lies an enormous spectrum: spellcheck, grammar correction, rewriting, translation, summarisation, brainstorming, research, fact checking, code generation, reasoning and, eventually, autonomous systems performing entire workflows.

So when I see a label saying contains AI-generated content, I understand my sister’s instinct. The label tells her that something happened. It tells her almost nothing about whether she should trust the result.

Perhaps the important question is not:

Was AI involved?

Perhaps it is:

Why should I trust this output?

That is a much harder question, but also a much more useful one.

AI is not a magic machine

Part of solving this problem will require better technology: provenance, better citations, auditable workflows, reproducible outputs where possible and clearer uncertainty. We need systems that can show which documents influenced a conclusion instead of merely presenting a polished paragraph, and ways to distinguish generated claims from verified claims.

But technology alone will not fix this. We also need better public understanding of what these systems actually are. AI is often discussed as if we have built a mysterious electronic mind living somewhere inside a data centre.

We have not.

What we have built is arguably more impressive.

At its core, a language model is still software repeatedly estimating a probability distribution over what token should come next, based on the enormous amount of structure it learned during training and the context currently available to it. That simple description dramatically understates what emerges from the process. Modern models can write software, analyse documents, use tools, interpret images, translate languages and reason through problems that would have looked like science fiction surprisingly recently.

But none of that turns them into magical truth machines.

They are computational systems. Extremely advanced ones, but computational systems still have inputs, assumptions, failure modes and operating conditions. The quality of their output depends on the model, the training, the context, the tools, the prompt, the surrounding software and the human deciding what happens next.

Understanding this does not make AI less impressive. It makes using it responsibly possible.

The next challenge is verification

We spent the last few years teaching people how to use AI: write better prompts, give it context, ask follow-up questions, connect it to tools, build agents, automate workflows. That will continue.

But I suspect the more important challenge of the next few years will be teaching people when to trust what comes out of it.

The goal cannot be to make AI infallible. We never managed to make humans infallible either. The goal should be to build systems where claims can be traced, assumptions can be inspected, sources can be verified, uncertainty can be communicated and trust can be earned rather than implied. That is partly a technical problem, partly an educational one and, eventually, probably a cultural one.

Because the strange little AI fingerprints I notice today will disappear. Models will get better at matching individual writing styles. AI-assisted text will stop having such obvious patterns. The emoji habits will change. The suspicious punctuation will disappear.

We will no longer be able to look at a paragraph and think, I can tell ChatGPT wrote this.

Strangely, that may make the underlying problem more visible rather than less. Once we can no longer guess how something was created, we will finally have to ask the question that mattered all along:

What makes this worth believing?

The AI revolution’s hardest problem may not be generating information.

It may be rebuilding the chain of trust around it.