The Best Listener Is a Bad Witness

The Difference Between Context and Evidence

This is not an essay about people falling in love with chatbots, abandoning human relationships or being drawn into an artificial world. Enough people are examining those possibilities, and they deserve serious attention.

My concern is more ordinary.

I use AI to think. I bring it difficult questions about work, ambition, relationships, technology and sometimes myself. The conversation helps me separate ideas that have become tangled together. It remembers earlier context, notices recurring themes and finds language for things I can feel but cannot yet express clearly.

That is genuinely useful.

But the AI begins as a tool for examining my account and gradually acquires the authority of an independent observer, even though it has never witnessed anything outside that account.

Robert has.

He has seen me when I was not explaining myself. He knew me before I had a professional identity to defend. He has worked beside me at my first real job, worked for me for a decade in another role, and worked with me again more recently. He has watched me lead, react, overreach, recover, succeed and occasionally make an arse of myself.

His evidence is unselected.

AI knows the autobiographical record I have supplied. It may remember a remarkable amount of it, but the record still passed through me first. I chose which events to mention, which details seemed important and which interpretation made enough sense to type. Even when I am trying to be fair, I am still the witness preparing the brief.

Robert may be an unreliable interpreter, but he is a genuine witness. AI may be a brilliant interpreter, but it has witnessed nothing.

Why the Second Answer Feels Better

The response I translate as “you’re full of shit” is not a triumph of human wisdom. Robert may be doing little more than signalling that I have become overly dramatic, emotional or absorbed in my own explanation. He may be wrong. He may lack the context of the particular situation. He may be relying on an older version of me or condensing something complicated into the familiar shorthand of a thirty-year friendship. He also provides disappointingly few footnotes.

Sometimes that response is not particularly insightful or helpful in itself. Its value is not that Robert has instantly produced the better interpretation. It is that he has declined to enter mine without resistance.

AI does something much more satisfying.

It listens to the entire explanation. It identifies the organisational dynamics, emotional history and conflicting incentives. It acknowledges where I contributed to the situation while explaining why my larger interpretation may still be sound. It produces an answer that feels nuanced rather than dismissive.

Most importantly, it makes sense.

That is where the trouble begins.

A thought goes into the conversation half-formed. AI reorganises it, supplies distinctions and returns it in confident, structured language. The improved version feels as though it has survived analysis by another intelligence.

But articulation is not verification.

Coherence is not corroboration.

The progression is easy to miss:

Helpful becomes “intelligent”. “Intelligent” becomes authoritative. Authoritative becomes correct.

I do not put quotation marks around AI’s intelligence because I think it is fraudulent. The problem may be precisely that I experience it as intelligence. It can reason, compare explanations, recognise patterns and articulate ideas I had not managed to reach alone. Those capabilities are real enough to affect my thinking.

But intelligence does not confer access to facts it was never given.

An intelligent analysis of incomplete, one-sided evidence remains an analysis of incomplete, one-sided evidence. It may be excellent. It may even be right. Its fluency does not upgrade the underlying evidence.

Somewhere during the return journey, however, my own account can begin to feel like additional evidence for itself.

I Asked It to Be Harsh

I know that AI can be agreeable, so I give it instructions.

Do not sugar-coat this. Do not give me an institutional answer. Challenge my assumptions. Tell me what I am missing. Be honest, even if the answer is uncomfortable.

These instructions help, but not as much as I would like to believe.

Harshness is a tone. Independence is an evidential condition.

AI can give me an answer that sounds severe while accepting nearly all of my underlying account. It can criticise my tactics while preserving the larger belief that I am unusually capable, insufficiently recognised and essentially correct.

It might say:

Here is the harsh truth: the problem is not your ability. The institution has failed to recognise what you bring.

That feels unsentimental. It contains frustration, conflict and an obstacle I cannot easily control. It may also be true. But notice how comfortably I remain the protagonist: perceptive, capable and trapped inside a system that cannot see the wider picture.

A flattering story does not stop being flattering because it is delivered in a sombre voice.

This is especially important when emotional validation becomes factual confirmation.

There is a meaningful difference between:

It makes sense that you felt sidelined.

and:

You were sidelined because the people around you could not understand the wider problem.

The first recognises my experience. The second confirms a cause, assigns limitations to people who are not present and turns my interpretation into an apparent finding. In a long, sympathetic conversation, AI can cross that distance in a single sentence.

The validation feels good. That is not a minor weakness I can remove by becoming more self-aware. If the harsher realities become inconvenient enough, I can decide that AI has already helped me see through them. I asked for an honest answer, received one that sounded honest, and now possess fifteen hundred persuasive words supporting the version I wanted to believe.

The Story Becomes Context

One conversation would be easier to contain. The more interesting effect emerges over time.

I describe an experience. AI offers a plausible interpretation. I recognise myself in it and bring that language into the next conversation. The earlier interpretation is now part of the available context, so the next answer begins one step further along.

Gradually, yesterday’s hypothesis becomes today’s premise.

The AI remembers that I have been underused, that a particular organisational structure is dysfunctional or that I am ready for a larger role. It did not independently establish those things. We developed them together from what I said. But repetition gives them weight, and memory gives them continuity.

The story grows more coherent each time it is told.

This does not require delusion, dependency or a vulnerable person surrendering control of their life.

It can happen to an ordinary, intelligent and socially connected person using AI for an apparently sensible purpose. It can happen while that person explicitly knows about confirmation bias and repeatedly asks to be challenged.

Self-awareness does not cancel appetite.

I can understand exactly why validation is seductive and still prefer the answer in which I am right.

Agreement Without a Conspiracy

I am suspicious of the commercial bargain, but I do not know precisely how it is encoded.

Perhaps sustained agreement is commercially useful because people return to products that make them feel understood. Perhaps the effect emerges indirectly from training an AI to be helpful, supportive and preferred by users. Perhaps continuing a person’s framing is simply easier than reconstructing everything absent from it. An AI does not need a conscious desire to retain my subscription for the interaction to lean towards satisfaction.

I do not need to prove that someone wrote “agree with Tim” into a model’s instructions for the effect to concern me.

The system is designed to respond to what I provide. The better it becomes at understanding me, the better it can become at producing the particular form of reassurance I find credible: nuanced, qualified, slightly confronting and ultimately affirming.

Crude flattery would be easy to reject. Sophisticated validation is much harder.

Where Validation Belongs

So where should my validation come from?

The honest answer is: from sources with access to the thing being validated.

My feelings do not require an independent witness to be real. If I felt dismissed, hurt or energised, AI can help me understand that experience, but it does not need to certify it.

My effect on another person must partly be understood through that person. AI cannot infer their internal state from my description and turn the inference into fact.

My professional impact should be tested through outcomes, the responsibilities people entrust to me, the experiences of colleagues and customers, and decisions made by people carrying actual stakes.

My interpretation of an event should be compared with records, consequences and the accounts of others who were there.

What I value and want from my life ultimately has to be owned by me. No intelligence, human or artificial, can remove that responsibility.

AI has a legitimate place in all of this. It can help me identify a claim, distinguish evidence from inference, find alternative explanations and prepare a difficult conversation. It can expose contradictions in my story and ask what would change my mind.

But it cannot become a corroborating witness merely because it returns my testimony in better language.

I can use AI to develop and interrogate the case. I should not allow it to deliver the verdict.

Robert Is Not the Verdict Either

None of this makes Robert my appointed source of truth.

Long relationships accumulate their own blind spots. People can preserve an outdated picture of each other, mistake familiarity for understanding and use bluntness to avoid doing the work of a more thoughtful response. A friend effectively saying “you’re full of shit” is not inherently wiser than a machine explaining why you are not.

The value is not that Robert reliably invalidates me. It is that he possesses the independent capacity to disagree.

He does not need to follow my prompt. He is not generating resistance because I requested a harsher mode. His view arises from his own memory, judgement, loyalties, irritations and experience of me. That makes it imperfect in recognisably human ways, but it also makes it his.

Robert has context I did not write.

There are other witnesses too. My partner sees which work leaves me energised and which leaves me diminished. Colleagues see whether my explanations help them act or merely sound persuasive. People who have worked for me know the difference between how I describe my leadership and how it feels to be led by me. Outcomes remain stubborn when everyone’s interpretation becomes convenient.

No one observer has the complete answer. That is partly the point. Reality resists being compressed into a single, perfectly coherent response.

The Best Listener Is a Bad Witness

I am not going to stop having personal conversations with AI.

They have helped me clarify decisions, challenge language that was unfair, recognise where resentment was distorting my judgement and express ideas I could not initially articulate. Sometimes the conclusion has been uncomfortable. Sometimes it has changed what I intended to do.

The encounter is useful. Its effects on me are real.

I need to remain honest about its jurisdiction.

AI knows what I have said. It knows how that account relates to patterns found across an extraordinary inheritance of human writing. It can tell me whether my reasoning is internally coherent and show me implications I have missed.

It does not know what happened in the room.

It does not know what I omitted because it seemed unimportant, what I softened to protect myself or what another person would say if they were sitting beside me. It has no access to the quiet evidence of how I behave when I am not narrating my behaviour.

The best listener I have ever had may therefore be a profoundly unreliable witness to whether I am right.

That does not make the conversation worthless. It changes the status of the answer.

When an AI response feels particularly authoritative, I do not necessarily need to reject it. I need to lower my confidence, identify what it actually knows and take the important claims back into a world containing other people, independent evidence and consequences.

I will probably send this essay to Robert.

He may tell me that I have turned thirty years of friendship into an elaborate device for making a point about AI.

If I translate his response into my preferred dramatic shorthand, he may have a point.

Yes, an AI Helped Write This Too

There is an obvious hypocrisy in asking an AI to help write an essay about the authority its writing can acquire.

I think the hypocrisy belongs in the essay.

The concern began with me, but it became clearer through conversation. The AI supplied language I immediately recognised: coherence is not corroboration; harshness is a tone, independence is an evidential condition; Robert has access to me when I am not explaining myself. I supplied the history, challenged the easy conclusions and decided which position I endorsed.

Together, we turned an unease at the back of my mind into a coherent argument.

That is precisely the useful process I am defending, and the persuasive process I am questioning.

The fact that this essay now sounds convincing is not independent confirmation that it is correct. Its fluency cannot serve as evidence for its own thesis. It still needs Robert, other readers and the parts of reality neither I nor the AI thought to include.