AI Strategy · Memory & Judgment

Differently Sighted, Not Objective

I pointed an AI at my own career and it revalued a project I’d shipped seventeen years ago — along an axis I’d never valued it on. I reached for the word objective. It was the wrong word. The AI isn’t neutral; it over-weights whatever got written down. What it actually is, is un-invested — and that’s the truer, more useful claim.

By Scott Farrell · LeverageAI

A little while ago I pointed a fairly smart AI at a wiki I’d built over the exhaust of my own career — every project, every proposal, a couple of decades of email, let’s just say. I asked it about a project I’d done for a client something like seventeen years ago. I had a story about that project already: I thought I’d executed it well, and that the pivot I’d found for the client — a disintermediation play — was the genuinely clever part.

The AI found the proposal, which I expected. Then it did something I didn’t expect. It surfaced an internal email from back then, where I’d explained to my staff and my business partner how I was going to frame the idea, and why. That email never went to the client. It never made it into the proposal. It was the deliberation behind the deliverable — and the machine handed it back to me, in my own words, timestamped. Off the back of it, the AI revalued the whole project along an axis I had honestly never valued it on. It saw a different worth, on a different dimension, than I had for seventeen years.

Sitting there, the word I reached for was objective. It felt like the machine had done a clean-room inspection of my past — no ego, no rose-coloured glasses, none of the fuzz and self-regard that clouds my own memory. And that word is where I want to stop you, because I think it’s wrong. Not slightly wrong. Wrong in a way that matters, because if you call this thing “objective” you will overtrust it, and you’ll miss what it’s actually good for.

The word I reached for was the wrong word

“Objective” smuggles in a claim it can’t back: a view from nowhere. A neutral eye that sees your history as it truly was. The AI does not have that eye, and it’s easy to prove.

The corpus it read isn’t reality. It’s the subset of reality that got written down. Proposals that were formalised, emails somebody actually sent, documents that survived a dead application and a dozen migrations. Everything that lived only in a meeting, a phone call, a corridor conversation, a decision made and never minuted — none of that is in there. So the AI over-weights the written record, systematically, because the written record is all it has. That is a skew. A real one, with a shape you can name.

Which means it isn’t neutral. It’s biased toward the documented. So why did it feel so clean?

Because the bias it doesn’t have is the one I was drowning in. My view of that project was invested. I had a stake in the story — a version of myself I was quietly protecting. We reason toward the conclusions we’re motivated to reach, and a flattering account of your own past is about as motivated as reasoning gets.² The AI has no such stake. It doesn’t need that project to have meant what I’d decided it meant. It wasn’t defending anyone.

The right word isn’t objective. It’s un-invested. The AI has no identity stake in your story — so it won’t defend the flattering version. That’s not the same as seeing from nowhere.

The distinction isn’t pedantry. The two skews aren’t symmetric, and that’s the whole point. The corpus’s bias is structural and disclosable — “I can only see what was written down” is a sentence the system can say out loud, and you can correct for it. My ego’s bias is invested and invisible to me — it defends itself, and it never announces its presence. A structural skew you can name is a tool. An invested skew you can’t see is a trap. Un-invested beats unbiased not because it’s perfectly clear-eyed, but because its blind spot is one you can account for and mine can’t.

Your self-narrative is a cached compilation

Here’s the part that actually rearranged something for me. I’d been treating my sense of my own career as a fact — a settled account of what I’d done and what it was worth. It isn’t a fact. It’s a cache.

Years ago, I compiled an answer to the question who am I, and what has my work meant. I ran that query once, under the conditions of one particular moment, and I got back a story. And I’ve been serving that same stored answer ever since, mostly without recompiling it. When someone asks about my career, I don’t re-derive it from the evidence — I read out the cache.

The trouble is the way the cache was built. It was built from free recall, under time pressure, and free recall is the weakest retrieval channel you own. It hands you the recent years and the headline roles — whatever comes to mind most easily — and quietly drops the rest.¹ A whole lattice of forgotten depth just isn’t in the summary, not because it wasn’t valuable but because it wasn’t available the day you compiled. I’ve written before about how a CV works exactly this way: not a memoir you compose from memory, but a query you run against an indexed life — and the moment you index the exhaust, recognition returns depth that recall never could.

What the AI-plus-wiki does is refuse to read the cache. It recompiles the past from the source records, live — and it can recompile under a different query than the one I originally ran. That is a specific, nameable capability: reconstructing a past state from distributed traces that no single record ever held, delivered fast enough to land inside the same thought that summoned it. The internal email wasn’t “found” the way you find a file. It was compiled back into view as part of answering a question I hadn’t quite asked.

Once you see your self-narrative as a cache and not a fact, the uncanny experience stops being uncanny. The AI didn’t see my career “more truly.” It ran a fresher query against a fuller index than the one I’d been quoting from since 2009.

Rose glasses off, blue glasses on — the lens is a parameter now

Here’s where the “objective” instinct really falls apart. When I brought a new client’s situation to that old project — a live deal where the seventeen-year-old work might be relevant — the AI reframed and re-appreciated the whole thing for that new context. My old rose-coloured glasses came off. A fresh set went on: call them the blue glasses of the new client’s needs. And the project looked different again.

The tempting way to describe that is “the AI removed my bias and showed me the truth.” But that’s not what happened, and pretending it is would put us straight back into the objectivity trap. There is no lens-free reading of that project. There’s the rose lens, and the blue lens, and a hundred others I could load — each one a different query, each one surfacing a different worth. Reframing for the new client isn’t a truer reading. It’s another one, chosen on purpose.

You don’t take the glasses off. You make the glasses a parameter — a knob you can turn on purpose — and the honest deliverable stops being “neutral” and becomes “disclosed.”

That is the real upgrade, and it’s a subtler one than “de-biasing.” For most of my life the lens on my past was fixed and invisible — I couldn’t see it, so I couldn’t change it, so it just was the past. The wiki turns the lens into something I can name and swap. Rose off, blue on. Finance lens, then risk lens. The lens that asks “what did I do well” and the one that asks “what would I never do again.” The bias doesn’t vanish; it becomes a setting. And a bias you can dial deliberately is worth infinitely more than a neutrality you can only pretend to have.

Why the reframe landed instead of offended

There’s a version of this that should have felt awful. Imagine a machine announcing that the work you were proud of actually mattered for a completely different reason than you thought — that your own sense of your career was, let’s be blunt, a bit off. Told to me as a verdict, that would have stung. Who asked it? What does a model know about my life?

It didn’t sting, and the reason is a design property, not a matter of tone. The system obeyed what I’ve called the dignity constraint: show the exhibit, not the conclusion; store pointers, not verdicts; keep the human as the terminal judge. It didn’t tell me I’d misunderstood my own career. It handed me a timestamped email — my words, my framing, from a decade and a half ago — and let the realisation arrive on its own. The revaluation was mine. The machine just made the evidence reachable while the thought was still warm.

You cannot argue with your own old email. A verdict from a machine is cheap and slightly insulting; an exhibit you can open and read is something else entirely — it carries its own warrant. That’s what let an un-invested reappraisal feel like a gift instead of a correction. It never asked me to believe it. It showed me the receipt and got out of the way.

And notice what the receipt was, because this is the quietly important part. The reusable judgment in that project didn’t live in the proposal — the proposal only says how we’d do the work. It lived in the internal email: here’s how I’m framing this, and why. That’s the situation and the move; the proposal is just the outcome. Organisations only ever showcase the shipped artifact, so the actual reusable thinking — the deliberation — evaporates. No case-study interview would ever have recovered it, because nobody retroactively rereads staff email from seventeen years ago. The wiki did. The un-invested reader went to the one place the honest answer was still sitting, precisely because it had no stake in only quoting the polished version.

The honest limits — because “un-invested” still isn’t “omniscient”

Two limits keep this from being magic, and both follow directly from getting the word right.

First: it can only reappraise what the exhaust actually holds. The disintermediation pivot survived because I’d written it down to my staff. If the cleverest thing I ever did had lived only in a conversation nobody recorded, there would be no receipt to compile, and the correct output would be a documented absence — “nothing in the record supports this” — not a manufactured find. A system worth trusting has to be able to come back with nothing and say so cleanly. That’s the same virtue as un-investedness wearing a different hat: a witness that could also have found against you, or found nothing, is the only kind whose confirmation is worth having.

Second: this is a retrieval story, not a storage one. My past wasn’t lost — it was sitting in files from an application that doesn’t even exist anymore, data I’d been hoarding for years without knowing why. What was broken was my ability to read it: to find the right record and re-appreciate it at the right moment. Swap “my aging memory” for “an organisation’s ten years of dead SharePoint” and the whole argument transfers untouched. The storage was always fine. The reading was never affordable — until an un-invested reader could do it, continuously, for free.

Closing: the truer claim

I built the wiki, and the AI to walk it, to find the thing I didn’t know existed — and it did. But the lesson that stuck wasn’t about the find. It was about the word I nearly used to describe it. “Objective” would have flattered the machine and misled me. It would have told me to trust a view from nowhere that doesn’t exist, and to forget that the whole apparatus can only ever see what got written down.

“Un-invested” is the honest word, and it happens to be the more useful one. It tells you exactly what to lean on — a reader with no stake in your flattering story — and exactly what to keep in your own hands: the exhibit, the lens, the final call. Your self-narrative is a cache; let it be recompiled. Your bias is a fixed pair of glasses; make it a parameter. And your history isn’t waiting to be seen truly — it’s waiting to be seen differently, by something that doesn’t need you to be who you’ve always said you were.

Differently sighted beats unbiased. And it’s the truer claim.

The one thing to carry out of here: if an AI ever tells you it sees your past objectively, don’t trust the sight — trust the exhibit it’s willing to show you, and stay the judge.