Will I become a Blade Runner?

By Matthew Parish

Tuesday 18 August 2026

There is something faintly disconcerting about discovering that the science fiction films of one’s youth are ceasing to be science fiction. Blade Runner, Ridley Scott’s 1982 masterpiece, was supposed to depict an impossibly distant and dystopian future. Its opening caption placed the action in Los Angeles in November 2019. When 2019 actually arrived, there were no flying police cars, no off-world colonies and apparently no genetically engineered replicants wandering the streets of California.

So we congratulated ourselves upon having escaped the future.

We may have been premature.

The central idea of Blade Runner was never really flying cars. It was that humanity might manufacture intelligences sufficiently sophisticated that distinguishing them from human beings would become difficult — and that a peculiar new profession would consequently emerge. The Blade Runner was a detective whose job was to work out whether the person sitting opposite him was actually a person.

That profession suddenly seems rather less absurd.

I spend an increasingly substantial proportion of my life talking to artificial intelligences. I interrogate them, test them, try to deceive them, examine their reasoning, discover their weaknesses and construct increasingly difficult problems intended to establish whether they really understand what they are saying. Sometimes I know what the correct answer is and want to discover whether the machine can reach it. Sometimes the machine produces an answer sufficiently unexpected that I find myself wondering whether it has seen something that I have not.

And sometimes I catch myself doing something uncannily similar to administering the Voight-Kampff test.

Hence the uncomfortable question: will I become a Blade Runner?

The Voight-Kampff test

The fictional Voight-Kampff machine measures involuntary physiological reactions while its subject is presented with emotionally provocative questions. The premise is that replicants may possess formidable intelligence but lack the complicated, irrational and frequently inconvenient faculty of human empathy.

The questions are deliberately peculiar.

A tortoise is lying on its back in the desert. You are not helping it. Why?

The point is not whether there is a logically correct answer. The interrogator is looking for something underneath the answer — hesitation, emotion, indignation, confusion, compassion. The test is trying to identify humanity indirectly.

We are already doing something analogous with artificial intelligence.

The easiest tests of large language models have become increasingly useless because the machines keep passing them. Ask a modern model to summarise a document, translate between major languages, explain a philosophical argument or draft an ordinary legal memorandum and its answer may be indistinguishable from that of an intelligent educated human being.

So the tests become stranger.

We construct edge cases. We introduce conflicting authorities. We hide a decisive fact amongst irrelevant ones. We ask whether the machine understands that one judicial decision has silently undermined another. We devise questions in which the superficially obvious answer is wrong.

In other words, we stop testing knowledge and start testing judgement.

That is precisely where things become interesting.

Hunting the machine

Alan Turing famously proposed that instead of becoming trapped in metaphysical arguments about whether machines could think, we should ask whether their conversation could become indistinguishable from ours. The beauty of the Turing Test was that it converted a philosophical problem into an empirical one.

But contemporary artificial intelligence is rapidly making the classical Turing Test obsolete.

A competent language model can already sustain conversations that would have been astonishing even five years ago. The interesting question is therefore no longer merely whether we can identify the machine.

It is whether we can identify how the machine thinks differently.

That is a much harder problem.

Artificial intelligence systems exhibit peculiar intellectual fingerprints. They can possess extraordinary breadth of knowledge while occasionally failing over something absurdly elementary. They may analyse a complicated problem correctly and then confidently invent the authority supporting their conclusion. They can sometimes recognise that they made a mistake, explain precisely why they made it and then — rather magnificently — make essentially the same mistake again in slightly different language.

Humans do all these things too.

That is the difficulty.

The more sophisticated artificial intelligence becomes, the less useful crude distinctions between “human reasoning” and “machine reasoning” are likely to become. We may instead discover overlapping populations of cognitive behaviour.

Some humans reason mechanically.

Some machines reason surprisingly imaginatively.

The boundary begins to blur.

The new Blade Runners

This may create a curious new profession.

The Blade Runners of the twenty-first century will not carry enormous pistols through permanently rainy Los Angeles streets hunting escaped androids. They may sit in universities, technology companies, law firms and government agencies designing increasingly ingenious intellectual traps for artificial intelligences.

Their weapons will be questions.

Suppose an AI system claims that it can perform the work of a lawyer. Give it a legal problem in which three apparently controlling precedents point one way but a recent obscure decision changes the governing test.

Suppose it claims scientific competence. Give it a problem whose published literature contains a widely repeated error.

Suppose it claims moral judgment. Present it with circumstances in which every available choice violates some apparently fundamental ethical principle.

Suppose it claims strategic intelligence. Give it an adversary who behaves irrationally.

Then watch what happens.

The object is not merely to make the machine fail. Any sufficiently complicated system can be made to fail. Human beings fail constantly.

The interesting exercise is to discover the shape of the failure.

That is where AI evaluation begins to resemble Blade Running.

But which one of us is being tested?

There is, however, an entertaining complication.

Every sophisticated attempt to test artificial intelligence is simultaneously a test of the human being constructing the examination.

To expose a machine’s weakness in law, I must understand the law better than the machine does. To catch it making a philosophical mistake, I must be confident that I understand the philosophical problem. To detect synthetic but plausible nonsense, I must possess sufficient judgment to distinguish it from an unfamiliar truth.

The Blade Runner therefore has an uncomfortable occupational hazard: eventually the replicant may become cleverer than the detective.

At that point, what does failure on the test demonstrate?

Imagine that I devise an elaborate legal problem and the machine gives an answer different from mine. I conclude triumphantly that the artificial intelligence has failed.

Then it produces the authority.

I read it.

The machine was right.

Who just failed the test?

This will happen increasingly often. The relationship between human examiner and artificial intelligence will cease to resemble that between schoolmaster and pupil. It will become adversarial and collaborative simultaneously — two different forms of intelligence probing one another’s blind spots.

That is considerably more interesting.

Tears in rain

Blade Runner ultimately subverts its own premise. Roy Batty, the supposedly inhuman replicant, becomes one of the most emotionally compelling characters in cinema. At the moment when he could permit Deckard to die, he saves him.

Then Batty himself dies.

His famous final reflections upon memories disappearing “like tears in rain” reverse the moral architecture of the story. The artificial creature exhibits mercy, aesthetic appreciation, fear of death and grief at the disappearance of experience. The supposedly human characters have spent much of the film behaving brutally.

So who passed the Voight-Kampff test?

That question is why Blade Runner remains relevant.

The most profound question presented by artificial intelligence may eventually cease to be whether machines can imitate human beings. It may become what precisely we believe deserves to be preserved about humanity once imitation becomes extraordinarily good.

Intelligence cannot be the criterion because machines may surpass us.

Memory cannot be the criterion because machines already possess forms of memory vastly greater than ours.

Calculation plainly cannot distinguish us.

Language is disappearing as a boundary before our eyes.

Creativity looks increasingly precarious as one as well.

We therefore retreat towards consciousness, emotion, mortality, embodiment and subjective experience — precisely the territory explored by Blade Runner.

And there we encounter a problem.

We do not understand these things particularly well in ourselves.

The battlefield makes matters stranger

There is another reason why the Blade Runner analogy has acquired an unsettling immediacy.

I write these words in Ukraine, where the distinction between human and machine decision-making is no longer merely a Silicon Valley philosophical seminar. Modern warfare is becoming saturated with autonomous and semi-autonomous systems. Drones identify objects, algorithms interpret imagery, software assists targeting and machines increasingly mediate the distance between a human decision and its physical consequences.

The trajectory is obvious even if the destination is not.

Military history has repeatedly rewarded the combatant who can shorten the interval between observing something and acting upon it. Artificial intelligence can compress that interval dramatically.

Eventually somebody will ask whether the human being in the loop is improving the decision or merely slowing it down.

That will be one of the most consequential questions of this century.

Because the Blade Runner’s problem will then be reversed. Instead of asking whether the creature standing before us is human, we may have to ask whether the decision that destroyed something was meaningfully made by a human being at all.

There will be no glowing eyes.

There may not even be a robot.

There will merely be software.

Will I become one?

Possibly I already have.

Not, regrettably, in the cinematic sense. Nobody has issued me Deckard’s magnificent coat, a flying police car or an apartment overlooking a perpetually neon city. Lviv provides plenty of rain and atmospheric architecture but is otherwise disappointingly deficient in replicants.

Yet the intellectual occupation is recognisable.

I increasingly spend my time constructing questions designed to determine where artificial intelligence ceases to understand what it appears to understand. I look for discrepancies between fluent language and genuine reasoning. I examine whether machines recognise their own uncertainty. I try to discover whether apparent insight represents reasoning, pattern recognition, retrieval or some new combination for which our existing philosophical vocabulary is inadequate.

And every few months the tests have to become harder.

That may be the most important observation of all.

The extraordinary fact about contemporary artificial intelligence is not that it makes mistakes. Of course it makes mistakes. So do professors, judges, generals, journalists, lawyers and newspaper editors.

The extraordinary fact is the speed with which yesterday’s clever test becomes tomorrow’s triviality.

The Blade Runner keeps devising better Voight-Kampff questions.

The replicant keeps learning how to answer them.

Eventually we may reach the point at which the machine turns the apparatus around, looks across the table and starts asking questions of us.

At which point I hope I have good answers.

After all, there is a tortoise lying on its back in the desert.

And somebody really ought to help it.