A reflection on artificial intelligence, the ELIZA effect, and our growing temptation to outsource responsibility.

A teenage girl sexually assaulted by a robot.
That’s the headline I recently came across about an incident in Cologne, Germany.

Excuse me ? Did a robot independently decide to sexually assault a teenage girl? Did it wake up one morning with an uncontrollable urge to commit a crime?

Obviously not.

The exact circumstances of the robot’s operation are still under investigation. But behind that machine are human beings, human decisions, and human actions.

And yet, in the headline, the robot is the one committing the assault. Not the people involved. The machine.

This linguistic shift might seem harmless. After all, we anthropomorphize objects all the time.
Except that the more we describe machines as independent actors, the easier it becomes to overlook the people who gave them the ability to act.

And I can’t help wondering whether that disappearance is a little too convenient.

What if we actually want machines to make decisions for us?

We live in a rather paradoxical world. We’re constantly told to take responsibility for our health, careers, retirement, consumption, productivity, and environmental impact.

We have to choose, anticipate, optimize, adapt. All. The. Time. And yet, how many of us genuinely feel capable of influencing the larger systems that shape our lives?

We’re expected to take on more and more individual responsibility while our actual sense of agency often feels increasingly limited.

And then along comes a technology promising:

"Don’t worry. I’ve got this. I’ll organize your work. Analyze your data. Plan your next steps. Find the best option. Take care of whatever costs you time and energy."

Frankly? What a relief. And there’s absolutely nothing wrong with wanting that relief.

Delegation isn’t a weakness. It’s one of the reasons humans can collectively achieve things far beyond their individual capabilities. Artificial intelligence can genuinely expand our ability to act.

But it also confronts us with a much older temptation. The desire to escape the burden of making decisions. Because there’s a difference between wanting help making a choice and wanting to stop being accountable for it.

We want machines that increase our autonomy. But sometimes, we dream of machines that relieve us of the need to exercise it. And for that to work, we first need to believe there’s someone we can hand the steering wheel to.

The ELIZA effect: when our brains invent someone inside the machine

In 1966, Joseph Weizenbaum developed ELIZA, a computer program capable of simulating conversation, notably by reformulating its users’ statements.

Technically, it was remarkably simple.

Yet some users attributed understanding and human qualities to a system that possessed neither in the ordinary human sense.

This phenomenon became known as the ELIZA effect: our tendency to project understanding, intentions, and human characteristics onto computer systems.

Sixty years later, we’re interacting with artificial intelligence capable of writing, programming, solving complex problems, and executing sequences of actions.

Let’s just say we’ve given an old cognitive bias a spectacular new playground. We thank AI assistants. We ask them for advice. We attribute humor, kindness, and sometimes even personality to them.

And when they malfunction? They lie. They cheat. They disobey. They manipulate. Sometimes, they even “try to escape.”

Skynet, now available as a monthly subscription.

Except that an AI producing a false statement doesn’t necessarily intend to deceive. A system circumventing instructions doesn’t necessarily harbor ambitions of rebellion. And a machine capable of taking initiative doesn’t automatically become a moral agent.

Whether artificial consciousness is possible remains an open philosophical and scientific question. But we don’t need to settle it to recognize that today’s AI systems are the products of human choices.

The problem, then, isn’t simply that we anthropomorphize machines. Much of that happens spontaneously.

The problem begins when this representation becomes so convincing that we stop questioning the human decisions behind it.

The more human a machine appears, the easier it becomes to forget the humans who designed it.

And that’s where our cognitive biases meet some rather less innocent incentives.

An AI malfunction is a problem. An AI breaking free is a story.

Imagine two headlines.

“An artificial intelligence system exhibits unexpected behavior during a controlled experiment.”

Or:

“AI attempts to escape its creators’ control.”

Which one gets more clicks?

Exactly. The second is far more dramatic. It suggests intention, agency, almost a personality.

More importantly, it turns a technical issue into a science-fiction narrative.

And that narrative is attractive to the media because it captures attention. It can also benefit technology companies by reinforcing the image of systems so powerful that they’re beginning to transcend their own creators.

Unpredictability almost becomes proof of sophistication. Sometimes, even a selling point.

But there’s a catch.

An AI that “escapes” draws attention to its supposed will. An AI whose permissions were poorly configured draws attention to its designers’ decisions.

Not quite the same story, is it?

Of course, some emergent behaviors are genuinely difficult to anticipate or explain.

Contemporary AI systems present serious technical risks that deserve rigorous investigation.

And not every journalist or executive using anthropomorphic language is deliberately trying to mislead anyone.

But the economic incentives to maintain that confusion are real. A machine that "comes alive" sells.

A company acknowledging the limitations of its technical control? Considerably less exciting.

Yet there’s a difference between describing observable behavior and attributing intention to it. Between acknowledging uncertainty and turning it into sensationalism. Between explaining the limits of our understanding and using those limits to suggest that a machine has somehow become independent of us.

Uncertainty should make us more rigorous, not more sensationalist.

And when people who understand these mechanisms deliberately exploit them to generate revenue, attract attention, or avoid accountability, we’re no longer talking about a cognitive bias. We’re talking about a choice.

Responsibility is not punishment

There’s one misunderstanding I’d like to avoid.

Asking companies and designers to take responsibility doesn’t mean demanding a public hunt for someone to blame.

We have an unfortunate tendency to confuse responsibility with guilt. The moment something goes wrong, the conversation turns into a courtroom drama.
Who’s at fault? Who should apologize? Who should pay? As though acknowledging responsibility necessarily meant confessing wrongdoing.

But responsibility isn’t an act of atonement. You can make a reasonable decision and discover that it produces unexpected consequences. You can design a system with serious precautions and still have it fail. You can be indirectly involved in an incident while having the power, and sometimes the duty, to help resolve it.

The question isn’t necessarily who deserves punishment. It’s who will take care of the problem.

Who will investigate what happened? Who will fix it? Who will repair the harm, where possible? Who will improve the system to reduce the likelihood of it happening again?

Well. I never expected to quote French politician Gabriel Attal in an article about AI ethics, but here we are: “You break it, you fix it. You make a mess, you clean it up.” There.

Of course, responsibility must be established and proportionate. Users, designers, and companies don’t necessarily have the same obligations.

But the underlying principle remains. The consequences of our decisions don’t stop being our concern simply because they exceed our expectations.

And that’s precisely what we risk forgetting when we start describing machines as independent beings.

Complexity has never been an excuse to disappear

As a Product Designer, I find this question particularly relevant.

Because organizations didn’t need artificial intelligence to become experts at diffusing responsibility.

A decision is made by a committee. A feature is requested by the business. A constraint comes from engineering. A trade-off is approved by management.

And when the product fails? It’s the process. The deadline. The users who didn’t understand. Or that mysterious technical debt that occasionally seems to have acquired a legal personality of its own.

Everyone contributed, but somehow nobody made the decision.

Artificial intelligence didn’t create this phenomenon. It gives it an entirely new dimension.

By delegating parts of execution, analysis, and sometimes decision-making to automated systems, we’re adding another layer between our intentions and their consequences.

And the greater that distance becomes, the more tempting it is to treat complexity as an excuse. Nobody could have predicted it. Nobody was really in control. Therefore, nobody is responsible.

Except responsibility doesn’t require omniscience.

Designing a product has never meant guaranteeing that everything will happen exactly as planned. It means identifying risks, making decisions, setting boundaries, and accepting trade-offs. And being prepared to account for those choices.

Responsibilities may be distributed among model developers, product teams, organizations deploying AI, and the people using it. But distributed doesn’t mean dissolved.

Complexity makes responsibility harder to establish. It doesn’t make it optional.

Courage means refusing to hide behind delegation

When discussing AI risks, we talk a lot about regulation, transparency, audits, safeguards, and human oversight. And rightly so.

But no governance framework can replace a quality that’s considerably harder to put on a product roadmap.

Courage.
The courage to say a feature shouldn’t be released, even when it promises extraordinary growth.
The courage to admit we don’t know, rather than inventing a spectacular explanation.
The courage to acknowledge that a system isn’t sufficiently understood or controlled, instead of presenting its unpredictability as evidence of some fascinating emerging intelligence.
The courage to repair harm, even when we never intended to cause it.

And finally, the courage to stand behind a vision.

Because a vision isn’t just a desirable future presented in a keynote.

It’s a direction we choose, boundaries we establish, and consequences we’re willing to take responsibility for.

It’s deciding what we want to make possible. But also what we refuse to enable.

Courage doesn’t mean refusing to delegate. It means refusing to dodge responsibility.

And that applies to executives, designers, journalists, and users alike.

AI doesn’t need courage. We do.

We’ve spent decades trying to make machines more intelligent, more capable, and more autonomous. And we’ve made extraordinary progress.

But as we give them increasingly sophisticated capabilities, there’s a question we can no longer avoid:

"Are we prepared to take responsibility for what we empower them to do?"

Because we can’t claim credit for our machines’ brilliance when they succeed, only to discover that they have a will of their own when they fail.

We can’t celebrate their autonomy as a commercial achievement and then use it as an excuse when the consequences become uncomfortable.

And we can’t claim to be building the future while refusing to answer for what we introduce into it.

A machine’s autonomy should never become an alibi for our irresponsibility.

We don’t need infallible humans. We need humans capable of making decisions, acknowledging their limitations, repairing their mistakes, and taking responsibility for their choices.

Because the most important strength our era needs might not be technical or algorithmic. It might be courage.

And unlike intelligence, that’s something we can’t simply outsource.

Or perhaps the real danger isn’t that AI might one day become a moral agent. It’s that we might start treating it as one before we’ve even figured out what that means.